diff --git a/.github/workflows/bench.yml.disabled b/.github/workflows/bench.yml.disabled index bfdbb4ef5..1c8787ef7 100644 --- a/.github/workflows/bench.yml.disabled +++ b/.github/workflows/bench.yml.disabled @@ -27,10 +27,10 @@ on: push: branches: - master - paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp'] + paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp'] pull_request_target: types: [opened, synchronize, reopened] - paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp'] + paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp'] schedule: - cron: '04 2 * * *' diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index c71d422e7..423173b97 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -19,6 +19,11 @@ concurrency: group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} cancel-in-progress: true +# Fine-grant permission +# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token +permissions: + contents: write # for creating release + env: BRANCH_NAME: ${{ github.head_ref || github.ref_name }} GGML_NLOOP: 3 diff --git a/.github/workflows/close-issue.yml b/.github/workflows/close-issue.yml index 69c9f4f69..f63860d14 100644 --- a/.github/workflows/close-issue.yml +++ b/.github/workflows/close-issue.yml @@ -3,6 +3,11 @@ on: schedule: - cron: "42 0 * * *" +# Fine-grant permission +# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token +permissions: + issues: write + jobs: close-issues: runs-on: ubuntu-latest diff --git a/.github/workflows/nix-ci-aarch64.yml b/.github/workflows/nix-ci-aarch64.yml index 4aa4b2379..0da6acdf1 100644 --- a/.github/workflows/nix-ci-aarch64.yml +++ b/.github/workflows/nix-ci-aarch64.yml @@ -21,6 +21,13 @@ concurrency: group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} cancel-in-progress: true +# Fine-grant permission +# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token +permissions: + # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub + id-token: write + contents: read + jobs: nix-build-aarch64: runs-on: ubuntu-latest diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml index 8955f38d0..8ecbbe53b 100644 --- a/.github/workflows/nix-ci.yml +++ b/.github/workflows/nix-ci.yml @@ -12,6 +12,13 @@ concurrency: group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} cancel-in-progress: true +# Fine-grant permission +# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token +permissions: + # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub + id-token: write + contents: read + jobs: nix-eval: strategy: diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 3d7c6f86c..4c882c254 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -1,24 +1,23 @@ # Pull requests (for contributors) - Test your changes: - - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library + - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the `ggml` library - Execute [the full CI locally on your machine](ci/README.md) before publishing -- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs. - - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience -- Consider allowing write access to your branch for faster review +- Optionally rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs +- Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly - If your PR becomes stale, don't hesitate to ping the maintainers in the comments # Pull requests (for collaborators) - Squash-merge PRs - Use the following format for the squashed commit title: ` : (#)`. For example: `utils : fix typo in utils.py (#1234)` -- Optionally, pick a `` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules +- Optionally pick a `` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules # Coding guidelines - Avoid adding third-party dependencies, extra files, extra headers, etc. - Always consider cross-compatibility with other operating systems and architectures -- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple +- Avoid fancy-looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a` - Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963) diff --git a/Makefile b/Makefile index 8a903d7ed..2793978c3 100644 --- a/Makefile +++ b/Makefile @@ -5,7 +5,6 @@ BUILD_TARGETS = \ llama-batched \ llama-batched-bench \ llama-bench \ - llama-benchmark-matmult \ llama-cli \ llama-convert-llama2c-to-ggml \ llama-embedding \ @@ -68,7 +67,7 @@ TEST_TARGETS = \ # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \ simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \ - retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm + retrieval speculative infill tokenize parallel export-lora lookahead lookup passkey gritlm # Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them. # We don't want to clutter things too much, so we only build replacements for the most commonly used binaries. @@ -1055,10 +1054,11 @@ ggml/src/ggml-alloc.o: \ $(CC) $(CFLAGS) -c $< -o $@ ggml/src/ggml-backend.o: \ - ggml/src/ggml-backend.c \ + ggml/src/ggml-backend.cpp \ + ggml/src/ggml-backend-impl.h \ ggml/include/ggml.h \ ggml/include/ggml-backend.h - $(CC) $(CFLAGS) -c $< -o $@ + $(CXX) $(CXXFLAGS) -c $< -o $@ ggml/src/ggml-quants.o: \ ggml/src/ggml-quants.c \ @@ -1523,16 +1523,6 @@ common/build-info.o: common/build-info.cpp tests: $(TEST_TARGETS) -llama-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp \ - $(OBJ_GGML) common/build-info.o - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - -run-benchmark-matmult: llama-benchmark-matmult - ./$@ - -.PHONY: run-benchmark-matmult swift - tests/test-arg-parser: tests/test-arg-parser.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) diff --git a/Package.swift b/Package.swift index 1d90b47bf..3a17e6c34 100644 --- a/Package.swift +++ b/Package.swift @@ -11,7 +11,7 @@ var sources = [ "src/unicode-data.cpp", "ggml/src/ggml.c", "ggml/src/ggml-alloc.c", - "ggml/src/ggml-backend.c", + "ggml/src/ggml-backend.cpp", "ggml/src/ggml-quants.c", "ggml/src/ggml-aarch64.c", ] diff --git a/README.md b/README.md index ecc2df8ca..41e5e5448 100644 --- a/README.md +++ b/README.md @@ -92,6 +92,7 @@ Typically finetunes of the base models below are supported as well. - [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct) - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a) - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat) +- [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a) (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md)) @@ -168,6 +169,7 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [AIKit](https://github.com/sozercan/aikit) (MIT) - [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL) - [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT) +- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL) *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)* diff --git a/ci/run.sh b/ci/run.sh index 7d241ecc0..e06778219 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -1,4 +1,4 @@ -#/bin/bash +#!/bin/bash # # sample usage: # @@ -751,7 +751,8 @@ function gg_run_rerank_tiny { model_f16="${path_models}/ggml-model-f16.gguf" - (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?hi\nwhat is panda?it's a bear\nwhat is panda?The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log + # for this model, the SEP token is "" + (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?hi\nwhat is panda?it's a bear\nwhat is panda?The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log # sample output # rerank score 0: 0.029 @@ -774,7 +775,7 @@ function gg_run_rerank_tiny { check_score "rerank score 0" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 0")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log - check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.15" | tee -a $OUT/${ci}-rk-f16.log + check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.30" | tee -a $OUT/${ci}-rk-f16.log set +e } diff --git a/common/arg.cpp b/common/arg.cpp index 8266a16c2..2a85ad845 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -911,7 +911,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, ).set_sparam()); add_opt(llama_arg( {"-s", "--seed"}, "SEED", - format("RNG seed (default: %u, use random seed for %u)", params.sparams.seed, LLAMA_DEFAULT_SEED), + format("RNG seed (default: %d, use random seed for %d)", params.sparams.seed, LLAMA_DEFAULT_SEED), [](gpt_params & params, const std::string & value) { params.sparams.seed = std::stoul(value); } diff --git a/common/common.cpp b/common/common.cpp index a0611f3d1..29df16c95 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -838,6 +838,31 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { return iparams; } + if (params.reranking) { + bool ok = true; + + if (llama_token_bos(model) == LLAMA_TOKEN_NULL) { + LOG_WRN("%s: warning: model does not have a BOS token, reranking will not work\n", __func__); + ok = false; + } + + if (llama_token_eos(model) == LLAMA_TOKEN_NULL) { + LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__); + ok = false; + } + + if (llama_token_sep(model) == LLAMA_TOKEN_NULL) { + LOG_WRN("%s: warning: model does not have a SEP token, reranking will not work\n", __func__); + ok = false; + } + + if (!ok) { + llama_free_model(model); + + return iparams; + } + } + auto cparams = llama_context_params_from_gpt_params(params); llama_context * lctx = llama_new_context_with_model(model, cparams); @@ -855,6 +880,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { if (cvec.n_embd == -1) { llama_free(lctx); llama_free_model(model); + return iparams; } @@ -867,6 +893,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { if (err) { llama_free(lctx); llama_free_model(model); + return iparams; } } @@ -889,7 +916,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { llama_lora_adapters_apply(lctx, iparams.lora_adapters); } - if (params.sparams.ignore_eos && llama_token_eos(model) == -1) { + if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) { LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__); params.sparams.ignore_eos = false; } @@ -930,6 +957,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { iparams.model = model; iparams.context = lctx; + return iparams; } diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index f3857d487..da5feb25b 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -15,6 +15,7 @@ from enum import IntEnum from pathlib import Path from hashlib import sha256 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast +from itertools import chain import math import numpy as np @@ -64,7 +65,6 @@ class Model: model_name: str | None metadata_override: Path | None dir_model_card: Path - is_lora: bool # subclasses should define this! model_arch: gguf.MODEL_ARCH @@ -72,7 +72,7 @@ class Model: def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False, use_temp_file: bool = False, eager: bool = False, metadata_override: Path | None = None, model_name: str | None = None, - split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False, is_lora: bool = False): + split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False): if type(self) is Model: raise TypeError(f"{type(self).__name__!r} should not be directly instantiated") @@ -94,7 +94,6 @@ class Model: self.metadata_override = metadata_override self.model_name = model_name self.dir_model_card = dir_model # overridden in convert_lora_to_gguf.py - self.is_lora = is_lora # true if model is used inside convert_lora_to_gguf.py # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type if self.ftype == gguf.LlamaFileType.GUESSED: @@ -270,10 +269,14 @@ class Model: return False + # some models need extra generated tensors (like rope_freqs) + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + return () + def prepare_tensors(self): max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,") - for name, data_torch in self.get_tensors(): + for name, data_torch in chain(self.generate_extra_tensors(), self.get_tensors()): # we don't need these if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): continue @@ -1617,7 +1620,7 @@ class LlamaModel(Model): return [(self.map_tensor_name(name), data_torch)] - def prepare_tensors(self): + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): if rope_scaling.get("rope_type", '').lower() == "llama3": base = self.hparams.get("rope_theta", 10000.0) @@ -1644,9 +1647,9 @@ class LlamaModel(Model): smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) rope_factors.append(1 / ((1 - smooth) / factor + smooth)) - if not self.is_lora: - self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32)) + def prepare_tensors(self): super().prepare_tensors() if self._experts is not None: @@ -1870,8 +1873,6 @@ class MiniCPM3Model(Model): def set_gguf_parameters(self): hparams = self.hparams - rope_dims = hparams["qk_rope_head_dim"] - self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hparams["hidden_size"]) @@ -1887,24 +1888,25 @@ class MiniCPM3Model(Model): self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: rope_scaling = self.find_hparam(['rope_scaling'], True) - if rope_scaling is None: - return + if rope_scaling is not None: + rope_dims = self.hparams["qk_rope_head_dim"] - long_factors = rope_scaling.get('long_factor', None) - short_factors = rope_scaling.get('short_factor', None) + long_factors = rope_scaling.get('long_factor', None) + short_factors = rope_scaling.get('short_factor', None) - if long_factors is None or short_factors is None: - raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor') + if long_factors is None or short_factors is None: + raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor') - if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: - raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') + if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: + raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32)) - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32)) def set_vocab(self): - self._set_vocab_llama_hf() + self._set_vocab_sentencepiece() def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor: if n_kv_head is not None and n_head != n_kv_head: @@ -2216,6 +2218,13 @@ class Phi3MiniModel(Model): self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"])) + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + n_embd = self.find_hparam(["hidden_size", "n_embd"]) + n_head = self.find_hparam(["num_attention_heads", "n_head"]) + max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"]) + orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"]) + rope_dims = n_embd // n_head + # write rope scaling for long context (128k) model rope_scaling = self.find_hparam(['rope_scaling'], True) if rope_scaling is None: @@ -2245,9 +2254,8 @@ class Phi3MiniModel(Model): if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') - if not self.is_lora: - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32)) - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32)) @Model.register("PlamoForCausalLM") @@ -4071,7 +4079,7 @@ class ExaoneModel(Model): self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"]) - def prepare_tensors(self): + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): if rope_scaling.get("rope_type", '').lower() == "llama3": base = self.hparams.get("rope_theta", 10000.0) @@ -4098,10 +4106,7 @@ class ExaoneModel(Model): smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) rope_factors.append(1 / ((1 - smooth) / factor + smooth)) - if not self.is_lora: - self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32)) - - super().prepare_tensors() + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32)) @Model.register("GraniteForCausalLM") diff --git a/convert_lora_to_gguf.py b/convert_lora_to_gguf.py index d1c94e580..439a78de1 100755 --- a/convert_lora_to_gguf.py +++ b/convert_lora_to_gguf.py @@ -331,6 +331,10 @@ if __name__ == '__main__': self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha) super().set_gguf_parameters() + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + # Never add extra tensors (e.g. rope_freqs) for LoRA adapters + return () + def get_tensors(self) -> Iterator[tuple[str, Tensor]]: tensor_map: dict[str, PartialLoraTensor] = {} @@ -392,7 +396,6 @@ if __name__ == '__main__': dry_run=args.dry_run, dir_lora_model=dir_lora, lora_alpha=alpha, - is_lora=True, ) logger.info("Exporting model...") diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md index bc266f7d8..ea34182e4 100644 --- a/docs/backend/SYCL.md +++ b/docs/backend/SYCL.md @@ -26,7 +26,7 @@ ### Llama.cpp + SYCL -The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it could support other vendor GPUs: Nvidia GPU (*AMD GPU coming*). +The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it also supports other vendor GPUs: Nvidia and AMD. ## Recommended Release @@ -111,10 +111,18 @@ SYCL backend supports Intel GPU Family: **Verified devices** -| Nvidia GPU | Status | Verified Model | -|--------------------------|---------|----------------| -| Ampere Series | Support | A100, A4000 | -| Ampere Series *(Mobile)* | Support | RTX 40 Series | +| Nvidia GPU | Status | Verified Model | +|--------------------------|-----------|----------------| +| Ampere Series | Supported | A100, A4000 | +| Ampere Series *(Mobile)* | Supported | RTX 40 Series | + +| AMD GPU | Status | Verified Model | +|--------------------------|--------------|----------------| +| Radeon Pro | Experimental | W6800 | +| Radeon RX | Experimental | 6700 XT | + +Note: AMD GPU support is highly experimental and is incompatible with F16. +Additionally, it only supports GPUs with a sub_group_size (warp size) of 32. ## Docker The docker build option is currently limited to *intel GPU* targets. @@ -186,6 +194,10 @@ Platform #0: Intel(R) OpenCL HD Graphics In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed. +- **AMD GPU** + +To target AMD GPUs with SYCL, the ROCm stack must be installed first. + 2. **Install IntelĀ® oneAPI Base toolkit** - **For Intel GPU** @@ -212,6 +224,19 @@ cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENAB cmake --build buildWithCublas --config Release ``` +- **Adding support to AMD GPUs** + +**oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit. + +**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs. + +```sh +git clone https://github.com/oneapi-src/oneMKL +cd oneMKL +# Find your HIPTARGET with rocminfo, under the key 'Name:' +cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas +cmake --build buildWithrocBLAS --config Release +``` 3. **Verify installation and environment** @@ -223,22 +248,32 @@ sycl-ls - **Intel GPU** -When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`ext_oneapi_level_zero:gpu:0`] in the sample output below: +When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`level_zero:gpu`] in the sample output below: ``` -[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.10.0.17_160000] -[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000] -[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO [23.30.26918.50] -[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918] +[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.10.0.17_160000] +[opencl:cpu][opencl:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000] +[opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO [23.30.26918.50] +[level_zero:gpu][level_zero:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918] ``` - **Nvidia GPU** -Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow: +Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`cuda:gpu`] as below: + ``` -[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.12.0.12_195853.xmain-hotfix] -[opencl:cpu:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix] -[ext_oneapi_cuda:gpu:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.2] +[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.12.0.12_195853.xmain-hotfix] +[opencl:cpu][opencl:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix] +[cuda:gpu][cuda:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.5] +``` + +- **AMD GPU** + +For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]: + +``` +[opencl:cpu][opencl:0] Intel(R) OpenCL, 12th Gen Intel(R) Core(TM) i9-12900K OpenCL 3.0 (Build 0) [2024.18.6.0.02_160000] +[hip:gpu][hip:0] AMD HIP BACKEND, AMD Radeon PRO W6800 gfx1030 [HIP 60140.9] ``` ### II. Build llama.cpp @@ -266,6 +301,7 @@ cmake --build build --config Release -j -v ``` #### Nvidia GPU + ```sh # Export relevant ENV variables export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH @@ -283,7 +319,25 @@ cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx - # build all binary cmake --build build --config Release -j -v +``` +#### AMD GPU + +```sh +# Export relevant ENV variables +export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH +export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH +export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR + +# Build LLAMA with rocBLAS acceleration through SYCL + +## AMD +# Use FP32, FP16 is not supported +# Find your GGML_SYCL_HIP_TARGET with rocminfo, under the key 'Name:' +cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_HIP_TARGET=${GGML_SYCL_HIP_TARGET} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx + +# build all binary +cmake --build build --config Release -j -v ``` ### III. Run the inference @@ -586,11 +640,11 @@ use 1 SYCL GPUs: [0] with Max compute units:512 #### Build -| Name | Value | Function | -|--------------------|-----------------------------------|---------------------------------------------| -| GGML_SYCL | ON (mandatory) | Enable build with SYCL code path.
FP32 path - recommended for better perforemance than FP16 on quantized model| -| GGML_SYCL_TARGET | INTEL *(default)* \| NVIDIA | Set the SYCL target device type. | -| GGML_SYCL_F16 | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path. | +| Name | Value | Function | +|--------------------|---------------------------------------|---------------------------------------------| +| GGML_SYCL | ON (mandatory) | Enable build with SYCL code path.
FP32 path - recommended for better perforemance than FP16 on quantized model| +| GGML_SYCL_TARGET | INTEL *(default)* \| NVIDIA \| AMD | Set the SYCL target device type. | +| GGML_SYCL_F16 | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path. | | CMAKE_C_COMPILER | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path. | | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)* | Set `icpx/icx` compiler for SYCL code path. | diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 67b3d2774..ead630661 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -16,7 +16,6 @@ else() add_subdirectory(baby-llama) add_subdirectory(batched-bench) add_subdirectory(batched) - add_subdirectory(benchmark) add_subdirectory(convert-llama2c-to-ggml) add_subdirectory(embedding) add_subdirectory(eval-callback) diff --git a/examples/benchmark/CMakeLists.txt b/examples/benchmark/CMakeLists.txt deleted file mode 100644 index 34a58cc02..000000000 --- a/examples/benchmark/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -set(TARGET llama-bench-matmult) -add_executable(${TARGET} benchmark-matmult.cpp) -install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT}) -target_include_directories(${TARGET} PRIVATE ../../common) -target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp deleted file mode 100644 index 922daf528..000000000 --- a/examples/benchmark/benchmark-matmult.cpp +++ /dev/null @@ -1,275 +0,0 @@ -#include "common.h" -#include "ggml.h" - -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include - -#if defined(_MSC_VER) -#pragma warning(disable: 4244 4267) // possible loss of data -#endif - -static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { - struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr); - - if (plan.work_size > 0) { - buf.resize(plan.work_size); - plan.work_data = buf.data(); - } - - ggml_graph_compute(graph, &plan); -} - -static float tensor_sum_elements(const ggml_tensor * tensor) { - double sum = 0; - if (tensor->type == GGML_TYPE_F32) { - for (int j = 0; j < tensor->ne[1]; j++) { - for (int k = 0; k < tensor->ne[0]; k++) { - sum += ((float *) tensor->data)[j*tensor->ne[0] + k]; - } - } - } - return sum; -} - -static void tensor_dump(const ggml_tensor * tensor, const char * name) { - printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name, - tensor->type, ggml_type_name(tensor->type), - tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]); - float sum = tensor_sum_elements(tensor); - printf("Sum of tensor %s is %6.2f\n", name, sum); -} - -#define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor) - -struct benchmark_params_struct { - int n_threads = 1; - int32_t n_iterations = 10; -}; - -static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) { - fprintf(stderr, "usage: %s [options]\n", argv[0]); - fprintf(stderr, "\n"); - fprintf(stderr, "options:\n"); - fprintf(stderr, " -h, --help show this help message and exit\n"); - fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads); - fprintf(stderr, " -i N, --iter N number of iterations to use during computation (default: %d)\n", params.n_iterations); - fprintf(stderr, "\n"); -} - -int main(int argc, char ** argv) { - struct benchmark_params_struct benchmark_params; - - bool invalid_param = false; - std::string arg; - for (int i = 1; i < argc; i++) { - arg = argv[i]; - - if (arg == "-t" || arg == "--threads") { - if (++i >= argc) { - invalid_param = true; - break; - } - benchmark_params.n_threads = std::stoi(argv[i]); - } else if (arg == "-i" || arg == "--iter") { - if (++i >= argc) { - invalid_param = true; - break; - } - benchmark_params.n_iterations = std::stoi(argv[i]); - } else if (arg == "-h" || arg == "--help") { - print_usage(argc, argv, benchmark_params); - exit(0); - } - } - if (invalid_param) { - fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); - print_usage(argc, argv, benchmark_params); - exit(1); - } - - print_build_info(); - printf("Starting Test\n"); - - // create the ggml context - struct ggml_context * ctx; - //const int sizex = 4096; - //const int sizey = 11008; - -#undef VERBOSE_DEBUGGING -#ifndef VERBOSE_DEBUGGING - const int sizey = 4096; - const int sizex = 11008; - const int sizez = 128; -#else - /* Working - let's increase size */ - const int sizey = 1; - const int sizex = (8*32); - const int sizez = 1; - - /*const int sizey = 1; - const int sizex = 3*(8*32); - const int sizez = 1;*/ -#endif - - //printf("Memsize required = %i\n", sizex*sizex); - - // TODO: perform the bench for all types or for a user specified type - const ggml_type qtype = GGML_TYPE_Q4_1; - - size_t ctx_size = 0; - ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); - ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); - ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez); - ctx_size += ggml_row_size(qtype, sizex*sizey); - ctx_size += ggml_row_size(qtype, sizex*sizey); - ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS - ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS - ctx_size += 1024*1024*16; - - printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024)); - - struct ggml_init_params params = { - /*.mem_size =*/ ctx_size, - /*.mem_buffer =*/ NULL, - /* no_alloc =*/ 0 - }; - - ctx = ggml_init(params); - if (!ctx) { - fprintf(stderr, "%s: ggml_init() failed\n", __func__); - return 1; - } - - - printf("Creating new tensors\n"); - // printf("Creating new tensor m1\n"); - struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey); - ggml_set_f32(m11, 1.0f); - - // printf("Creating new tensor m1\n"); - struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey); - ggml_set_f32(m12, 1.5f); - - // printf("Creating new tensor m2\n"); - struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez); - ggml_set_f32(m2, 2.0f); - - printf("\n------ Test 1 - Matrix Mult via F32 code\n"); - // printf("Creating new tensor m11xm2\n"); - struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2); - - // printf("Creating compute graph\n"); - struct ggml_cgraph * gf = ggml_new_graph(ctx); - ggml_build_forward_expand(gf, m11xm2); - - printf("n_threads=%i\n", benchmark_params.n_threads); - - TENSOR_DUMP(m11); - TENSOR_DUMP(m2); - - std::vector work_buffer; - - ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads); - - TENSOR_DUMP(ggml_graph_node(gf, 0)); - - printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype)); - - int32_t nelements = sizex*sizey; - - // Set up a the benchmark matrices - // printf("Creating new tensor q11 & Running quantize\n"); - struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey); - ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr); - - // Set up a the compute graph - // printf("Creating new tensor q31\n"); - struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2); - - // printf("Creating compute graph\n"); - struct ggml_cgraph * gf31 = ggml_new_graph(ctx); - ggml_build_forward_expand(gf31, q31); - - // Set up a second graph computation to make sure we override the CPU cache lines - // printf("Creating new tensor q12 & Running quantize\n"); - struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey); - ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr); - - // printf("Creating new tensor q32\n"); - struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2); - - //printf("Creating compute graph\n"); - struct ggml_cgraph * gf32 = ggml_new_graph(ctx); - ggml_build_forward_expand(gf32, q32); - printf("n_threads=%i\n", benchmark_params.n_threads); - - const int dimx = sizex; - const int dimy = sizey; - const int dimz = sizez; - long long int flops_per_dot_product = dimy + dimy; - long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ; - printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000); - - - // Let's use the F32 result from above as a reference for the quantized multiplication - float sum_of_F32_reference = tensor_sum_elements(ggml_graph_node(gf, 0)); - - printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n"); - printf("=====================================================================================\n"); - - double gflops_sum = 0; - for (int i=0;i allowed_delta) { - printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n", - sum_of_F32_reference, - sum_of_Q4_result, - delta, - allowed_delta - ); - exit(0); - } - - // Running a different graph computation to make sure we override the CPU cache lines - ggml_graph_compute_helper(work_buffer, gf32, benchmark_params.n_threads); - } - printf("\n"); - printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations)); - printf("=====================================================================================\n"); -} diff --git a/examples/cvector-generator/pca.hpp b/examples/cvector-generator/pca.hpp index a969c486d..f6e307fbc 100644 --- a/examples/cvector-generator/pca.hpp +++ b/examples/cvector-generator/pca.hpp @@ -204,13 +204,6 @@ static ggml_status compute_piter( ggml_backend_cpu_set_n_threads(model.backend, params.n_threads); } -// TODO: enable GPU support when support for GGML_OP_SQRT is added -//#ifdef GGML_USE_METAL -// if (ggml_backend_is_metal(model.backend)) { -// ggml_backend_metal_set_n_cb(model.backend, params.n_threads); -// } -//#endif - ggml_status res = ggml_backend_graph_compute(model.backend, gf); if (res == GGML_STATUS_SUCCESS) { auto extract_i = [](std::string prefix, std::string str) -> int { diff --git a/examples/gguf-split/gguf-split.cpp b/examples/gguf-split/gguf-split.cpp index 82c239b83..7e62657e1 100644 --- a/examples/gguf-split/gguf-split.cpp +++ b/examples/gguf-split/gguf-split.cpp @@ -22,12 +22,20 @@ #endif enum split_operation : uint8_t { - SPLIT_OP_SPLIT, - SPLIT_OP_MERGE, + OP_NONE, + OP_SPLIT, + OP_MERGE, +}; + +enum split_mode : uint8_t { + MODE_NONE, + MODE_TENSOR, + MODE_SIZE, }; struct split_params { - split_operation operation = SPLIT_OP_SPLIT; + split_operation operation = OP_NONE; + split_mode mode = MODE_NONE; size_t n_bytes_split = 0; int n_split_tensors = 128; std::string input; @@ -87,59 +95,52 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p } bool arg_found = false; - bool is_op_set = false; - bool is_mode_set = false; if (arg == "-h" || arg == "--help") { split_print_usage(argv[0]); exit(0); - } - if (arg == "--version") { + } else if (arg == "--version") { fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT); fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET); exit(0); - } - if (arg == "--dry-run") { + } else if (arg == "--dry-run") { arg_found = true; params.dry_run = true; - } - if (arg == "--no-tensor-first-split") { + } else if (arg == "--no-tensor-first-split") { arg_found = true; params.no_tensor_first_split = true; - } - - if (is_op_set) { - throw std::invalid_argument("error: either --split or --merge can be specified, but not both"); - } - if (arg == "--merge") { + } else if (arg == "--merge") { arg_found = true; - is_op_set = true; - params.operation = SPLIT_OP_MERGE; - } - if (arg == "--split") { + if (params.operation != OP_NONE && params.operation != OP_MERGE) { + throw std::invalid_argument("error: either --split or --merge can be specified, but not both"); + } + params.operation = OP_MERGE; + } else if (arg == "--split") { arg_found = true; - is_op_set = true; - params.operation = SPLIT_OP_SPLIT; - } - - if (is_mode_set) { - throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both"); - } - if (arg == "--split-max-tensors") { + if (params.operation != OP_NONE && params.operation != OP_SPLIT) { + throw std::invalid_argument("error: either --split or --merge can be specified, but not both"); + } + params.operation = OP_SPLIT; + } else if (arg == "--split-max-tensors") { if (++arg_idx >= argc) { invalid_param = true; break; } arg_found = true; - is_mode_set = true; + if (params.mode != MODE_NONE && params.mode != MODE_TENSOR) { + throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both"); + } + params.mode = MODE_TENSOR; params.n_split_tensors = atoi(argv[arg_idx]); - } - if (arg == "--split-max-size") { + } else if (arg == "--split-max-size") { if (++arg_idx >= argc) { invalid_param = true; break; } arg_found = true; - is_mode_set = true; + if (params.mode != MODE_NONE && params.mode != MODE_SIZE) { + throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both"); + } + params.mode = MODE_SIZE; params.n_bytes_split = split_str_to_n_bytes(argv[arg_idx]); } @@ -148,6 +149,15 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p } } + // the operation is split if not specified + if (params.operation == OP_NONE) { + params.operation = OP_SPLIT; + } + // the split mode is by tensor if not specified + if (params.mode == MODE_NONE) { + params.mode = MODE_TENSOR; + } + if (invalid_param) { throw std::invalid_argument("error: invalid parameter for argument: " + arg); } @@ -265,13 +275,15 @@ struct split_strategy { } bool should_split(int i_tensor, size_t next_size) { - if (params.n_bytes_split > 0) { + if (params.mode == MODE_SIZE) { // split by max size per file return next_size > params.n_bytes_split; - } else { + } else if (params.mode == MODE_TENSOR) { // split by number of tensors per file return i_tensor > 0 && i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0; } + // should never happen + GGML_ABORT("invalid mode"); } void print_info() { @@ -559,9 +571,9 @@ int main(int argc, const char ** argv) { split_params_parse(argc, argv, params); switch (params.operation) { - case SPLIT_OP_SPLIT: gguf_split(params); + case OP_SPLIT: gguf_split(params); break; - case SPLIT_OP_MERGE: gguf_merge(params); + case OP_MERGE: gguf_merge(params); break; default: split_print_usage(argv[0]); exit(EXIT_FAILURE); diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 8aa7b0750..14e02c8dd 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -2444,12 +2444,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima ggml_backend_cpu_set_n_threads(ctx->backend, n_threads); } -#ifdef GGML_USE_METAL - if (ggml_backend_is_metal(ctx->backend)) { - ggml_backend_metal_set_n_cb(ctx->backend, n_threads); - } -#endif - ggml_backend_graph_compute(ctx->backend, gf); // the last node is the embedding tensor diff --git a/examples/main/README.md b/examples/main/README.md index 6730effdf..f0c3031ab 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -69,7 +69,7 @@ In this section, we cover the most commonly used options for running the `llama- - `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. - `-mli, --multiline-input`: Allows you to write or paste multiple lines without ending each in '\' - `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has. -- - `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance. +- `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance. ## Input Prompts diff --git a/examples/rpc/rpc-server.cpp b/examples/rpc/rpc-server.cpp index 6342e6488..355125831 100644 --- a/examples/rpc/rpc-server.cpp +++ b/examples/rpc/rpc-server.cpp @@ -6,6 +6,10 @@ #include "ggml-metal.h" #endif +#ifdef GGML_USE_VULKAN +#include "ggml-vulkan.h" +#endif + #include "ggml-rpc.h" #ifdef _WIN32 # include @@ -79,6 +83,12 @@ static ggml_backend_t create_backend() { if (!backend) { fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__); } +#elif GGML_USE_VULKAN + fprintf(stderr, "%s: using Vulkan backend\n", __func__); + backend = ggml_backend_vk_init(0); // init device 0 + if (!backend) { + fprintf(stderr, "%s: ggml_backend_vulkan_init() failed\n", __func__); + } #endif // if there aren't GPU Backends fallback to CPU backend @@ -92,6 +102,8 @@ static ggml_backend_t create_backend() { static void get_backend_memory(size_t * free_mem, size_t * total_mem) { #ifdef GGML_USE_CUDA ggml_backend_cuda_get_device_memory(0, free_mem, total_mem); +#elif GGML_USE_VULKAN + ggml_backend_vk_get_device_memory(0, free_mem, total_mem); #else #ifdef _WIN32 MEMORYSTATUSEX status; diff --git a/examples/server/README.md b/examples/server/README.md index 951c4a44c..6253de43c 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -100,7 +100,7 @@ The project is under active development, and we are [looking for feedback and co | Argument | Explanation | | -------- | ----------- | | `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'
(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) | -| `-s, --seed SEED` | RNG seed (default: 4294967295, use random seed for 4294967295) | +| `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) | | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) | | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) | | `--penalize-nl` | penalize newline tokens (default: false) | diff --git a/examples/server/server.cpp b/examples/server/server.cpp index f343cc252..13e54e501 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2027,7 +2027,7 @@ struct server_context { continue; } - // prompt: querydoc + // prompt: [BOS]query[EOS][SEP]doc[EOS] prompt_tokens.clear(); prompt_tokens.push_back(llama_token_bos(model)); { @@ -2035,7 +2035,7 @@ struct server_context { prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end()); } prompt_tokens.push_back(llama_token_eos(model)); - prompt_tokens.push_back(llama_token_bos(model)); + prompt_tokens.push_back(llama_token_sep(model)); { const auto part = tokenize(slot.prompt[1], false); prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end()); diff --git a/ggml/include/ggml-alloc.h b/ggml/include/ggml-alloc.h index 0dff47d65..23600eea9 100644 --- a/ggml/include/ggml-alloc.h +++ b/ggml/include/ggml-alloc.h @@ -24,7 +24,7 @@ GGML_API void ggml_tallocr_alloc(struct ggml_tallocr * talloc, st // Graph allocator /* Example usage: - ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type()); + ggml_gallocr_t galloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type()); // optional: create a worst-case graph and reserve the buffers to avoid reallocations ggml_gallocr_reserve(galloc, build_graph(max_batch)); diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h index 71c0bef8e..152b9adb0 100644 --- a/ggml/include/ggml-backend.h +++ b/ggml/include/ggml-backend.h @@ -12,43 +12,52 @@ extern "C" { typedef struct ggml_backend_event * ggml_backend_event_t; typedef struct ggml_backend * ggml_backend_t; typedef void * ggml_backend_graph_plan_t; + typedef struct ggml_backend_reg * ggml_backend_reg_t; + typedef struct ggml_backend_device * ggml_backend_dev_t; + + + // + // Backend buffer type + // + + GGML_API const char * ggml_backend_buft_name (ggml_backend_buffer_type_t buft); + GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer (ggml_backend_buffer_type_t buft, size_t size); + GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft); + GGML_API size_t ggml_backend_buft_get_max_size (ggml_backend_buffer_type_t buft); + GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); + GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft); + GGML_API ggml_backend_dev_t ggml_backend_buft_get_device (ggml_backend_buffer_type_t buft); // // Backend buffer // - // buffer type - GGML_API const char * ggml_backend_buft_name (ggml_backend_buffer_type_t buft); - GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer (ggml_backend_buffer_type_t buft, size_t size); - GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft); - GGML_API size_t ggml_backend_buft_get_max_size (ggml_backend_buffer_type_t buft); - GGML_API GGML_CALL size_t ggml_backend_buft_get_alloc_size (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); - GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft); - - // buffer enum ggml_backend_buffer_usage { GGML_BACKEND_BUFFER_USAGE_ANY = 0, GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1, GGML_BACKEND_BUFFER_USAGE_COMPUTE = 2, }; - GGML_API const char * ggml_backend_buffer_name (ggml_backend_buffer_t buffer); - GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer); - GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer); - GGML_API GGML_CALL void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_max_size (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value); - GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer); - GGML_API void ggml_backend_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage); - GGML_API enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage (ggml_backend_buffer_t buffer); - GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type (ggml_backend_buffer_t buffer); - GGML_API void ggml_backend_buffer_reset (ggml_backend_buffer_t buffer); + GGML_API const char * ggml_backend_buffer_name (ggml_backend_buffer_t buffer); + GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer); + GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer); + GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer); + GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); + GGML_API size_t ggml_backend_buffer_get_max_size (ggml_backend_buffer_t buffer); + GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value); + GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer); + GGML_API void ggml_backend_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage); + GGML_API enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage (ggml_backend_buffer_t buffer); + GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type (ggml_backend_buffer_t buffer); + GGML_API void ggml_backend_buffer_reset (ggml_backend_buffer_t buffer); + + // tensor copy between different backends + GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst); // - // Backend + // Backend (stream) // GGML_API ggml_guid_t ggml_backend_guid(ggml_backend_t backend); @@ -64,9 +73,9 @@ extern "C" { GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); // "offset" refers to the offset of the tensor data for setting/getting data - GGML_API GGML_CALL void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - GGML_API GGML_CALL void ggml_backend_tensor_memset( struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size); + GGML_API void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + GGML_API void ggml_backend_tensor_memset( struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size); GGML_API void ggml_backend_synchronize(ggml_backend_t backend); @@ -76,65 +85,120 @@ extern "C" { GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan); GGML_API enum ggml_status ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph); GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph); + + // NOTE: will be removed, use device version instead GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op); GGML_API bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft); GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op); - // tensor copy between different backends - GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst); - // asynchronous copy // the copy is performed after all the currently queued operations in backend_src // backend_dst will wait for the copy to complete before performing other operations // automatic fallback to sync copy if async is not supported GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst); - // events - GGML_API ggml_backend_event_t ggml_backend_event_new (ggml_backend_t backend); - GGML_API void ggml_backend_event_free (ggml_backend_event_t event); - GGML_API void ggml_backend_event_record (ggml_backend_event_t event); - GGML_API void ggml_backend_event_synchronize(ggml_backend_event_t event); - GGML_API void ggml_backend_event_wait (ggml_backend_t backend, ggml_backend_event_t event); + GGML_API ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend); // - // CPU backend + // Events // - GGML_API ggml_backend_t ggml_backend_cpu_init(void); + GGML_API ggml_backend_event_t ggml_backend_event_new(ggml_backend_dev_t device); + GGML_API void ggml_backend_event_free(ggml_backend_event_t event); + GGML_API void ggml_backend_event_record(ggml_backend_event_t event, ggml_backend_t backend); + GGML_API void ggml_backend_event_synchronize(ggml_backend_event_t event); + GGML_API void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event); - GGML_API GGML_CALL bool ggml_backend_is_cpu (ggml_backend_t backend); - GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads); - GGML_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool); - GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data); + // + // Backend device + // - // Create a backend buffer from an existing pointer - GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size); + enum ggml_backend_dev_type { + GGML_BACKEND_DEVICE_TYPE_CPU, + GGML_BACKEND_DEVICE_TYPE_GPU, + // devices with full capabilities (excludes backends such as BLAS that only support matrix multiplication) + GGML_BACKEND_DEVICE_TYPE_CPU_FULL, + GGML_BACKEND_DEVICE_TYPE_GPU_FULL + }; - GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void); + // functionality supported by the device + struct ggml_backend_dev_caps { + // asynchronous operations + bool async; + // pinned host buffer + bool host_buffer; + // creating buffers from host ptr + bool buffer_from_host_ptr; + // event synchronization + bool events; + }; -#ifdef GGML_USE_CPU_HBM - GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void); -#endif + // all the device properties + struct ggml_backend_dev_props { + const char * name; + const char * description; + size_t memory_free; + size_t memory_total; + enum ggml_backend_dev_type type; + struct ggml_backend_dev_caps caps; + }; + + GGML_API const char * ggml_backend_dev_name(ggml_backend_dev_t device); + GGML_API const char * ggml_backend_dev_description(ggml_backend_dev_t device); + GGML_API void ggml_backend_dev_memory(ggml_backend_dev_t device, size_t * free, size_t * total); + GGML_API enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device); + GGML_API void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props); + GGML_API ggml_backend_reg_t ggml_backend_dev_backend_reg(ggml_backend_dev_t device); + GGML_API ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * params); + GGML_API ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device); + GGML_API ggml_backend_buffer_type_t ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device); + GGML_API ggml_backend_buffer_t ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size); + + GGML_API bool ggml_backend_dev_supports_op(ggml_backend_dev_t device, const struct ggml_tensor * op); + GGML_API bool ggml_backend_dev_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft); + GGML_API bool ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_tensor * op); + + // + // Backend (reg) + // + + GGML_API const char * ggml_backend_reg_name(ggml_backend_reg_t reg); + GGML_API size_t ggml_backend_reg_dev_count(ggml_backend_reg_t reg); + GGML_API ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index); + GGML_API void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name); + + + // Functions that may be obtained using ggml_backend_reg_get_proc_address + typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(const float *); // // Backend registry // - // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way + // Backend (reg) enumeration + GGML_API size_t ggml_backend_reg_count(void); + GGML_API ggml_backend_reg_t ggml_backend_reg_get(size_t index); + GGML_API ggml_backend_reg_t ggml_backend_reg_by_name(const char * name); - GGML_API size_t ggml_backend_reg_get_count(void); - GGML_API size_t ggml_backend_reg_find_by_name(const char * name); // returns index of backend with name, or SIZE_MAX if not found - GGML_API ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is backend_name:params (params is optional) - GGML_API const char * ggml_backend_reg_get_name(size_t i); - GGML_API ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific - GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i); - GGML_API ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size); + // Device enumeration + GGML_API size_t ggml_backend_dev_count(void); + GGML_API ggml_backend_dev_t ggml_backend_dev_get(size_t index); + GGML_API ggml_backend_dev_t ggml_backend_dev_by_name(const char * name); + GGML_API ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type); + + // Direct backend (stream) initialization + // = ggml_backend_dev_init(ggml_backend_dev_by_name(name), params) + GGML_API ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params); + // = ggml_backend_dev_init(ggml_backend_dev_by_type(type), params) + GGML_API ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params); + // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU_FULL) OR ggml_backend_dev_by_type(CPU_FULL), NULL) + GGML_API ggml_backend_t ggml_backend_init_best(void); // // Backend scheduler // - // The backend scheduler allows for multiple backends to be used together + // The backend scheduler allows for multiple backend devices to be used together // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends // The backends are selected based on: // - the backend that supports the operation @@ -169,9 +233,9 @@ extern "C" { } */ - struct ggml_backend_sched; typedef struct ggml_backend_sched * ggml_backend_sched_t; + // Evaluation callback for each node in the graph (set with ggml_backend_sched_set_eval_callback) // when ask == true, the scheduler wants to know if the user wants to observe this node // this allows the scheduler to batch nodes together in order to evaluate them in a single call // @@ -185,7 +249,7 @@ extern "C" { GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); // Initialize backend buffers from a measure graph - GGML_API bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); + GGML_API bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); // returns success GGML_API int ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched); GGML_API ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i); @@ -200,7 +264,7 @@ extern "C" { GGML_API ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node); // Allocate and compute graph on the backend scheduler - GGML_API bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph); + GGML_API bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph); // returns success GGML_API enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph); GGML_API enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph); GGML_API void ggml_backend_sched_synchronize(ggml_backend_sched_t sched); @@ -226,7 +290,7 @@ extern "C" { GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph); GGML_API void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy); - typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data); + typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data); // Compare the output of two backends GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data); @@ -235,6 +299,26 @@ extern "C" { GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr); GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor); + // + // CPU backend + // + + GGML_API ggml_backend_t ggml_backend_cpu_init(void); + + GGML_API bool ggml_backend_is_cpu (ggml_backend_t backend); + GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads); + GGML_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool); + GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data); + + // Create a backend buffer from an existing pointer + GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size); + GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void); + + GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void); + +#ifdef GGML_USE_CPU_HBM + GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void); +#endif #ifdef __cplusplus } diff --git a/ggml/include/ggml-blas.h b/ggml/include/ggml-blas.h index f2e37de06..dd612860d 100644 --- a/ggml/include/ggml-blas.h +++ b/ggml/include/ggml-blas.h @@ -9,13 +9,13 @@ extern "C" { #endif // backend API -GGML_API GGML_CALL ggml_backend_t ggml_backend_blas_init(void); +GGML_API ggml_backend_t ggml_backend_blas_init(void); -GGML_API GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend); +GGML_API bool ggml_backend_is_blas(ggml_backend_t backend); // number of threads used for conversion to float // for openblas and blis, this will also set the number of threads used for blas operations -GGML_API GGML_CALL void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads); +GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads); #ifdef __cplusplus diff --git a/ggml/include/ggml-cann.h b/ggml/include/ggml-cann.h index 031ad1ce2..95bdaf10d 100644 --- a/ggml/include/ggml-cann.h +++ b/ggml/include/ggml-cann.h @@ -44,7 +44,7 @@ extern "C" { * @param device The index of the device to initialize. * @return A pointer to the initialized backend instance, or nullptr on failure. */ -GGML_API GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device); +GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device); /** * @brief Checks if a given backend is a CANN backend. @@ -55,7 +55,7 @@ GGML_API GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device); * @param backend The backend instance to check. * @return True if the backend is a CANN backend, false otherwise. */ -GGML_API GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend); +GGML_API bool ggml_backend_is_cann(ggml_backend_t backend); /** * @brief Retrieves the CANN buffer type for a specified device. @@ -67,7 +67,7 @@ GGML_API GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend); * @return A pointer to the buffer type interface for the specified device, or * nullptr if the device index is out of range. */ -GGML_API GGML_CALL ggml_backend_buffer_type_t +GGML_API ggml_backend_buffer_type_t ggml_backend_cann_buffer_type(int32_t device); /** @@ -78,14 +78,14 @@ ggml_backend_cann_buffer_type(int32_t device); * * @return The number of CANN devices available. */ -GGML_API GGML_CALL int32_t ggml_backend_cann_get_device_count(void); +GGML_API int32_t ggml_backend_cann_get_device_count(void); /** * @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU. * * @return A pointer to the host buffer type interface. */ -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void); +GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void); /** * @brief Retrieves the description of a specific CANN device. @@ -97,7 +97,7 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type * @param description Pointer to a buffer where the description will be written. * @param description_size Size of the description buffer. */ -GGML_API GGML_CALL void ggml_backend_cann_get_device_description( +GGML_API void ggml_backend_cann_get_device_description( int32_t device, char* description, size_t description_size); /** @@ -112,20 +112,9 @@ GGML_API GGML_CALL void ggml_backend_cann_get_device_description( * @param total Pointer to a variable where the total memory size will be * stored. */ -GGML_API GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device, - size_t* free, - size_t* total); - -/** - * @brief Set the logging callback for GGML. - * - * This function sets the logging callback and user data for logging. - * - * @param log_callback The logging callback to set. - * @param user_data User data to pass to the logging callback. - */ -GGML_API void ggml_backend_cann_log_set_callback(ggml_log_callback log_callback, - void* user_data); +GGML_API void ggml_backend_cann_get_device_memory(int32_t device, + size_t* free, + size_t* total); #ifdef __cplusplus } diff --git a/ggml/include/ggml-cuda.h b/ggml/include/ggml-cuda.h index 71bb6dcf0..f44d8f4e6 100644 --- a/ggml/include/ggml-cuda.h +++ b/ggml/include/ggml-cuda.h @@ -3,6 +3,10 @@ #include "ggml.h" #include "ggml-backend.h" +#ifdef __cplusplus +extern "C" { +#endif + #ifdef GGML_USE_HIPBLAS #define GGML_CUDA_NAME "ROCm" #define GGML_CUBLAS_NAME "hipBLAS" @@ -13,35 +17,31 @@ #define GGML_CUDA_NAME "CUDA" #define GGML_CUBLAS_NAME "cuBLAS" #endif - -#ifdef __cplusplus -extern "C" { -#endif - #define GGML_CUDA_MAX_DEVICES 16 // backend API -GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device); +GGML_API ggml_backend_t ggml_backend_cuda_init(int device); -GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend); +GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend); // device buffer -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device); +GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device); // split tensor buffer that splits matrices by rows across multiple devices -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split); +GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split); // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void); +GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void); -GGML_API GGML_CALL int ggml_backend_cuda_get_device_count(void); -GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size); -GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total); +GGML_API int ggml_backend_cuda_get_device_count(void); +GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size); +GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total); -GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size); -GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer); +GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size); +GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer); + +GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void); -GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data); #ifdef __cplusplus } #endif diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h index d483cf1ac..b8d3f678b 100644 --- a/ggml/include/ggml-metal.h +++ b/ggml/include/ggml-metal.h @@ -1,3 +1,5 @@ +// Note: this description is outdated +// // An interface allowing to compute ggml_cgraph with Metal // // This is a fully functional interface that extends ggml with GPU support for Apple devices. @@ -25,9 +27,6 @@ #include #include -// max memory buffers that can be mapped to the device -#define GGML_METAL_MAX_BUFFERS 64 - struct ggml_tensor; struct ggml_cgraph; @@ -40,19 +39,17 @@ extern "C" { // user-code should use only these functions // -GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data); - GGML_API ggml_backend_t ggml_backend_metal_init(void); GGML_API bool ggml_backend_is_metal(ggml_backend_t backend); -GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size); - -GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb); +GGML_DEPRECATED( + GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size), + "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713"); GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data); -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void); +GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void); // helper to check if the device supports a specific family // ideally, the user code should be doing these checks @@ -62,6 +59,8 @@ GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int fam // capture all command buffers committed the next time `ggml_backend_graph_compute` is called GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend); +GGML_API ggml_backend_reg_t ggml_backend_metal_reg(void); + #ifdef __cplusplus } #endif diff --git a/ggml/include/ggml-rpc.h b/ggml/include/ggml-rpc.h index aa144832a..64cde7f13 100644 --- a/ggml/include/ggml-rpc.h +++ b/ggml/include/ggml-rpc.h @@ -10,14 +10,14 @@ extern "C" { #define GGML_RPC_MAX_SERVERS 16 // backend API -GGML_API GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint); -GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend); +GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint); +GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend); -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint); +GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint); -GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total); +GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total); -GGML_API GGML_CALL void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem); +GGML_API void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem); #ifdef __cplusplus } diff --git a/ggml/include/ggml-sycl.h b/ggml/include/ggml-sycl.h index 43ab1519c..03b698e61 100644 --- a/ggml/include/ggml-sycl.h +++ b/ggml/include/ggml-sycl.h @@ -23,20 +23,20 @@ GGML_API ggml_backend_t ggml_backend_sycl_init(int device); GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device); // split tensor buffer that splits matrices by rows across multiple devices -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split); +GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split); // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void); -GGML_API void ggml_backend_sycl_print_sycl_devices(void); -GGML_API GGML_CALL void ggml_sycl_get_gpu_list(int *id_list, int max_len); -GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description, size_t description_size); -GGML_API GGML_CALL int ggml_backend_sycl_get_device_count(); -GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total); +GGML_API void ggml_backend_sycl_print_sycl_devices(void); +GGML_API void ggml_sycl_get_gpu_list(int *id_list, int max_len); +GGML_API void ggml_sycl_get_device_description(int device, char *description, size_t description_size); +GGML_API int ggml_backend_sycl_get_device_count(); +GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total); // SYCL doesn't support registering host memory, keep here for reference -// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size); -// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer); +// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size); +// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer); #ifdef __cplusplus } #endif diff --git a/ggml/include/ggml-vulkan.h b/ggml/include/ggml-vulkan.h index af661c2d7..e074042ef 100644 --- a/ggml/include/ggml-vulkan.h +++ b/ggml/include/ggml-vulkan.h @@ -13,16 +13,16 @@ extern "C" { GGML_API void ggml_vk_instance_init(void); // backend API -GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num); +GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num); -GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend); -GGML_API GGML_CALL int ggml_backend_vk_get_device_count(void); -GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size); -GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total); +GGML_API bool ggml_backend_is_vk(ggml_backend_t backend); +GGML_API int ggml_backend_vk_get_device_count(void); +GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size); +GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total); -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num); +GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num); // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void); +GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void); #ifdef __cplusplus } diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index f46d4a8a6..e7678d071 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -187,16 +187,6 @@ # define GGML_API #endif -#ifdef GGML_MULTIPLATFORM -# if defined(_WIN32) -# define GGML_CALL -# else -# define GGML_CALL __attribute__((__ms_abi__)) -# endif -#else -# define GGML_CALL -#endif - // TODO: support for clang #ifdef __GNUC__ # define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint))) @@ -340,7 +330,7 @@ extern "C" { }; // get ggml_status name string - GGML_API GGML_CALL const char * ggml_status_to_string(enum ggml_status status); + GGML_API const char * ggml_status_to_string(enum ggml_status status); // ieee 754-2008 half-precision float16 // todo: make this not an integral type @@ -466,6 +456,7 @@ extern "C" { GGML_OP_SUM_ROWS, GGML_OP_MEAN, GGML_OP_ARGMAX, + GGML_OP_COUNT_EQUAL, GGML_OP_REPEAT, GGML_OP_REPEAT_BACK, GGML_OP_CONCAT, @@ -577,10 +568,10 @@ extern "C" { // this tensor... enum ggml_tensor_flag { - GGML_TENSOR_FLAG_INPUT = 1, // ...is an input for the GGML compute graph - GGML_TENSOR_FLAG_OUTPUT = 2, // ...is an output for the GGML compute graph - GGML_TENSOR_FLAG_PARAM = 4, // ...contains trainable parameters - GGML_TENSOR_FLAG_LOSS = 8, // ...defines loss for numerical optimization (multiple loss tensors add up) + GGML_TENSOR_FLAG_INPUT = 1, // ...is an input for the GGML compute graph + GGML_TENSOR_FLAG_OUTPUT = 2, // ...is an output for the GGML compute graph + GGML_TENSOR_FLAG_PARAM = 4, // ...contains trainable parameters + GGML_TENSOR_FLAG_LOSS = 8, // ...defines loss for numerical optimization (multiple loss tensors add up) }; // n-dimensional tensor @@ -716,46 +707,46 @@ extern "C" { GGML_API void ggml_print_object (const struct ggml_object * obj); GGML_API void ggml_print_objects(const struct ggml_context * ctx); - GGML_API GGML_CALL int64_t ggml_nelements (const struct ggml_tensor * tensor); - GGML_API GGML_CALL int64_t ggml_nrows (const struct ggml_tensor * tensor); - GGML_API GGML_CALL size_t ggml_nbytes (const struct ggml_tensor * tensor); - GGML_API size_t ggml_nbytes_pad (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN + GGML_API int64_t ggml_nelements (const struct ggml_tensor * tensor); + GGML_API int64_t ggml_nrows (const struct ggml_tensor * tensor); + GGML_API size_t ggml_nbytes (const struct ggml_tensor * tensor); + GGML_API size_t ggml_nbytes_pad(const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN - GGML_API GGML_CALL int64_t ggml_blck_size(enum ggml_type type); - GGML_API GGML_CALL size_t ggml_type_size(enum ggml_type type); // size in bytes for all elements in a block - GGML_API GGML_CALL size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row + GGML_API int64_t ggml_blck_size(enum ggml_type type); + GGML_API size_t ggml_type_size(enum ggml_type type); // size in bytes for all elements in a block + GGML_API size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row GGML_DEPRECATED( GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float "use ggml_row_size() instead"); - GGML_API GGML_CALL const char * ggml_type_name(enum ggml_type type); - GGML_API GGML_CALL const char * ggml_op_name (enum ggml_op op); - GGML_API const char * ggml_op_symbol(enum ggml_op op); + GGML_API const char * ggml_type_name(enum ggml_type type); + GGML_API const char * ggml_op_name (enum ggml_op op); + GGML_API const char * ggml_op_symbol(enum ggml_op op); - GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op); - GGML_API GGML_CALL const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name + GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op); + GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name - GGML_API GGML_CALL size_t ggml_element_size(const struct ggml_tensor * tensor); + GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor); - GGML_API GGML_CALL bool ggml_is_quantized(enum ggml_type type); + GGML_API bool ggml_is_quantized(enum ggml_type type); // TODO: temporary until model loading of ggml examples is refactored GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype); - GGML_API GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor); - GGML_API GGML_CALL bool ggml_is_permuted (const struct ggml_tensor * tensor); - GGML_API GGML_CALL bool ggml_is_empty (const struct ggml_tensor * tensor); - GGML_API bool ggml_is_scalar (const struct ggml_tensor * tensor); - GGML_API bool ggml_is_vector (const struct ggml_tensor * tensor); - GGML_API bool ggml_is_matrix (const struct ggml_tensor * tensor); - GGML_API bool ggml_is_3d (const struct ggml_tensor * tensor); - GGML_API int ggml_n_dims (const struct ggml_tensor * tensor); // returns 1 for scalars + GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor); + GGML_API bool ggml_is_permuted (const struct ggml_tensor * tensor); + GGML_API bool ggml_is_empty (const struct ggml_tensor * tensor); + GGML_API bool ggml_is_scalar (const struct ggml_tensor * tensor); + GGML_API bool ggml_is_vector (const struct ggml_tensor * tensor); + GGML_API bool ggml_is_matrix (const struct ggml_tensor * tensor); + GGML_API bool ggml_is_3d (const struct ggml_tensor * tensor); + GGML_API int ggml_n_dims (const struct ggml_tensor * tensor); // returns 1 for scalars - GGML_API GGML_CALL bool ggml_is_contiguous (const struct ggml_tensor * tensor); - GGML_API GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous() - GGML_API GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1 - GGML_API GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2 + GGML_API bool ggml_is_contiguous (const struct ggml_tensor * tensor); + GGML_API bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous() + GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1 + GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2 GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1); GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1); @@ -847,7 +838,7 @@ extern "C" { GGML_API void * ggml_get_data (const struct ggml_tensor * tensor); GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor); - GGML_API GGML_CALL enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor); + GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor); GGML_API const char * ggml_get_name (const struct ggml_tensor * tensor); GGML_API struct ggml_tensor * ggml_set_name ( struct ggml_tensor * tensor, const char * name); @@ -1004,6 +995,12 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a); + // count number of equal elements in a and b + GGML_API struct ggml_tensor * ggml_count_equal( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + // if a is the same shape as b, and a is not parameter, return a // otherwise, return a new tensor: repeat(a) to fit in b GGML_API struct ggml_tensor * ggml_repeat( @@ -1410,14 +1407,14 @@ extern "C" { // supports 3D: a->ne[2] == b->ne[1] GGML_API struct ggml_tensor * ggml_get_rows( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); + struct ggml_tensor * a, // data + struct ggml_tensor * b); // row indices GGML_API struct ggml_tensor * ggml_get_rows_back( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c); + struct ggml_tensor * a, // gradients of ggml_get_rows result + struct ggml_tensor * b, // row indices + struct ggml_tensor * c); // data for ggml_get_rows, only used for its shape GGML_API struct ggml_tensor * ggml_diag( struct ggml_context * ctx, @@ -1561,16 +1558,16 @@ extern "C" { "use ggml_rope_ext_inplace instead"); // compute correction dims for YaRN RoPE scaling - GGML_CALL void ggml_rope_yarn_corr_dims( + void ggml_rope_yarn_corr_dims( int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]); // rotary position embedding backward, i.e compute dx from dy // a - dy GGML_API struct ggml_tensor * ggml_rope_back( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, + struct ggml_tensor * a, // gradients of ggml_rope result + struct ggml_tensor * b, // positions + struct ggml_tensor * c, // freq factors int n_dims, int mode, int n_ctx_orig, @@ -2036,15 +2033,15 @@ extern "C" { // loss function GGML_API struct ggml_tensor * ggml_cross_entropy_loss( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); + struct ggml_context * ctx, + struct ggml_tensor * a, // logits + struct ggml_tensor * b); // labels GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c); + struct ggml_context * ctx, + struct ggml_tensor * a, // logits + struct ggml_tensor * b, // labels + struct ggml_tensor * c); // gradients of cross_entropy_loss result // AdamW optimizer step // Paper: https://arxiv.org/pdf/1711.05101v3.pdf @@ -2052,6 +2049,7 @@ extern "C" { GGML_API struct ggml_tensor * ggml_opt_step_adamw( struct ggml_context * ctx, struct ggml_tensor * a, + struct ggml_tensor * grad, float alpha, float beta1, float beta2, @@ -2066,7 +2064,7 @@ extern "C" { GGML_API void ggml_set_loss(struct ggml_tensor * tensor); GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor); - GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate, bool keep); + GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate); GGML_API void ggml_build_opt_adamw( struct ggml_context * ctx, @@ -2176,6 +2174,10 @@ extern "C" { typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel); typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data); + // Set callback for all future logging events. + // If this is not called, or NULL is supplied, everything is output on stderr. + GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data); + // optimization parameters // // see ggml.c (ggml_opt_default_params) for default values diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index cbc349500..286bec255 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -511,8 +511,8 @@ if (GGML_HIPBLAS) endif() if (GGML_SYCL) - if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA)$") - message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL or NVIDIA") + if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$") + message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD") endif() check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL) @@ -532,6 +532,9 @@ if (GGML_SYCL) list(APPEND GGML_CDEF_PUBLIC GGML_USE_SYCL) if (GGML_SYCL_F16) + if (GGML_SYCL_TARGET STREQUAL "AMD") + message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.") + endif() add_compile_definitions(GGML_SYCL_F16) endif() @@ -543,6 +546,12 @@ if (GGML_SYCL) if (GGML_SYCL_TARGET STREQUAL "NVIDIA") add_compile_definitions(GGML_SYCL_WARP_SIZE=32) + elseif (GGML_SYCL_TARGET STREQUAL "AMD") + # INFO: Allowed Sub_group_sizes are not consistent through all + # hip targets. For example, 64 is used for certain models, but the backend + # does not support it. + # Target archs tested working: gfx1030, gfx1031, (Only tested sub_group_size = 32) + add_compile_definitions(GGML_SYCL_WARP_SIZE=32) else() add_compile_definitions(GGML_SYCL_WARP_SIZE=16) endif() @@ -576,6 +585,12 @@ if (GGML_SYCL) elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda") list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl pthread m dl onemkl) + elseif (GGML_SYCL_TARGET STREQUAL "AMD") + if (GGML_SYCL_HIP_TARGET STREQUAL "") + message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_HIP_TARGET has not been set.") + endif() + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=${GGML_SYCL_HIP_TARGET}") + list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl pthread m dl onemkl) endif() endif() endif() @@ -1310,7 +1325,7 @@ add_library(ggml ../include/ggml-backend.h ggml.c ggml-alloc.c - ggml-backend.c + ggml-backend.cpp ggml-quants.c ggml-quants.h ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA} diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h index b0d4141cc..ba2e26999 100644 --- a/ggml/src/ggml-backend-impl.h +++ b/ggml/src/ggml-backend-impl.h @@ -9,145 +9,226 @@ extern "C" { #endif // - // Backend buffer + // Backend buffer type // - // buffer type - typedef void * ggml_backend_buffer_type_context_t; - struct ggml_backend_buffer_type_i { - const char * (*GGML_CALL get_name) (ggml_backend_buffer_type_t buft); + const char * (*get_name) (ggml_backend_buffer_type_t buft); // allocate a buffer of this type - ggml_backend_buffer_t (*GGML_CALL alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size); + ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size); // tensor alignment - size_t (*GGML_CALL get_alignment) (ggml_backend_buffer_type_t buft); - // max buffer size that can be allocated - size_t (*GGML_CALL get_max_size) (ggml_backend_buffer_type_t buft); - // data size needed to allocate the tensor, including padding - size_t (*GGML_CALL get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); - // check if tensor data is in host memory - bool (*GGML_CALL is_host) (ggml_backend_buffer_type_t buft); + size_t (*get_alignment) (ggml_backend_buffer_type_t buft); + // (optional) max buffer size that can be allocated (defaults to SIZE_MAX) + size_t (*get_max_size) (ggml_backend_buffer_type_t buft); + // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes) + size_t (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); + // (optional) check if tensor data is in host memory (defaults to false) + bool (*is_host) (ggml_backend_buffer_type_t buft); }; struct ggml_backend_buffer_type { struct ggml_backend_buffer_type_i iface; - ggml_backend_buffer_type_context_t context; + ggml_backend_dev_t device; + void * context; }; - // buffer - typedef void * ggml_backend_buffer_context_t; + // + // Backend buffer + // struct ggml_backend_buffer_i { - const char * (*GGML_CALL get_name) (ggml_backend_buffer_t buffer); - void (*GGML_CALL free_buffer) (ggml_backend_buffer_t buffer); - void * (*GGML_CALL get_base) (ggml_backend_buffer_t buffer); - void (*GGML_CALL init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - void (*GGML_CALL memset_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size); - void (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - void (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - bool (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer - void (*GGML_CALL clear) (ggml_backend_buffer_t buffer, uint8_t value); - void (*GGML_CALL reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras + const char * (*get_name) (ggml_backend_buffer_t buffer); + // (optional) free the buffer + void (*free_buffer) (ggml_backend_buffer_t buffer); + // base address of the buffer + void * (*get_base) (ggml_backend_buffer_t buffer); + // (optional) initialize a tensor in the buffer (eg. add tensor extras) + void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + // tensor data access + void (*memset_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size); + void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + // (optional) tensor copy: dst is in the buffer, src may be in any buffer, including buffers from a different backend (return false if not supported) + bool (*cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); + // clear the entire buffer + void (*clear) (ggml_backend_buffer_t buffer, uint8_t value); + // (optional) reset any internal state due to tensor initialization, such as tensor extras + void (*reset) (ggml_backend_buffer_t buffer); }; struct ggml_backend_buffer { struct ggml_backend_buffer_i iface; ggml_backend_buffer_type_t buft; - ggml_backend_buffer_context_t context; + void * context; size_t size; enum ggml_backend_buffer_usage usage; }; - GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init( - ggml_backend_buffer_type_t buft, - struct ggml_backend_buffer_i iface, - ggml_backend_buffer_context_t context, - size_t size); + ggml_backend_buffer_t ggml_backend_buffer_init( + ggml_backend_buffer_type_t buft, + struct ggml_backend_buffer_i iface, + void * context, + size_t size); // do not use directly, use ggml_backend_tensor_copy instead bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst); + // multi-buffer // buffer that contains a collection of buffers - GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers); - GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer); - GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage); + ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers); + bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer); + void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage); // - // Backend + // Backend (stream) // - typedef void * ggml_backend_context_t; - struct ggml_backend_i { - const char * (*GGML_CALL get_name)(ggml_backend_t backend); + const char * (*get_name)(ggml_backend_t backend); - void (*GGML_CALL free)(ggml_backend_t backend); + void (*free)(ggml_backend_t backend); // buffer allocation - ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend); + ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend); // (optional) asynchronous tensor data access - void (*GGML_CALL set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst); + void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst); // (optional) complete all pending operations - void (*GGML_CALL synchronize)(ggml_backend_t backend); + void (*synchronize)(ggml_backend_t backend); - // compute graph with a plan (not used currently) - // create a new plan for a graph - ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph); - void (*GGML_CALL graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); + // (optional) compute graph with a plan (not used currently) + ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph); + void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology - void (*GGML_CALL graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph); + void (*graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph); // compute the graph with the plan - enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); + enum ggml_status (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); - // compute graph without a plan (async) - enum ggml_status (*GGML_CALL graph_compute) (ggml_backend_t backend, struct ggml_cgraph * cgraph); + // compute graph (always async if supported by the backend) + enum ggml_status (*graph_compute) (ggml_backend_t backend, struct ggml_cgraph * cgraph); + // IMPORTANT: these functions have been moved to the device interface and will be removed from the backend interface + // new backends should implement the device interface instead + + // These functions are being moved to the device interface // check if the backend can compute an operation - bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op); + bool (*supports_op) (ggml_backend_t backend, const struct ggml_tensor * op); // check if the backend can use tensors allocated in a buffer type - bool (*GGML_CALL supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft); + bool (*supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft); // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer // these should be expensive operations with large batch sizes that may benefit from running on this backend // even if the weight has to be copied from the CPU temporarily - bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op); + bool (*offload_op) (ggml_backend_t backend, const struct ggml_tensor * op); // (optional) event synchronization - // create a new event that can record events on this backend instance - ggml_backend_event_t (*GGML_CALL event_new) (ggml_backend_t backend); - void (*GGML_CALL event_free) (ggml_backend_event_t event); - // record an event on the backend instance that created it - void (*GGML_CALL event_record) (ggml_backend_event_t event); - // wait for an event on on a different backend instance - void (*GGML_CALL event_wait) (ggml_backend_t backend, ggml_backend_event_t event); - // block until an event is recorded - void (*GGML_CALL event_synchronize) (ggml_backend_event_t event); + // record an event on this stream + void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event); + // wait for an event on on a different stream + void (*event_wait) (ggml_backend_t backend, ggml_backend_event_t event); }; struct ggml_backend { ggml_guid_t guid; - struct ggml_backend_i iface; - ggml_backend_context_t context; + ggml_backend_dev_t device; + void * context; }; struct ggml_backend_event { - ggml_backend_t backend; + struct ggml_backend_device * device; void * context; }; // - // Backend registry + // Backend device // - typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data); + // Note: if additional properties are needed, we should add a struct with all of them + // the current functions to obtain the properties can remain, since they are more convenient for often used properties + struct ggml_backend_device_i { + // device name: short identifier for this device, such as "CPU" or "CUDA0" + const char * (*get_name)(ggml_backend_dev_t dev); - GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data); + // device description: short informative description of the device, could be the model name + const char * (*get_description)(ggml_backend_dev_t dev); + + // device memory in bytes + void (*get_memory)(ggml_backend_dev_t dev, size_t * free, size_t * total); + + // device type + enum ggml_backend_dev_type (*get_type)(ggml_backend_dev_t dev); + + // device properties + void (*get_props)(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props); + + // backend (stream) initialization + ggml_backend_t (*init_backend)(ggml_backend_dev_t dev, const char * params); + + // preferred buffer type + ggml_backend_buffer_type_t (*get_buffer_type)(ggml_backend_dev_t dev); + + // (optional) host buffer type (in system memory, typically this is a pinned memory buffer for faster transfers between host and device) + ggml_backend_buffer_type_t (*get_host_buffer_type)(ggml_backend_dev_t dev); + + // (optional) buffer from pointer: create a buffer from a host pointer (useful for memory mapped models and importing data from other libraries) + ggml_backend_buffer_t (*buffer_from_host_ptr)(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size); + + // check if the backend can compute an operation + bool (*supports_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op); + + // check if the backend can use tensors allocated in a buffer type + bool (*supports_buft)(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft); + + // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer + // these should be expensive operations with large batch sizes that may benefit from running on this backend + // even if the weight has to be copied from the CPU temporarily + bool (*offload_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op); + + // (optional) event synchronization + ggml_backend_event_t (*event_new) (ggml_backend_dev_t dev); + void (*event_free) (ggml_backend_dev_t dev, ggml_backend_event_t event); + void (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event); + }; + + struct ggml_backend_device { + struct ggml_backend_device_i iface; + ggml_backend_reg_t reg; + void * context; + }; + + // + // Backend (reg) + // + + struct ggml_backend_reg_i { + const char * (*get_name)(ggml_backend_reg_t reg); + + // enumerate available devices + size_t (*get_device_count)(ggml_backend_reg_t reg); + ggml_backend_dev_t (*get_device)(ggml_backend_reg_t reg, size_t index); + + // (optional) get a pointer to a function in the backend + // backends can add custom functions that are not part of the standard ggml-backend interface + void * (*get_proc_address)(ggml_backend_reg_t reg, const char * name); + }; + + struct ggml_backend_reg { + // int api_version; // TODO: for dynamic loading + struct ggml_backend_reg_i iface; + void * context; + }; + + + // Internal backend registry API + void ggml_backend_register(ggml_backend_reg_t reg); + void ggml_backend_device_register(ggml_backend_dev_t device); + // TODO: backends can be loaded as a dynamic library, in which case it needs to export this function + // typedef ggml_backend_register_t * (*ggml_backend_init)(void); #ifdef __cplusplus } diff --git a/ggml/src/ggml-backend.c b/ggml/src/ggml-backend.cpp similarity index 75% rename from ggml/src/ggml-backend.c rename to ggml/src/ggml-backend.cpp index ba280e064..4f3e9374c 100644 --- a/ggml/src/ggml-backend.c +++ b/ggml/src/ggml-backend.cpp @@ -1,3 +1,13 @@ +// Note: porting this file to C++ is a work in progress + +#ifdef _WIN32 +#define WIN32_LEAN_AND_MEAN +#ifndef NOMINMAX +# define NOMINMAX +#endif +#include +#endif + #include "ggml-backend-impl.h" #include "ggml-alloc.h" #include "ggml-impl.h" @@ -8,9 +18,14 @@ #include #include #include +#include +#include +#ifdef __APPLE__ +#include +#include +#endif -#define MAX(a, b) ((a) > (b) ? (a) : (b)) // backend buffer type @@ -18,7 +33,7 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) { return buft->iface.get_name(buft); } -GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { return buft->iface.alloc_buffer(buft, size); } @@ -34,7 +49,7 @@ size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) { return SIZE_MAX; } -GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) { +size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) { // get_alloc_size is optional, defaults to ggml_nbytes if (buft->iface.get_alloc_size) { size_t size = buft->iface.get_alloc_size(buft, tensor); @@ -51,16 +66,18 @@ bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) { return false; } +ggml_backend_dev_t ggml_backend_buft_get_device(ggml_backend_buffer_type_t buft) { + return buft->device; +} + // backend buffer -GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init( - ggml_backend_buffer_type_t buft, - struct ggml_backend_buffer_i iface, - ggml_backend_buffer_context_t context, - size_t size) { - ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer)); - - (*buffer) = (struct ggml_backend_buffer) { +ggml_backend_buffer_t ggml_backend_buffer_init( + ggml_backend_buffer_type_t buft, + struct ggml_backend_buffer_i iface, + void * context, + size_t size) { + ggml_backend_buffer_t buffer = new ggml_backend_buffer { /* .interface = */ iface, /* .buft = */ buft, /* .context = */ context, @@ -83,7 +100,7 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) { if (buffer->iface.free_buffer != NULL) { buffer->iface.free_buffer(buffer); } - free(buffer); + delete buffer; } size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) { @@ -98,14 +115,14 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) { return base; } -GGML_CALL void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { +void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { // init_tensor is optional if (buffer->iface.init_tensor) { buffer->iface.init_tensor(buffer, tensor); } } -size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) { +size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) { return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer)); } @@ -218,7 +235,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten } } -GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; GGML_ASSERT(buf != NULL && "tensor buffer not set"); @@ -232,7 +249,7 @@ GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * buf->iface.set_tensor(buf, tensor, data, offset, size); } -GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { +void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; GGML_ASSERT(buf != NULL && "tensor buffer not set"); @@ -246,7 +263,7 @@ GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * buf->iface.get_tensor(buf, tensor, data, offset, size); } -GGML_API GGML_CALL void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { +GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; GGML_ASSERT(buf != NULL && "tensor buffer not set"); @@ -299,20 +316,39 @@ enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct } bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { + // helper to ease transition to device interface + if (backend->device) { + return ggml_backend_dev_supports_op(backend->device, op); + } + return backend->iface.supports_op(backend, op); } bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { + // helper to ease transition to device interface + if (backend->device) { + return ggml_backend_dev_supports_buft(backend->device, buft); + } + return backend->iface.supports_buft(backend, buft); } bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) { + // helper to ease transition to device interface + if (backend->device) { + return ggml_backend_dev_offload_op(backend->device, op); + } + if (backend->iface.offload_op != NULL) { return backend->iface.offload_op(backend, op); } return false; } +ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) { + return backend->device; +} + // backend copy static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) { @@ -375,30 +411,31 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b // events -ggml_backend_event_t ggml_backend_event_new(ggml_backend_t backend) { - if (backend->iface.event_new == NULL) { +ggml_backend_event_t ggml_backend_event_new(ggml_backend_dev_t device) { + // null device is allowed for the transition period to the device interface + if (device == NULL || device->iface.event_new == NULL) { return NULL; } - return backend->iface.event_new(backend); + return device->iface.event_new(device); } void ggml_backend_event_free(ggml_backend_event_t event) { if (event == NULL) { return; } - event->backend->iface.event_free(event); + event->device->iface.event_free(event->device, event); } -void ggml_backend_event_record(ggml_backend_event_t event) { - GGML_ASSERT(event->backend->iface.event_record != NULL); +void ggml_backend_event_record(ggml_backend_event_t event, ggml_backend_t backend) { + GGML_ASSERT(backend->iface.event_record != NULL); - event->backend->iface.event_record(event); + backend->iface.event_record(backend, event); } void ggml_backend_event_synchronize(ggml_backend_event_t event) { - GGML_ASSERT(event->backend->iface.event_synchronize != NULL); + GGML_ASSERT(event->device->iface.event_synchronize); - event->backend->iface.event_synchronize(event); + event->device->iface.event_synchronize(event->device, event); } void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) { @@ -407,170 +444,235 @@ void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) backend->iface.event_wait(backend, event); } -// backend registry +// Backend device -#define GGML_REG_MAX_BACKENDS 64 - -struct ggml_backend_reg { - char name[128]; - ggml_backend_init_fn init_fn; - ggml_backend_buffer_type_t default_buffer_type; - void * user_data; -}; - -static struct ggml_backend_reg ggml_backend_registry[GGML_REG_MAX_BACKENDS]; -static size_t ggml_backend_registry_count = 0; - -GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data); - -GGML_CALL static void ggml_backend_registry_init(void) { - static bool initialized = false; - - if (initialized) { - return; - } - - initialized = true; - - ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL); - - // add forward decls here to avoid including the backend headers -#ifdef GGML_USE_CUDA - extern GGML_CALL void ggml_backend_cuda_reg_devices(void); - ggml_backend_cuda_reg_devices(); -#endif - -#ifdef GGML_USE_SYCL - extern void ggml_backend_sycl_reg_devices(void); - ggml_backend_sycl_reg_devices(); -#endif - -#ifdef GGML_USE_METAL - extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); - extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void); - ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL); -#endif - -#ifdef GGML_USE_VULKAN - extern GGML_CALL int ggml_backend_vk_reg_devices(void); - ggml_backend_vk_reg_devices(); -#endif - -#ifdef GGML_USE_KOMPUTE - extern GGML_CALL void ggml_backend_kompute_reg_devices(void); - ggml_backend_kompute_reg_devices(); -#endif - -#ifdef GGML_USE_CANN - extern GGML_CALL int ggml_backend_cann_reg_devices(void); - ggml_backend_cann_reg_devices(); -#endif +const char * ggml_backend_dev_name(ggml_backend_dev_t device) { + return device->iface.get_name(device); } -GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) { - GGML_ASSERT(ggml_backend_registry_count < GGML_REG_MAX_BACKENDS); - - size_t id = ggml_backend_registry_count; - - ggml_backend_registry[id] = (struct ggml_backend_reg) { - /* .name = */ {0}, - /* .fn = */ init_fn, - /* .default_buffer_type = */ default_buffer_type, - /* .user_data = */ user_data, - }; - - snprintf(ggml_backend_registry[id].name, sizeof(ggml_backend_registry[id].name), "%s", name); - -#ifndef NDEBUG - fprintf(stderr, "%s: registered backend %s\n", __func__, name); -#endif - - ggml_backend_registry_count++; +const char * ggml_backend_dev_description(ggml_backend_dev_t device) { + return device->iface.get_description(device); } -size_t ggml_backend_reg_get_count(void) { - ggml_backend_registry_init(); - - return ggml_backend_registry_count; +void ggml_backend_dev_memory(ggml_backend_dev_t device, size_t * free, size_t * total) { + device->iface.get_memory(device, free, total); } -size_t ggml_backend_reg_find_by_name(const char * name) { - ggml_backend_registry_init(); - - for (size_t i = 0; i < ggml_backend_registry_count; i++) { - // TODO: case insensitive in a portable way - if (strcmp(ggml_backend_registry[i].name, name) == 0) { - return i; - } - } - - // not found - return SIZE_MAX; +enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) { + return device->iface.get_type(device); } -// init from backend:params string -ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) { - ggml_backend_registry_init(); +void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props) { + memset(props, 0, sizeof(*props)); + device->iface.get_props(device, props); +} - const char * params = strchr(backend_str, ':'); - char backend_name[128]; - if (params == NULL) { - snprintf(backend_name, sizeof(backend_name), "%s", backend_str); - params = ""; - } else { - snprintf(backend_name, sizeof(backend_name), "%.*s", (int)(params - backend_str), backend_str); - params++; - } +ggml_backend_reg_t ggml_backend_dev_backend_reg(ggml_backend_dev_t device) { + return device->reg; +} - size_t backend_i = ggml_backend_reg_find_by_name(backend_name); +ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * params) { + return device->iface.init_backend(device, params); +} - if (backend_i == SIZE_MAX) { - fprintf(stderr, "%s: backend %s not found\n", __func__, backend_name); +ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) { + return device->iface.get_buffer_type(device); +} + +ggml_backend_buffer_type_t ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device) { + if (device->iface.get_host_buffer_type == NULL) { return NULL; } - return ggml_backend_reg_init_backend(backend_i, params); + return device->iface.get_host_buffer_type(device); } -const char * ggml_backend_reg_get_name(size_t i) { - ggml_backend_registry_init(); - - GGML_ASSERT(i < ggml_backend_registry_count); - return ggml_backend_registry[i].name; +ggml_backend_buffer_t ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size) { + return device->iface.buffer_from_host_ptr(device, ptr, size, max_tensor_size); } -ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params) { - ggml_backend_registry_init(); - - GGML_ASSERT(i < ggml_backend_registry_count); - return ggml_backend_registry[i].init_fn(params, ggml_backend_registry[i].user_data); +bool ggml_backend_dev_supports_op(ggml_backend_dev_t device, const struct ggml_tensor * op) { + return device->iface.supports_op(device, op); } -ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i) { - ggml_backend_registry_init(); - - GGML_ASSERT(i < ggml_backend_registry_count); - return ggml_backend_registry[i].default_buffer_type; +bool ggml_backend_dev_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft) { + return device->iface.supports_buft(device, buft); } -ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) { - ggml_backend_registry_init(); +bool ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_tensor * op) { + return device->iface.offload_op(device, op); +} - GGML_ASSERT(i < ggml_backend_registry_count); - return ggml_backend_buft_alloc_buffer(ggml_backend_registry[i].default_buffer_type, size); +// Backend (reg) + +const char * ggml_backend_reg_name(ggml_backend_reg_t reg) { + return reg->iface.get_name(reg); +} + +size_t ggml_backend_reg_dev_count(ggml_backend_reg_t reg) { + return reg->iface.get_device_count(reg); +} + +ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index) { + return reg->iface.get_device(reg, index); +} + +void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) { + if (!reg->iface.get_proc_address) { + return NULL; + } + return reg->iface.get_proc_address(reg, name); +} + +// Backend registry + +#ifdef GGML_USE_CUDA +#include "ggml-cuda.h" +#endif + +#ifdef GGML_USE_METAL +#include "ggml-metal.h" +#endif + +struct ggml_backend_registry { + std::vector backends; + std::vector devices; + + ggml_backend_registry() { +#ifdef GGML_USE_CUDA + register_backend(ggml_backend_cuda_reg()); +#endif +#ifdef GGML_USE_METAL + register_backend(ggml_backend_metal_reg()); +#endif + + register_backend(ggml_backend_cpu_reg()); + + // TODO: sycl, vulkan, kompute, cann + } + + void register_backend(ggml_backend_reg_t reg) { +#ifndef NDEBUG + fprintf(stderr, "%s: registered backend %s (%zu devices)\n", + __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg)); +#endif + backends.push_back(reg); + for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) { + register_device(ggml_backend_reg_dev_get(reg, i)); + } + } + + void register_device(ggml_backend_dev_t device) { +#ifndef NDEBUG + fprintf(stderr, "%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device)); +#endif + devices.push_back(device); + } +}; + +static ggml_backend_registry & get_reg() { + static ggml_backend_registry reg; + return reg; +} + +// Internal API +void ggml_backend_register(ggml_backend_reg_t reg) { + get_reg().register_backend(reg); +} + +void ggml_backend_device_register(ggml_backend_dev_t device) { + get_reg().register_device(device); +} + +// Backend (reg) enumeration +size_t ggml_backend_reg_count() { + return get_reg().backends.size(); +} + +ggml_backend_reg_t ggml_backend_reg_get(size_t index) { + GGML_ASSERT(index < ggml_backend_reg_count()); + return get_reg().backends[index]; +} + +ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) { + for (size_t i = 0; i < ggml_backend_reg_count(); i++) { + ggml_backend_reg_t reg = ggml_backend_reg_get(i); + if (strcmp(ggml_backend_reg_name(reg), name) == 0) { + return reg; + } + } + return NULL; +} + +// Device enumeration +size_t ggml_backend_dev_count() { + return get_reg().devices.size(); +} + +ggml_backend_dev_t ggml_backend_dev_get(size_t index) { + GGML_ASSERT(index < ggml_backend_dev_count()); + return get_reg().devices[index]; +} + +ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) { + for (size_t i = 0; i < ggml_backend_dev_count(); i++) { + ggml_backend_dev_t dev = ggml_backend_dev_get(i); + if (strcmp(ggml_backend_dev_name(dev), name) == 0) { + return dev; + } + } + return NULL; +} + +ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) { + for (size_t i = 0; i < ggml_backend_dev_count(); i++) { + ggml_backend_dev_t dev = ggml_backend_dev_get(i); + if (ggml_backend_dev_type(dev) == type) { + return dev; + } + } + return NULL; +} + +// Convenience functions +ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) { + ggml_backend_dev_t dev = ggml_backend_dev_by_name(name); + if (!dev) { + return NULL; + } + return ggml_backend_dev_init(dev, params); +} + +ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) { + ggml_backend_dev_t dev = ggml_backend_dev_by_type(type); + if (!dev) { + return NULL; + } + return ggml_backend_dev_init(dev, params); +} + +ggml_backend_t ggml_backend_init_best(void) { + ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL); + if (!dev) { + dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU_FULL); + } + if (!dev) { + return NULL; + } + return ggml_backend_dev_init(dev, NULL); } // backend CPU static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment -GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_cpu_buffer_get_name(ggml_backend_buffer_t buffer) { return "CPU"; GGML_UNUSED(buffer); } -GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) { +static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) { uintptr_t data = (uintptr_t)buffer->context; // align the buffer @@ -581,29 +683,29 @@ GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t b return (void *)data; } -GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) { free(buffer->context); } -GGML_CALL static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { +static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { memset((char *)tensor->data + offset, value, size); GGML_UNUSED(buffer); } -GGML_CALL static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { memcpy((char *)tensor->data + offset, data, size); GGML_UNUSED(buffer); } -GGML_CALL static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { +static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { memcpy(data, (const char *)tensor->data + offset, size); GGML_UNUSED(buffer); } -GGML_CALL static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { +static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { if (ggml_backend_buffer_is_host(src->buffer)) { memcpy(dst->data, src->data, ggml_nbytes(src)); return true; @@ -613,12 +715,12 @@ GGML_CALL static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t b GGML_UNUSED(buffer); } -GGML_CALL static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { +static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { memset(buffer->context, value, buffer->size); } -static struct ggml_backend_buffer_i cpu_backend_buffer_i = { - /* .get_name = */ ggml_backend_cpu_buffer_name, +static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = { + /* .get_name = */ ggml_backend_cpu_buffer_get_name, /* .free_buffer = */ ggml_backend_cpu_buffer_free_buffer, /* .get_base = */ ggml_backend_cpu_buffer_get_base, /* .init_tensor = */ NULL, // no initialization required @@ -630,9 +732,8 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i = { /* .reset = */ NULL, }; -// for buffers from ptr, free is not called -static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = { - /* .get_name = */ ggml_backend_cpu_buffer_name, +static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = { + /* .get_name = */ ggml_backend_cpu_buffer_get_name, /* .free_buffer = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed /* .get_base = */ ggml_backend_cpu_buffer_get_base, /* .init_tensor = */ NULL, // no initialization required @@ -644,13 +745,13 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = { /* .reset = */ NULL, }; -GGML_CALL static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) { return "CPU"; GGML_UNUSED(buft); } -GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h) if (data == NULL) { @@ -658,24 +759,24 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer return NULL; } - return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size); + return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, size); } -GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return TENSOR_ALIGNMENT; GGML_UNUSED(buft); } -GGML_CALL static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) { +static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) { return true; GGML_UNUSED(buft); } -GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) { +ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) { static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = { - /* .iface = */ { + /* .iface = */ { /* .get_name = */ ggml_backend_cpu_buffer_type_get_name, /* .alloc_buffer = */ ggml_backend_cpu_buffer_type_alloc_buffer, /* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment, @@ -683,6 +784,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) { /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes /* .is_host = */ ggml_backend_cpu_buffer_type_is_host, }, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0), /* .context = */ NULL, }; @@ -695,23 +797,23 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) { #include -GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) { return "CPU_HBM"; GGML_UNUSED(buft); } -GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) { +static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) { return "CPU_HBM"; GGML_UNUSED(buf); } -GGML_CALL static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) { hbw_free(buffer->context); } -GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { //void * ptr = hbw_malloc(size); void * ptr; int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size); @@ -749,27 +851,27 @@ struct ggml_backend_cpu_context { int n_threads; ggml_threadpool_t threadpool; - void * work_data; + uint8_t * work_data; size_t work_size; ggml_abort_callback abort_callback; void * abort_callback_data; }; -GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) { +static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) { return "CPU"; GGML_UNUSED(backend); } -GGML_CALL static void ggml_backend_cpu_free(ggml_backend_t backend) { +static void ggml_backend_cpu_free(ggml_backend_t backend) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; - free(cpu_ctx->work_data); - free(cpu_ctx); - free(backend); + delete[] cpu_ctx->work_data; + delete cpu_ctx; + delete backend; } -GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) { +static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) { return ggml_backend_cpu_buffer_type(); GGML_UNUSED(backend); @@ -780,18 +882,18 @@ struct ggml_backend_plan_cpu { struct ggml_cgraph cgraph; }; -GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) { +static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; - struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu)); + struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu; cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool); cpu_plan->cgraph = *cgraph; // FIXME: deep copy if (cpu_plan->cplan.work_size > 0) { - cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size); + cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size]; if (cpu_plan->cplan.work_data == NULL) { - free(cpu_plan); + delete cpu_plan; return NULL; } } @@ -802,16 +904,16 @@ GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(gg return cpu_plan; } -GGML_CALL static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { +static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan; - free(cpu_plan->cplan.work_data); - free(cpu_plan); + delete[] cpu_plan->cplan.work_data; + delete cpu_plan; GGML_UNUSED(backend); } -GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { +static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan; return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan); @@ -819,21 +921,21 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backe GGML_UNUSED(backend); } -GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool); if (cpu_ctx->work_size < cplan.work_size) { - free(cpu_ctx->work_data); - cpu_ctx->work_data = malloc(cplan.work_size); + delete[] cpu_ctx->work_data; + cpu_ctx->work_data = new uint8_t[cplan.work_size]; if (cpu_ctx->work_data == NULL) { cpu_ctx->work_size = 0; return GGML_STATUS_ALLOC_FAILED; } cpu_ctx->work_size = cplan.work_size; } - cplan.work_data = cpu_ctx->work_data; + cplan.work_data = (uint8_t *)cpu_ctx->work_data; cplan.abort_callback = cpu_ctx->abort_callback; cplan.abort_callback_data = cpu_ctx->abort_callback_data; @@ -841,35 +943,8 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t return ggml_graph_compute(cgraph, &cplan); } -GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { - switch (op->op) { - case GGML_OP_CPY: - return - op->type != GGML_TYPE_IQ2_XXS && - op->type != GGML_TYPE_IQ2_XS && - op->type != GGML_TYPE_IQ1_S && - op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float - case GGML_OP_MUL_MAT: - return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type; - case GGML_OP_ROPE_BACK: - return op->src[2] == NULL && (op->op_params[2] & 4) == 0; - case GGML_OP_IM2COL_BACK: - return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32; - default: - return true; - } - - GGML_UNUSED(backend); -} - -GGML_CALL static bool ggml_backend_cpu_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { - return ggml_backend_buft_is_host(buft); - - GGML_UNUSED(backend); -} - -static struct ggml_backend_i cpu_backend_i = { - /* .get_name = */ ggml_backend_cpu_name, +static const struct ggml_backend_i ggml_backend_cpu_i = { + /* .get_name = */ ggml_backend_cpu_get_name, /* .free = */ ggml_backend_cpu_free, /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type, /* .set_tensor_async = */ NULL, @@ -881,14 +956,11 @@ static struct ggml_backend_i cpu_backend_i = { /* .graph_plan_update = */ NULL, /* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute, /* .graph_compute = */ ggml_backend_cpu_graph_compute, - /* .supports_op = */ ggml_backend_cpu_supports_op, - /* .supports_buft = */ ggml_backend_cpu_supports_buft, + /* .supports_op = */ NULL, + /* .supports_buft = */ NULL, /* .offload_op = */ NULL, - /* .event_new = */ NULL, - /* .event_free = */ NULL, /* .event_record = */ NULL, /* .event_wait = */ NULL, - /* .event_synchronize = */ NULL, }; static ggml_guid_t ggml_backend_cpu_guid(void) { @@ -897,7 +969,7 @@ static ggml_guid_t ggml_backend_cpu_guid(void) { } ggml_backend_t ggml_backend_cpu_init(void) { - struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context)); + struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context; if (ctx == NULL) { return NULL; } @@ -909,21 +981,22 @@ ggml_backend_t ggml_backend_cpu_init(void) { ctx->abort_callback = NULL; ctx->abort_callback_data = NULL; - ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend)); + ggml_backend_t cpu_backend = new ggml_backend { + /* .guid = */ ggml_backend_cpu_guid(), + /* .interface = */ ggml_backend_cpu_i, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0), + /* .context = */ ctx, + }; + if (cpu_backend == NULL) { - free(ctx); + delete ctx; return NULL; } - *cpu_backend = (struct ggml_backend) { - /* .guid = */ ggml_backend_cpu_guid(), - /* .interface = */ cpu_backend_i, - /* .context = */ ctx - }; return cpu_backend; } -GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) { +bool ggml_backend_is_cpu(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid()); } @@ -954,16 +1027,227 @@ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_ ctx->abort_callback_data = abort_callback_data; } -GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) { +ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) { GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned"); - return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size); + return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size); } -GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) { +//////////////////////// + +struct ggml_backend_cpu_device_context { + std::string description = "CPU"; + + ggml_backend_cpu_device_context() { +#ifdef __APPLE__ + size_t len = 0; + if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) { + description.resize(len); + sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT + } +#elif defined(__linux__) + FILE * f = fopen("/proc/cpuinfo", "r"); + if (f) { + char buf[1024]; + while (fgets(buf, sizeof(buf), f)) { + if (strncmp(buf, "model name", 10) == 0) { + char * p = strchr(buf, ':'); + if (p) { + p++; + while (std::isspace(*p)) { + p++; + } + while (std::isspace(p[strlen(p) - 1])) { + p[strlen(p) - 1] = '\0'; + } + description = p; + break; + } + } + } + fclose(f); + } +#elif defined(_WIN32) + HKEY hKey; + if (RegOpenKeyEx(HKEY_LOCAL_MACHINE, + TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"), + 0, + KEY_READ, + &hKey) == ERROR_SUCCESS) { + DWORD cpu_brand_size = 0; + if (RegQueryValueExA(hKey, + TEXT("ProcessorNameString"), + NULL, + NULL, + NULL, + &cpu_brand_size) == ERROR_SUCCESS) { + description.resize(cpu_brand_size); + if (RegQueryValueExA(hKey, + TEXT("ProcessorNameString"), + NULL, + NULL, + (LPBYTE)&description[0], // NOLINT + &cpu_brand_size) == ERROR_SUCCESS) { + if (description.find('\0') != std::string::npos) { + description.resize(description.find('\0')); + } + } + } + RegCloseKey(hKey); + } +#endif + } +}; + +static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) { + return "CPU"; + + GGML_UNUSED(dev); +} + +static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) { + struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context; + + return ctx->description.c_str(); +} + +static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { + // TODO + *free = 0; + *total = 0; + + GGML_UNUSED(dev); +} + +static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) { + return GGML_BACKEND_DEVICE_TYPE_CPU_FULL; + + GGML_UNUSED(dev); +} + +static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) { + props->name = ggml_backend_cpu_device_get_name(dev); + props->description = ggml_backend_cpu_device_get_description(dev); + props->type = ggml_backend_cpu_device_get_type(dev); + ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total); + props->caps = { + /* .async = */ false, + /* .host_buffer = */ false, + /* .buffer_from_host_ptr = */ true, + /* .events = */ false, + }; +} + +static ggml_backend_t ggml_backend_cpu_device_init(ggml_backend_dev_t dev, const char * params) { return ggml_backend_cpu_init(); + GGML_UNUSED(dev); GGML_UNUSED(params); - GGML_UNUSED(user_data); +} + +static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) { + return ggml_backend_cpu_buffer_type(); + + GGML_UNUSED(dev); +} + +static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) { + return ggml_backend_cpu_buffer_from_ptr(ptr, size); + + GGML_UNUSED(dev); + GGML_UNUSED(max_tensor_size); +} + +static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) { + switch (op->op) { + case GGML_OP_CPY: + return + op->type != GGML_TYPE_IQ2_XXS && + op->type != GGML_TYPE_IQ2_XS && + op->type != GGML_TYPE_IQ1_S && + op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float + case GGML_OP_MUL_MAT: + return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type; + case GGML_OP_ROPE_BACK: + return op->src[2] == NULL && (op->op_params[2] & 4) == 0; + case GGML_OP_IM2COL_BACK: + return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32; + case GGML_OP_OUT_PROD: + return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32; + default: + return true; + } + + GGML_UNUSED(dev); +} + +static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) { + return ggml_backend_buft_is_host(buft); + + GGML_UNUSED(dev); +} + +static const struct ggml_backend_device_i ggml_backend_cpu_device_i = { + /* .get_name = */ ggml_backend_cpu_device_get_name, + /* .get_description = */ ggml_backend_cpu_device_get_description, + /* .get_memory = */ ggml_backend_cpu_device_get_memory, + /* .get_type = */ ggml_backend_cpu_device_get_type, + /* .get_props = */ ggml_backend_cpu_device_get_props, + /* .init_backend = */ ggml_backend_cpu_device_init, + /* .get_buffer_type = */ ggml_backend_cpu_device_get_buffer_type, + /* .get_host_buffer_type = */ NULL, + /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_ptr, + /* .supports_op = */ ggml_backend_cpu_device_supports_op, + /* .supports_buft = */ ggml_backend_cpu_device_supports_buft, + /* .offload_op = */ NULL, + /* .event_new = */ NULL, + /* .event_free = */ NULL, + /* .event_synchronize = */ NULL, +}; + +//////////////////////// + +static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) { + return "CPU"; + + GGML_UNUSED(reg); +} + +static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) { + return 1; + + GGML_UNUSED(reg); +} + +static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) { + GGML_ASSERT(index == 0); + + static ggml_backend_cpu_device_context ctx; + static ggml_backend_device ggml_backend_cpu_device = { + /* .iface = */ ggml_backend_cpu_device_i, + /* .reg = */ reg, + /* .context = */ &ctx, + }; + + return &ggml_backend_cpu_device; + + GGML_UNUSED(reg); + GGML_UNUSED(index); +} + +static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = { + /* .get_name = */ ggml_backend_cpu_reg_get_name, + /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count, + /* .get_device = */ ggml_backend_cpu_reg_get_device, + /* .get_proc_address = */ NULL, +}; + +ggml_backend_reg_t ggml_backend_cpu_reg(void) { + static struct ggml_backend_reg ggml_backend_cpu_reg = { + /* .iface = */ ggml_backend_cpu_reg_i, + /* .context = */ NULL, + }; + + return &ggml_backend_cpu_reg; } // multi-buffer buffer @@ -973,16 +1257,14 @@ struct ggml_backend_multi_buffer_context { size_t n_buffers; }; -typedef struct ggml_backend_multi_buffer_context * ggml_backend_multi_buffer_context_t; - -GGML_CALL static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) { - ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context; +static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) { + ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context; return ctx->buffers[0]->iface.get_name(ctx->buffers[0]); } -GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) { - ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context; +static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) { + ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context; for (size_t i = 0; i < ctx->n_buffers; i++) { ggml_backend_buffer_free(ctx->buffers[i]); } @@ -991,32 +1273,28 @@ GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_ free(ctx); } -GGML_CALL static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { - ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context; +static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context; for (size_t i = 0; i < ctx->n_buffers; i++) { ggml_backend_buffer_clear(ctx->buffers[i], value); } } -static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface(void) { - static struct ggml_backend_buffer_i multi_backend_buffer_i = { - /* .get_name = */ ggml_backend_multi_buffer_get_name, - /* .free_buffer = */ ggml_backend_multi_buffer_free_buffer, - /* .get_base = */ NULL, - /* .init_tensor = */ NULL, - /* .memset_tensor = */ NULL, - /* .set_tensor = */ NULL, - /* .get_tensor = */ NULL, - /* .cpy_tensor = */ NULL, - /* .clear = */ ggml_backend_multi_buffer_clear, - /* .reset = */ NULL, - }; +static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = { + /* .get_name = */ ggml_backend_multi_buffer_get_name, + /* .free_buffer = */ ggml_backend_multi_buffer_free_buffer, + /* .get_base = */ NULL, + /* .init_tensor = */ NULL, + /* .memset_tensor = */ NULL, + /* .set_tensor = */ NULL, + /* .get_tensor = */ NULL, + /* .cpy_tensor = */ NULL, + /* .clear = */ ggml_backend_multi_buffer_clear, + /* .reset = */ NULL, +}; - return multi_backend_buffer_i; -} - -GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) { - ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) malloc(sizeof(struct ggml_backend_multi_buffer_context)); +ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) { + ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) malloc(sizeof(struct ggml_backend_multi_buffer_context)); ctx->n_buffers = n_buffers; ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t)); @@ -1028,16 +1306,16 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_back total_size += ggml_backend_buffer_get_size(buffers[i]); } - return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_context_interface(), ctx, total_size); + return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_i, ctx, total_size); } -GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) { +bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) { return buffer->iface.get_name == ggml_backend_multi_buffer_get_name; } -GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) { +void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) { GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer)); - ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context; + ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context; for (size_t i = 0; i < ctx->n_buffers; i++) { ggml_backend_buffer_set_usage(ctx->buffers[i], usage); } @@ -1592,7 +1870,8 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg i_split++; if (i_split >= sched->splits_capacity) { sched->splits_capacity *= 2; - sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split)); + sched->splits = (ggml_backend_sched_split *) + realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split)); GGML_ASSERT(sched->splits != NULL); } split = &sched->splits[i_split]; @@ -1678,11 +1957,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg sched->prev_leaf_backend_ids = tmp; } - int graph_size = MAX(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies; + int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies; if (sched->graph.size < graph_size) { sched->graph.size = graph_size; - sched->graph.nodes = realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *)); - sched->graph.leafs = realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *)); + sched->graph.nodes = (ggml_tensor **) realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *)); + sched->graph.leafs = (ggml_tensor **) realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *)); GGML_ASSERT(sched->graph.nodes != NULL); GGML_ASSERT(sched->graph.leafs != NULL); } @@ -1881,7 +2160,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s // record the event of this copy if (split->n_inputs > 0) { if (sched->events[split_backend_id][sched->cur_copy] != NULL) { - ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy]); + ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend); } } } @@ -1901,7 +2180,7 @@ ggml_backend_sched_t ggml_backend_sched_new( GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS); GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU - struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched)); + struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched)); sched->debug = getenv("GGML_SCHED_DEBUG") != NULL; sched->n_backends = n_backends; @@ -1910,21 +2189,21 @@ ggml_backend_sched_t ggml_backend_sched_new( // initialize hash table // FIXME: needs to be size*2 to account for leafs (do it in graph_split instead) sched->hash_set = ggml_hash_set_new(graph_size); - sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0])); - sched->hv_tensor_copies = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *)); + sched->hv_tensor_backend_ids = (int *) malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0])); + sched->hv_tensor_copies = (ggml_tensor **) malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *)); const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2; - sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0])); - sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0])); - sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0])); - sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0])); + sched->node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->node_backend_ids[0])); + sched->leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->leaf_backend_ids[0])); + sched->prev_node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0])); + sched->prev_leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0])); sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false); - sched->context_buffer = malloc(sched->context_buffer_size); + sched->context_buffer = (char *) malloc(sched->context_buffer_size); const int initial_splits_capacity = 16; - sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0])); + sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0])); sched->splits_capacity = initial_splits_capacity; for (int b = 0; b < n_backends; b++) { @@ -1933,7 +2212,7 @@ ggml_backend_sched_t ggml_backend_sched_new( GGML_ASSERT(ggml_backend_supports_buft(backends[b], sched->bufts[b])); if (sched->n_copies > 1) { for (int c = 0; c < sched->n_copies; c++) { - sched->events[b][c] = ggml_backend_event_new(backends[b]); + sched->events[b][c] = ggml_backend_event_new(backends[b]->device); } } } @@ -2169,8 +2448,8 @@ static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) { struct ggml_hash_set hash_set = ggml_hash_set_new(graph->visited_hash_set.size); - struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT - bool * node_init = calloc(hash_set.size, sizeof(node_init[0])); + struct ggml_tensor ** node_copies = (ggml_tensor **) calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT + bool * node_init = (bool *) calloc(hash_set.size, sizeof(node_init[0])); struct ggml_init_params params = { /* .mem_size = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false), @@ -2188,7 +2467,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s free(node_init); ggml_free(ctx_allocated); ggml_free(ctx_unallocated); - return (struct ggml_backend_graph_copy) { + return { /* .buffer = */ NULL, /* .ctx_allocated = */ NULL, /* .ctx_unallocated = */ NULL, @@ -2211,7 +2490,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s free(node_init); ggml_free(ctx_allocated); ggml_free(ctx_unallocated); - return (struct ggml_backend_graph_copy) { + return { /* .buffer = */ NULL, /* .ctx_allocated = */ NULL, /* .ctx_unallocated = */ NULL, @@ -2240,7 +2519,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s free(node_copies); free(node_init); - return (struct ggml_backend_graph_copy) { + return { /* .buffer = */ buffer, /* .ctx_allocated = */ ctx_allocated, /* .ctx_unallocated = */ ctx_unallocated, diff --git a/ggml/src/ggml-blas.cpp b/ggml/src/ggml-blas.cpp index 6d99c6bea..b850e6a8d 100644 --- a/ggml/src/ggml-blas.cpp +++ b/ggml/src/ggml-blas.cpp @@ -235,25 +235,25 @@ static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct g // backend interface -GGML_CALL static const char * ggml_backend_blas_name(ggml_backend_t backend) { +static const char * ggml_backend_blas_name(ggml_backend_t backend) { return "BLAS"; GGML_UNUSED(backend); } -GGML_CALL static void ggml_backend_blas_free(ggml_backend_t backend) { +static void ggml_backend_blas_free(ggml_backend_t backend) { ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context; delete ctx; delete backend; } -GGML_CALL static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) { +static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) { return ggml_backend_cpu_buffer_type(); GGML_UNUSED(backend); } -GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context; for (int i = 0; i < cgraph->n_nodes; i++) { @@ -285,7 +285,7 @@ GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t GGML_UNUSED(backend); } -GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { +static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { const struct ggml_tensor * src0 = op->src[0]; const struct ggml_tensor * src1 = op->src[1]; @@ -300,7 +300,7 @@ GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, cons GGML_UNUSED(backend); } -GGML_CALL static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { +static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { return ggml_backend_buft_is_host(buft); GGML_UNUSED(backend); @@ -322,11 +322,8 @@ static struct ggml_backend_i blas_backend_i = { /* .supports_op = */ ggml_backend_blas_supports_op, /* .supports_buft = */ ggml_backend_blas_supports_buft, /* .offload_op = */ NULL, - /* .event_new = */ NULL, - /* .event_free = */ NULL, /* .event_record = */ NULL, /* .event_wait = */ NULL, - /* .event_synchronize = */ NULL, }; static ggml_guid_t ggml_backend_blas_guid(void) { @@ -340,6 +337,7 @@ ggml_backend_t ggml_backend_blas_init(void) { ggml_backend_t backend = new ggml_backend { /* .guid = */ ggml_backend_blas_guid(), /* .interface = */ blas_backend_i, + /* .device = */ nullptr, /* .context = */ ctx, }; @@ -356,7 +354,7 @@ ggml_backend_t ggml_backend_blas_init(void) { return backend; } -GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend) { +bool ggml_backend_is_blas(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid()); } diff --git a/ggml/src/ggml-cann.cpp b/ggml/src/ggml-cann.cpp index d3ab78006..db5f8f186 100644 --- a/ggml/src/ggml-cann.cpp +++ b/ggml/src/ggml-cann.cpp @@ -39,69 +39,6 @@ #include "ggml-common.h" -/** - * @brief Default logging callback for GGML. - * - * This function is the default logging callback that logs messages to stderr. - * - * @param level The log level. - * @param msg The log message. - * @param user_data User data passed to the callback. - */ -static void ggml_cann_default_log_callback(enum ggml_log_level level, - const char* msg, void* user_data) { - GGML_UNUSED(level); - GGML_UNUSED(user_data); - fprintf(stderr, "%s", msg); -} - -ggml_log_callback ggml_cann_log_callback = ggml_cann_default_log_callback; -void* ggml_cann_log_user_data = NULL; - -GGML_API void ggml_backend_cann_log_set_callback(ggml_log_callback log_callback, - void* user_data) { - ggml_cann_log_callback = log_callback; - ggml_cann_log_user_data = user_data; -} - -#define GGML_CANN_LOG_INFO(...) ggml_cann_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__) -#define GGML_CANN_LOG_WARN(...) ggml_cann_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__) -#define GGML_CANN_LOG_ERROR(...) \ - ggml_cann_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) - -GGML_ATTRIBUTE_FORMAT(2, 3) - -/** - * @brief Log a message using the current logging callback. - * - * This function formats a log message and passes it to the current logging - * callback. - * - * @param level The log level. - * @param format The format string for the log message. - * @param ... The arguments for the format string. - */ -static void ggml_cann_log(enum ggml_log_level level, const char* format, ...) { - if (ggml_cann_log_callback != NULL) { - va_list args; - va_start(args, format); - char buffer[128]; - int len = vsnprintf(buffer, 128, format, args); - if (len < 128) { - ggml_cann_log_callback(level, buffer, ggml_cann_log_user_data); - } else { - // vsnprintf adds a null terminator - std::vector buffer2(len + 1); - va_end(args); - va_start(args, format); - vsnprintf(&buffer2[0], buffer2.size(), format, args); - ggml_cann_log_callback(level, buffer2.data(), - ggml_cann_log_user_data); - } - va_end(args); - } -} - /** * @brief Handles CANN errors by printing an error message and aborting. * @@ -116,10 +53,10 @@ static void ggml_cann_log(enum ggml_log_level level, const char* format, ...) { int32_t id = -1; aclrtGetDevice(&id); - GGML_CANN_LOG_ERROR("CANN error: %s\n", msg); - GGML_CANN_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func, + GGML_LOG_ERROR("CANN error: %s\n", msg); + GGML_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func, file, line); - GGML_CANN_LOG_ERROR(" %s\n", stmt); + GGML_LOG_ERROR(" %s\n", stmt); // abort with GGML_ASSERT to get a stack trace GGML_ABORT("CANN error"); } @@ -165,7 +102,7 @@ static ggml_cann_device_info ggml_cann_init() { aclError err = aclrtGetDeviceCount((uint32_t*)&info.device_count); if (err != ACL_SUCCESS) { - GGML_CANN_LOG_ERROR("%s: failed to initialize CANN: %s\n", + GGML_LOG_ERROR("%s: failed to initialize CANN: %s\n", __func__, aclGetRecentErrMsg()); return info; } @@ -315,7 +252,7 @@ struct ggml_cann_pool_leg : public ggml_cann_pool { *actual_size = look_ahead_size; pool_size += look_ahead_size; #ifdef DEBUG_CANN_MALLOC - GGML_CANN_LOG_INFO( + GGML_LOG_INFO( "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, " "requested %u MB\n", __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024), @@ -470,7 +407,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { // add to the pool pool_size += reserve_size; - // GGML_CANN_LOG_INFO("cann pool[%d]: size increased to %llu MB ( + // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB ( // reserved %llu MB)\n", // device, (unsigned long long) (pool_size/1024/1024), // (unsigned long long) (reserve_size/1024/1024)); @@ -483,7 +420,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { pool_used += size; #ifdef DEBUG_CANN_MALLOC - GGML_CANN_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device, + GGML_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device, (unsigned long long)size, (unsigned long long)ptr); #endif return ptr; @@ -497,7 +434,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { */ void free(void* ptr, size_t size) override { #ifdef DEBUG_CANN_MALLOC - GGML_CANN_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device, + GGML_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device, (unsigned long long)size, (unsigned long long)ptr); #endif @@ -560,7 +497,7 @@ struct ggml_backend_cann_buffer_context { * @return A pointer to a C-string containing the name of the buffer. */ -GGML_CALL static const char* ggml_backend_cann_buffer_get_name( +static const char* ggml_backend_cann_buffer_get_name( ggml_backend_buffer_t buffer) { return "CANN"; @@ -576,7 +513,7 @@ GGML_CALL static const char* ggml_backend_cann_buffer_get_name( * @param buffer The buffer to check. * @return true if the buffer is a CANN buffer, false otherwise. */ -GGML_CALL static bool ggml_backend_buffer_is_cann( +static bool ggml_backend_buffer_is_cann( ggml_backend_buffer_t buffer) { return buffer->iface.get_name == ggml_backend_cann_buffer_get_name; } @@ -589,7 +526,7 @@ GGML_CALL static bool ggml_backend_buffer_is_cann( * * @param buffer The CANN buffer to free. */ -GGML_CALL static void ggml_backend_cann_buffer_free_buffer( +static void ggml_backend_cann_buffer_free_buffer( ggml_backend_buffer_t buffer) { ggml_backend_cann_buffer_context* ctx = (ggml_backend_cann_buffer_context*)buffer->context; @@ -605,7 +542,7 @@ GGML_CALL static void ggml_backend_cann_buffer_free_buffer( * @param buffer The CANN buffer whose base pointer is to be retrieved. * @return A pointer to the base of the device memory allocated for the buffer. */ -GGML_CALL static void* ggml_backend_cann_buffer_get_base( +static void* ggml_backend_cann_buffer_get_base( ggml_backend_buffer_t buffer) { ggml_backend_cann_buffer_context* ctx = (ggml_backend_cann_buffer_context*)buffer->context; @@ -625,9 +562,9 @@ GGML_CALL static void* ggml_backend_cann_buffer_get_base( * @param dst Pointer to the destination buffer where transformed data will be * stored. */ -GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor, - const void* src, - void* dst) { +static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor, + const void* src, + void* dst) { int64_t n_elems = ggml_nelements(tensor); int64_t groups = n_elems / QK4_0; @@ -677,7 +614,7 @@ GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor, * @param dst Pointer to the destination buffer where the Q4.0 formatted data * will be stored. */ -GGML_CALL static void ggml_backend_cann_transform_back_q4_0( +static void ggml_backend_cann_transform_back_q4_0( const ggml_tensor* tensor, void* src, void* dst) { int64_t n_elems = ggml_nelements(tensor); @@ -726,9 +663,9 @@ GGML_CALL static void ggml_backend_cann_transform_back_q4_0( * @param dst Pointer to the destination buffer where transformed data will be * stored. */ -GGML_CALL static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor, - const void* src, - void* dst) { +static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor, + const void* src, + void* dst) { int64_t n_elems = ggml_nelements(tensor); int64_t groups = n_elems / QK8_0; size_t quant_bytes = n_elems * sizeof(uint8_t); @@ -760,7 +697,7 @@ GGML_CALL static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor, * @param dst Pointer to the destination buffer where the Q8.0 formatted data * will be stored. */ -GGML_CALL static void ggml_backend_cann_transform_back_q8_0( +static void ggml_backend_cann_transform_back_q8_0( const ggml_tensor* tensor, const void* src, void* dst) { int64_t n_elems = ggml_nelements(tensor); int64_t groups = n_elems / QK8_0; @@ -792,8 +729,8 @@ GGML_CALL static void ggml_backend_cann_transform_back_q8_0( * @param dst Pointer to the destination buffer where transformed data will be * stored. */ -GGML_CALL static void ggml_backend_cann_transform(ggml_tensor* tensor, - const void* src, void* dst) { +static void ggml_backend_cann_transform(ggml_tensor* tensor, + const void* src, void* dst) { switch (tensor->type) { case GGML_TYPE_Q4_0: ggml_backend_cann_transform_q4_0(tensor, src, dst); @@ -818,7 +755,7 @@ GGML_CALL static void ggml_backend_cann_transform(ggml_tensor* tensor, * @param dst Pointer to the destination buffer where transformed tensor data * will be stored. */ -GGML_CALL static void ggml_backend_cann_transform_back( +static void ggml_backend_cann_transform_back( const ggml_tensor* tensor, void* src, void* dst) { switch (tensor->type) { case GGML_TYPE_Q4_0: @@ -841,7 +778,7 @@ GGML_CALL static void ggml_backend_cann_transform_back( * @param type The tensor type to check. * @return true if transformation is needed, false otherwise. */ -GGML_CALL static bool need_transform(ggml_type type) { +static bool need_transform(ggml_type type) { switch (type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q8_0: @@ -860,7 +797,7 @@ GGML_CALL static bool need_transform(ggml_type type) { * @param buffer The CANN buffer from which to initialize the tensor. * @param tensor Pointer to the tensor to be initialized. */ -GGML_CALL static void ggml_backend_cann_buffer_init_tensor( +static void ggml_backend_cann_buffer_init_tensor( ggml_backend_buffer_t buffer, ggml_tensor* tensor) { if (tensor->view_src != NULL && tensor->view_offs == 0) { GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft); @@ -896,7 +833,7 @@ GGML_CALL static void ggml_backend_cann_buffer_init_tensor( * @param offset Offset in the source data from where to start copying. * @param size Size of the data to be copied, in bytes. */ -GGML_CALL static void ggml_backend_cann_buffer_set_tensor( +static void ggml_backend_cann_buffer_set_tensor( ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data, size_t offset, size_t size) { ggml_backend_cann_buffer_context *ctx = @@ -941,7 +878,7 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor( * @param offset Offset in the destination buffer where to start copying. * @param size Size of the data to be copied, in bytes. */ -GGML_CALL static void ggml_backend_cann_buffer_get_tensor( +static void ggml_backend_cann_buffer_get_tensor( ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data, size_t offset, size_t size) { ggml_backend_cann_buffer_context* ctx = @@ -975,7 +912,7 @@ GGML_CALL static void ggml_backend_cann_buffer_get_tensor( * @param dst Pointer to the destination tensor where the data will be copied. * @return true if the copy operation succeeded, false otherwise. */ -GGML_CALL static bool ggml_backend_cann_buffer_cpy_tensor( +static bool ggml_backend_cann_buffer_cpy_tensor( ggml_backend_buffer_t buffer, const ggml_tensor* src, ggml_tensor* dst) { if (ggml_backend_buffer_is_cann(src->buffer)) { ggml_backend_cann_buffer_context* src_ctx = @@ -1017,7 +954,7 @@ GGML_CALL static bool ggml_backend_cann_buffer_cpy_tensor( * @param buffer The CANN buffer to be cleared. * @param value The value to which each byte in the buffer will be set. */ -GGML_CALL static void ggml_backend_cann_buffer_clear( +static void ggml_backend_cann_buffer_clear( ggml_backend_buffer_t buffer, uint8_t value) { ggml_backend_cann_buffer_context* ctx = (ggml_backend_cann_buffer_context*)buffer->context; @@ -1065,7 +1002,7 @@ struct ggml_backend_cann_buffer_type_context { * @param buft Pointer to the buffer type context. * @return Const pointer to the C-style string containing the name. */ -GGML_CALL static const char* ggml_backend_cann_buffer_type_name( +static const char* ggml_backend_cann_buffer_type_name( ggml_backend_buffer_type_t buft) { return "CANN"; @@ -1082,7 +1019,7 @@ GGML_CALL static const char* ggml_backend_cann_buffer_type_name( * @param size Size in bytes of the buffer to allocate. * @return Pointer to the allocated buffer, or nullptr if allocation fails. */ -GGML_CALL static ggml_backend_buffer_t +static ggml_backend_buffer_t ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { ggml_backend_cann_buffer_type_context* buft_ctx = @@ -1095,7 +1032,7 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, void* dev_ptr; aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST); if (err != ACL_SUCCESS) { - GGML_CANN_LOG_ERROR( + GGML_LOG_ERROR( "%s: allocating %.2f MiB on device %d: aclrtMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, aclGetRecentErrMsg()); @@ -1121,7 +1058,7 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, * @return The alignment requirement in bytes (fixed at 128 bytes for CANN * buffers). */ -GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alignment( +static size_t ggml_backend_cann_buffer_type_get_alignment( ggml_backend_buffer_type_t buft) { return 128; @@ -1142,7 +1079,7 @@ GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alignment( * @return The total allocation size in bytes required for the tensor in the * CANN buffer. */ -GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alloc_size( +static size_t ggml_backend_cann_buffer_type_get_alloc_size( ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) { size_t size = ggml_nbytes(tensor); int64_t ne0 = tensor->ne[0]; @@ -1193,7 +1130,7 @@ static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = { * @return A pointer to the buffer type interface for the specified device, or * nullptr if the device index is out of range. */ -GGML_CALL ggml_backend_buffer_type_t +ggml_backend_buffer_type_t ggml_backend_cann_buffer_type(int32_t device) { static std::mutex mutex; std::lock_guard lock(mutex); @@ -1231,7 +1168,7 @@ ggml_backend_cann_buffer_type(int32_t device) { * @param buft Pointer to the host buffer type context. * @return Const pointer to the C-style string containing the name. */ -GGML_CALL static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) { return "CANN_Host"; GGML_UNUSED(buft); @@ -1246,7 +1183,7 @@ GGML_CALL static const char * ggml_backend_cann_host_buffer_type_name(ggml_backe * @param buft Pointer to the host buffer context. * @return Const pointer to the C-style string containing the name. */ -GGML_CALL static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) { return "CANN_Host"; GGML_UNUSED(buffer); @@ -1260,7 +1197,7 @@ GGML_CALL static const char * ggml_backend_cann_host_buffer_name(ggml_backend_bu * * @param buffer The CANN host buffer to free. */ -GGML_CALL static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) { +static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) { ACL_CHECK(aclrtFreeHost(buffer->context)); } @@ -1280,7 +1217,7 @@ static void * ggml_cann_host_malloc(size_t size) { aclError err = aclrtMallocHost((void **) &hostPtr, size); if (err != ACL_SUCCESS) { - GGML_CANN_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__, + GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__, size / 1024.0 / 1024.0, aclGetRecentErrMsg()); return nullptr; } @@ -1294,7 +1231,7 @@ static void * ggml_cann_host_malloc(size_t size) { * @param size Size in bytes of the host buffer to allocate. * @return Pointer to the allocated host buffer, or CPU buffer pointer if allocation fails. */ -GGML_CALL static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { void * hostPtr = ggml_cann_host_malloc(size); if (hostPtr == nullptr) { @@ -1316,7 +1253,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_ * Provides function pointers for allocating, querying properties, and managing * memory for CANN buffer types in the GGML backend. */ -GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() { +ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() { static struct ggml_backend_buffer_type ggml_backend_cann_buffer_type_host = { /* .iface = */ { /* .get_name = */ ggml_backend_cann_host_buffer_type_name, @@ -1326,6 +1263,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() { /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, }, + /* .device = */ nullptr, /* .context = */ nullptr, }; @@ -1495,7 +1433,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx, * @param backend Pointer to the CANN backend structure. * @return A pointer to a constant string representing the backend name. */ -GGML_CALL static const char* ggml_backend_cann_name(ggml_backend_t backend) { +static const char* ggml_backend_cann_name(ggml_backend_t backend) { ggml_backend_cann_context* cann_ctx = (ggml_backend_cann_context*)backend->context; @@ -1510,7 +1448,7 @@ GGML_CALL static const char* ggml_backend_cann_name(ggml_backend_t backend) { * * @param backend Pointer to the CANN backend structure to be freed. */ -GGML_CALL static void ggml_backend_cann_free(ggml_backend_t backend) { +static void ggml_backend_cann_free(ggml_backend_t backend) { ggml_backend_cann_context* cann_ctx = (ggml_backend_cann_context*)backend->context; ACL_CHECK(aclrtSynchronizeDevice()); @@ -1535,7 +1473,7 @@ GGML_CALL static void ggml_backend_cann_free(ggml_backend_t backend) { * @param backend Pointer to the CANN backend structure. * @return Pointer to the buffer type structure for the CANN backend. */ -GGML_CALL static ggml_backend_buffer_type_t +static ggml_backend_buffer_type_t ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) { ggml_backend_cann_context* cann_ctx = (ggml_backend_cann_context*)backend->context; @@ -1556,11 +1494,11 @@ ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) { * @param offset Offset in bytes within the host data. * @param size Size of the data to copy in bytes. */ -GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend, - ggml_tensor *tensor, - const void *data, - size_t offset, - size_t size) { +static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend, + ggml_tensor *tensor, + const void *data, + size_t offset, + size_t size) { ggml_backend_cann_context *cann_ctx = (ggml_backend_cann_context *)backend->context; @@ -1587,7 +1525,7 @@ GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend, } } -GGML_CALL static void ggml_backend_cann_get_tensor_async( +static void ggml_backend_cann_get_tensor_async( ggml_backend_t backend, const ggml_tensor *tensor, void *data, size_t offset, size_t size) { ggml_backend_cann_context *cann_ctx = @@ -1626,7 +1564,7 @@ GGML_CALL static void ggml_backend_cann_get_tensor_async( * @param dst Pointer to the destination tensor to copy data to. * @return true if the copy operation succeeds, false otherwise. */ -GGML_CALL static bool ggml_backend_cann_cpy_tensor_async( +static bool ggml_backend_cann_cpy_tensor_async( ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor* src, ggml_tensor* dst) { GGML_ASSERT(ggml_backend_is_cann(backend_src) || @@ -1694,7 +1632,7 @@ GGML_CALL static bool ggml_backend_cann_cpy_tensor_async( * * @param backend Pointer to the CANN backend structure to synchronize. */ -GGML_CALL static void ggml_backend_cann_synchronize(ggml_backend_t backend) { +static void ggml_backend_cann_synchronize(ggml_backend_t backend) { ggml_backend_cann_context* cann_ctx = (ggml_backend_cann_context*)backend->context; @@ -1715,7 +1653,7 @@ GGML_CALL static void ggml_backend_cann_synchronize(ggml_backend_t backend) { * @return enum ggml_status Returns GGML_STATUS_SUCCESS if computation * completes successfully, otherwise an appropriate error status. */ -GGML_CALL static enum ggml_status ggml_backend_cann_graph_compute( +static enum ggml_status ggml_backend_cann_graph_compute( ggml_backend_t backend, ggml_cgraph* cgraph) { ggml_backend_cann_context* cann_ctx = (ggml_backend_cann_context*)backend->context; @@ -1732,7 +1670,7 @@ GGML_CALL static enum ggml_status ggml_backend_cann_graph_compute( bool ok = ggml_cann_compute_forward(*cann_ctx, node); if (!ok) { - GGML_CANN_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, + GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); } GGML_ASSERT(ok); @@ -1753,7 +1691,7 @@ GGML_CALL static enum ggml_status ggml_backend_cann_graph_compute( * @return bool Returns true if the operation is supported by the backend, * otherwise false. */ -GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend, +static bool ggml_backend_cann_supports_op(ggml_backend_t backend, const ggml_tensor* op) { switch (op->op) { case GGML_OP_UNARY: @@ -1875,7 +1813,7 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) { * @return bool Returns true if the CANN backend supports the buffer type, * otherwise false. */ -GGML_CALL static bool ggml_backend_cann_supports_buft( +static bool ggml_backend_cann_supports_buft( ggml_backend_t backend, ggml_backend_buffer_type_t buft) { if (ggml_backend_buft_is_cann(buft)) { ggml_backend_cann_context * cann_ctx = @@ -1901,7 +1839,7 @@ GGML_CALL static bool ggml_backend_cann_supports_buft( * @return bool Returns true if the operation should be offloaded, otherwise * false. */ -GGML_CALL static bool ggml_backend_cann_offload_op(ggml_backend_t backend, +static bool ggml_backend_cann_offload_op(ggml_backend_t backend, const ggml_tensor* op) { const int min_batch_size = 32; GGML_UNUSED(backend); @@ -2021,11 +1959,8 @@ static ggml_backend_i ggml_backend_cann_interface = { /* .supports_op = */ ggml_backend_cann_supports_op, /* .supports_buft = */ ggml_backend_cann_supports_buft, /* .offload_op = */ ggml_backend_cann_offload_op, - /* .event_new = */ ggml_backend_cann_event_new, - /* .event_free = */ ggml_backend_cann_event_free, /* .event_record = */ ggml_backend_cann_event_record, /* .event_wait = */ ggml_backend_cann_event_wait, - /* .event_synchronize = */ ggml_backend_cann_event_synchronize, }; /** @@ -2042,91 +1977,46 @@ static ggml_guid_t ggml_backend_cann_guid() { return &guid; } -GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) { +ggml_backend_t ggml_backend_cann_init(int32_t device) { aclInit(nullptr); if (device < 0 || device >= ggml_backend_cann_get_device_count()) { - GGML_CANN_LOG_ERROR("%s: error: invalid device %d\n", __func__, device); + GGML_LOG_ERROR("%s: error: invalid device %d\n", __func__, device); return nullptr; } ggml_backend_cann_context* ctx = new ggml_backend_cann_context(device); if (ctx == nullptr) { - GGML_CANN_LOG_ERROR("%s: error: failed to allocate context\n", __func__); + GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__); return nullptr; } ggml_cann_set_device(ctx->device); ggml_backend_t cann_backend = new ggml_backend{/* .guid = */ ggml_backend_cann_guid(), /* .interface = */ ggml_backend_cann_interface, + /* .device = */ nullptr, /* .context = */ ctx}; return cann_backend; } -GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend) { +bool ggml_backend_is_cann(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cann_guid()); } -GGML_CALL int32_t ggml_backend_cann_get_device_count() { +int32_t ggml_backend_cann_get_device_count() { return ggml_cann_info().device_count; } -GGML_CALL void ggml_backend_cann_get_device_description( +void ggml_backend_cann_get_device_description( int32_t device, char* description, size_t description_size) { ggml_cann_set_device(device); const char* soc_name = aclrtGetSocName(); snprintf(description, description_size, "%s", soc_name); } -GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device, size_t* free, - size_t* total) { +void ggml_backend_cann_get_device_memory(int32_t device, size_t* free, + size_t* total) { ggml_cann_set_device(device); ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total)); } - -// backend registry -/** - * @brief Initializes a CANN backend based on the provided parameters. - * - * This function initializes a CANN backend using the device index and then - * initializes the backend using `ggml_backend_cann_init`. - * - * @param params Parameters for initialization (unused in this implementation). - * @param user_data User data containing the device index to initialize the - * backend. - * @return ggml_backend_t The initialized CANN backend. - */ -GGML_CALL static ggml_backend_t ggml_backend_reg_cann_init(const char* params, - void* user_data) { - ggml_backend_t cann_backend = - ggml_backend_cann_init((int)(intptr_t)user_data); - return cann_backend; - - GGML_UNUSED(params); -} - -extern "C" GGML_CALL int ggml_backend_cann_reg_devices(); - -/** - * @brief Registers CANN (Ascend) devices as backend options. - * - * This function initializes ACL, retrieves the number of available CANN - * devices, and registers each device as a backend option using - * `ggml_backend_register`. Each device is given a unique name based on - * `GGML_CANN_NAME` followed by its index. - * - * @return int The number of CANN devices registered. - */ -GGML_CALL int ggml_backend_cann_reg_devices() { - uint32_t device_count = ggml_backend_cann_get_device_count(); - // initialization - for (uint32_t i = 0; i < device_count; i++) { - char name[128]; - snprintf(name, sizeof(name), "CANN%d", i); - ggml_backend_register(name, ggml_backend_reg_cann_init, - ggml_backend_cann_buffer_type(i), - (void*)(intptr_t)i); - } - return device_count; -} diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 6efdab14c..edb61abdf 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -5,12 +5,14 @@ #include "ggml-cuda/common.cuh" #include "ggml-cuda/acc.cuh" #include "ggml-cuda/arange.cuh" +#include "ggml-cuda/argmax.cuh" #include "ggml-cuda/argsort.cuh" #include "ggml-cuda/binbcast.cuh" #include "ggml-cuda/clamp.cuh" #include "ggml-cuda/concat.cuh" #include "ggml-cuda/conv-transpose-1d.cuh" #include "ggml-cuda/convert.cuh" +#include "ggml-cuda/count-equal.cuh" #include "ggml-cuda/cpy.cuh" #include "ggml-cuda/cross-entropy-loss.cuh" #include "ggml-cuda/diagmask.cuh" @@ -56,54 +58,16 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size"); -static void ggml_cuda_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) { - GGML_UNUSED(level); - GGML_UNUSED(user_data); - fprintf(stderr, "%s", msg); -} - -ggml_log_callback ggml_cuda_log_callback = ggml_cuda_default_log_callback; -void * ggml_cuda_log_user_data = NULL; - -GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data) { - ggml_cuda_log_callback = log_callback; - ggml_cuda_log_user_data = user_data; -} - -#define GGML_CUDA_LOG_INFO(...) ggml_cuda_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__) -#define GGML_CUDA_LOG_WARN(...) ggml_cuda_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__) -#define GGML_CUDA_LOG_ERROR(...) ggml_cuda_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) - -GGML_ATTRIBUTE_FORMAT(2, 3) -static void ggml_cuda_log(enum ggml_log_level level, const char * format, ...) { - if (ggml_cuda_log_callback != NULL) { - va_list args; - va_start(args, format); - char buffer[128]; - int len = vsnprintf(buffer, 128, format, args); - if (len < 128) { - ggml_cuda_log_callback(level, buffer, ggml_cuda_log_user_data); - } else { - std::vector buffer2(len + 1); // vsnprintf adds a null terminator - va_end(args); - va_start(args, format); - vsnprintf(&buffer2[0], buffer2.size(), format, args); - ggml_cuda_log_callback(level, buffer2.data(), ggml_cuda_log_user_data); - } - va_end(args); - } -} - [[noreturn]] void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) { int id = -1; // in case cudaGetDevice fails cudaGetDevice(&id); - GGML_CUDA_LOG_ERROR("CUDA error: %s\n", msg); - GGML_CUDA_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func, file, line); - GGML_CUDA_LOG_ERROR(" %s\n", stmt); - // abort with GGML_ASSERT to get a stack trace - GGML_ABORT("CUDA error"); + GGML_LOG_ERROR(GGML_CUDA_NAME " error: %s\n", msg); + GGML_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func, file, line); + GGML_LOG_ERROR(" %s\n", stmt); + // abort with GGML_ABORT to get a stack trace + GGML_ABORT(GGML_CUDA_NAME " error"); } // this is faster on Windows @@ -166,7 +130,7 @@ static ggml_cuda_device_info ggml_cuda_init() { cudaError_t err = cudaGetDeviceCount(&info.device_count); if (err != cudaSuccess) { - GGML_CUDA_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err)); + GGML_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err)); return info; } @@ -174,16 +138,16 @@ static ggml_cuda_device_info ggml_cuda_init() { int64_t total_vram = 0; #ifdef GGML_CUDA_FORCE_MMQ - GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__); + GGML_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__); #else - GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ: no\n", __func__); + GGML_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ: no\n", __func__); #endif // GGML_CUDA_FORCE_MMQ #ifdef GGML_CUDA_FORCE_CUBLAS - GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: yes\n", __func__); + GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: yes\n", __func__); #else - GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: no\n", __func__); + GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: no\n", __func__); #endif // GGML_CUDA_FORCE_CUBLAS - GGML_CUDA_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count); + GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count); for (int id = 0; id < info.device_count; ++id) { int device_vmm = 0; @@ -204,7 +168,7 @@ static ggml_cuda_device_info ggml_cuda_init() { cudaDeviceProp prop; CUDA_CHECK(cudaGetDeviceProperties(&prop, id)); - GGML_CUDA_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no"); + GGML_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no"); info.default_tensor_split[id] = total_vram; total_vram += prop.totalGlobalMem; @@ -312,7 +276,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool { *actual_size = look_ahead_size; pool_size += look_ahead_size; #ifdef DEBUG_CUDA_MALLOC - GGML_CUDA_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz, + GGML_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024), (uint32_t)(pool_size / 1024 / 1024), (uint32_t)(size / 1024 / 1024)); #endif return ptr; @@ -327,7 +291,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool { return; } } - GGML_CUDA_LOG_WARN("Cuda buffer pool full, increase MAX_CUDA_BUFFERS\n"); + GGML_LOG_WARN(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n"); ggml_cuda_set_device(device); CUDA_CHECK(cudaFree(ptr)); pool_size -= size; @@ -457,26 +421,26 @@ struct ggml_backend_cuda_buffer_context { } }; -GGML_CALL static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; return ctx->name.c_str(); } -GGML_CALL static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) { +static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) { return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name; } -GGML_CALL static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; delete ctx; } -GGML_CALL static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) { +static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; return ctx->dev_ptr; } -GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; if (tensor->view_src != NULL) { @@ -496,7 +460,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t } } -GGML_CALL static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { +static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; ggml_cuda_set_device(ctx->device); @@ -504,7 +468,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); } -GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; ggml_cuda_set_device(ctx->device); @@ -512,7 +476,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); } -GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { +static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; ggml_cuda_set_device(ctx->device); @@ -520,7 +484,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); } -GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { +static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { if (ggml_backend_buffer_is_cuda(src->buffer)) { ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context; ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context; @@ -541,7 +505,7 @@ GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t GGML_UNUSED(buffer); } -GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { +static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; ggml_cuda_set_device(ctx->device); @@ -550,7 +514,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffe CUDA_CHECK(cudaDeviceSynchronize()); } -static ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = { +static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = { /* .get_name = */ ggml_backend_cuda_buffer_get_name, /* .free_buffer = */ ggml_backend_cuda_buffer_free_buffer, /* .get_base = */ ggml_backend_cuda_buffer_get_base, @@ -569,17 +533,17 @@ struct ggml_backend_cuda_buffer_type_context { std::string name; }; -GGML_CALL static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_cuda_buffer_type_get_name(ggml_backend_buffer_type_t buft) { ggml_backend_cuda_buffer_type_context * ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; return ctx->name.c_str(); } static bool ggml_backend_buft_is_cuda(ggml_backend_buffer_type_t buft) { - return buft->iface.get_name == ggml_backend_cuda_buffer_type_name; + return buft->iface.get_name == ggml_backend_cuda_buffer_type_get_name; } -GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; ggml_cuda_set_device(buft_ctx->device); @@ -591,7 +555,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe if (err != cudaSuccess) { // clear the error cudaGetLastError(); - GGML_CUDA_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err)); + GGML_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err)); return nullptr; } @@ -600,13 +564,13 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size); } -GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return 128; GGML_UNUSED(buft); } -GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { +static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { size_t size = ggml_nbytes(tensor); int64_t ne0 = tensor->ne[0]; @@ -621,8 +585,8 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backen GGML_UNUSED(buft); } -static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = { - /* .get_name = */ ggml_backend_cuda_buffer_type_name, +static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = { + /* .get_name = */ ggml_backend_cuda_buffer_type_get_name, /* .alloc_buffer = */ ggml_backend_cuda_buffer_type_alloc_buffer, /* .get_alignment = */ ggml_backend_cuda_buffer_type_get_alignment, /* .get_max_size = */ NULL, // defaults to SIZE_MAX @@ -630,7 +594,7 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = { /* .is_host = */ NULL, }; -GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) { +ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) { static std::mutex mutex; std::lock_guard lock(mutex); @@ -643,9 +607,10 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) { static bool ggml_backend_cuda_buffer_type_initialized = false; if (!ggml_backend_cuda_buffer_type_initialized) { - for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) { + for (int i = 0; i < ggml_backend_cuda_get_device_count(); i++) { ggml_backend_cuda_buffer_types[i] = { /* .iface = */ ggml_backend_cuda_buffer_type_interface, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), i), /* .context = */ new ggml_backend_cuda_buffer_type_context{i, GGML_CUDA_NAME + std::to_string(i)}, }; } @@ -715,7 +680,7 @@ struct ggml_backend_cuda_split_buffer_context { std::vector tensor_extras; }; -GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) { return GGML_CUDA_NAME "_Split"; GGML_UNUSED(buffer); @@ -726,19 +691,19 @@ static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) { GGML_UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds } -GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; delete ctx; } -GGML_CALL static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) { +static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) { // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced return (void *)0x1000; GGML_UNUSED(buffer); } -GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; @@ -786,7 +751,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_bu tensor->extra = extra; } -GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { // split tensors must always be set in their entirety at once GGML_ASSERT(offset == 0); GGML_ASSERT(size == ggml_nbytes(tensor)); @@ -824,7 +789,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buf } } -GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { +static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { // split tensors must always be set in their entirety at once GGML_ASSERT(offset == 0); GGML_ASSERT(size == ggml_nbytes(tensor)); @@ -862,12 +827,12 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buf } } -GGML_CALL static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { +static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { GGML_UNUSED(buffer); GGML_UNUSED(value); } -static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = { +static const ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = { /* .get_name = */ ggml_backend_cuda_split_buffer_get_name, /* .free_buffer = */ ggml_backend_cuda_split_buffer_free_buffer, /* .get_base = */ ggml_backend_cuda_split_buffer_get_base, @@ -882,17 +847,17 @@ static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = { // cuda split buffer type -GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_cuda_split_buffer_type_get_name(ggml_backend_buffer_type_t buft) { return GGML_CUDA_NAME "_Split"; GGML_UNUSED(buft); } static bool ggml_backend_buft_is_cuda_split(ggml_backend_buffer_type_t buft) { - return buft->iface.get_name == ggml_backend_cuda_split_buffer_type_name; + return buft->iface.get_name == ggml_backend_cuda_split_buffer_type_get_name; } -GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point // instead, we allocate them for each tensor separately in init_tensor // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated, @@ -902,13 +867,13 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size); } -GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return 128; GGML_UNUSED(buft); } -GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { +static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context; size_t total_size = 0; @@ -935,14 +900,14 @@ GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_ return total_size; } -GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) { +static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) { return false; GGML_UNUSED(buft); } -static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = { - /* .get_name = */ ggml_backend_cuda_split_buffer_type_name, +static const ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = { + /* .get_name = */ ggml_backend_cuda_split_buffer_type_get_name, /* .alloc_buffer = */ ggml_backend_cuda_split_buffer_type_alloc_buffer, /* .get_alignment = */ ggml_backend_cuda_split_buffer_type_get_alignment, /* .get_max_size = */ NULL, // defaults to SIZE_MAX @@ -950,7 +915,7 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface /* .is_host = */ ggml_backend_cuda_split_buffer_type_is_host, }; -GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) { +ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) { static std::mutex mutex; std::lock_guard lock(mutex); @@ -979,6 +944,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const f struct ggml_backend_buffer_type buft { /* .iface = */ ggml_backend_cuda_split_buffer_type_interface, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), 0), /* .context = */ new ggml_backend_cuda_split_buffer_type_context{tensor_split_arr}, }; @@ -988,19 +954,19 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const f // host buffer type -GGML_CALL static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) { return GGML_CUDA_NAME "_Host"; GGML_UNUSED(buft); } -GGML_CALL static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) { return GGML_CUDA_NAME "_Host"; GGML_UNUSED(buffer); } -GGML_CALL static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { CUDA_CHECK(cudaFreeHost(buffer->context)); } @@ -1014,7 +980,7 @@ static void * ggml_cuda_host_malloc(size_t size) { if (err != cudaSuccess) { // clear the error cudaGetLastError(); - GGML_CUDA_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__, + GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__, size / 1024.0 / 1024.0, cudaGetErrorString(err)); return nullptr; } @@ -1022,7 +988,7 @@ static void * ggml_cuda_host_malloc(size_t size) { return ptr; } -GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { void * ptr = ggml_cuda_host_malloc(size); if (ptr == nullptr) { @@ -1038,7 +1004,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_ return buffer; } -GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() { +ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() { static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = { /* .iface = */ { /* .get_name = */ ggml_backend_cuda_host_buffer_type_name, @@ -1048,6 +1014,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() { /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, }, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), 0), /* .context = */ nullptr, }; @@ -2178,6 +2145,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg } switch (dst->op) { + case GGML_OP_ARGMAX: + ggml_cuda_argmax(ctx, dst); + break; + case GGML_OP_COUNT_EQUAL: + ggml_cuda_count_equal(ctx, dst); + break; case GGML_OP_REPEAT: ggml_cuda_op_repeat(ctx, dst); break; @@ -2280,7 +2253,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg break; case GGML_OP_MUL_MAT: if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) { - GGML_CUDA_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]); + GGML_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]); return false; } else { ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst); @@ -2364,7 +2337,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg cudaError_t err = cudaGetLastError(); if (err != cudaSuccess) { - GGML_CUDA_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst)); + GGML_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst)); CUDA_CHECK(err); } @@ -2375,26 +2348,26 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg // backend -GGML_CALL static const char * ggml_backend_cuda_name(ggml_backend_t backend) { +static const char * ggml_backend_cuda_get_name(ggml_backend_t backend) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; return cuda_ctx->name.c_str(); } -GGML_CALL static void ggml_backend_cuda_free(ggml_backend_t backend) { +static void ggml_backend_cuda_free(ggml_backend_t backend) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; delete cuda_ctx; delete backend; } -GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) { +static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; return ggml_backend_cuda_buffer_type(cuda_ctx->device); } -GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; @@ -2403,7 +2376,7 @@ GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream())); } -GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { +static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; @@ -2412,7 +2385,7 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cuda_ctx->stream())); } -GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) { +static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) { ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer; ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer; @@ -2433,7 +2406,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_ if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) { #ifndef NDEBUG - GGML_CUDA_LOG_WARN("%s: backend and buffer devices do not match\n", __func__); + GGML_LOG_WARN("%s: backend and buffer devices do not match\n", __func__); #endif return false; } @@ -2467,7 +2440,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_ return true; } -GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) { +static void ggml_backend_cuda_synchronize(ggml_backend_t backend) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; CUDA_CHECK(cudaStreamSynchronize(cuda_ctx->stream())); @@ -2475,6 +2448,7 @@ GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) { GGML_UNUSED(backend); } +#ifdef USE_CUDA_GRAPH static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) { graph_node_properties->node_address = node->data; graph_node_properties->node_op = node->op; @@ -2525,8 +2499,9 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra return true; } +#endif -GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { +static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; ggml_cuda_set_device(cuda_ctx->device); @@ -2549,7 +2524,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) { cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true; #ifndef NDEBUG - GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__); + GGML_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__); #endif } } @@ -2600,14 +2575,14 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t if (node->src[0] && node->src[0]->buffer && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) { use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture #ifndef NDEBUG - GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__); + GGML_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__); #endif } if (node->op == GGML_OP_MUL_MAT_ID) { use_cuda_graph = false; // This node type is not supported by CUDA graph capture #ifndef NDEBUG - GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__); + GGML_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__); #endif } @@ -2616,7 +2591,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t // Changes in batch size or context size can cause changes to the grid size of some kernels. use_cuda_graph = false; #ifndef NDEBUG - GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]); + GGML_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]); #endif } @@ -2628,7 +2603,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t if (!ptr) { use_cuda_graph = false; #ifndef NDEBUG - GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to unsupported copy op\n", __func__); + GGML_LOG_WARN("%s: disabling CUDA graphs due to unsupported copy op\n", __func__); #endif } else { if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) { @@ -2652,7 +2627,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) { cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true; #ifndef NDEBUG - GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__); + GGML_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__); #endif } } @@ -2691,7 +2666,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t bool ok = ggml_cuda_compute_forward(*cuda_ctx, node); if (!ok) { - GGML_CUDA_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); + GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); } GGML_ASSERT(ok); } @@ -2710,7 +2685,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t use_cuda_graph = false; cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true; #ifndef NDEBUG - GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__); + GGML_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__); #endif } else { graph_evaluated_or_captured = true; // CUDA graph has been captured @@ -2777,7 +2752,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info); if (stat == cudaErrorGraphExecUpdateFailure) { #ifndef NDEBUG - GGML_CUDA_LOG_ERROR("%s: CUDA graph update failed\n", __func__); + GGML_LOG_ERROR("%s: CUDA graph update failed\n", __func__); #endif // The pre-existing graph exec cannot be updated due to violated constraints // so instead clear error and re-instantiate @@ -2798,8 +2773,188 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t return GGML_STATUS_SUCCESS; } -GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) { - ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context; +static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) { + ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; + + CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, cuda_ctx->stream())); +} + +static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) { + ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; + + if (ggml_backend_is_cuda(backend)) { + CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), (cudaEvent_t)event->context, 0)); + } else { +#if 0 + // untested + auto wait_fn = [](void * user_data) { + ggml_backend_event_t event = (ggml_backend_event_t)user_data; + ggml_backend_event_synchronize(event); + }; + + CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event)); +#endif + GGML_ABORT("fatal error"); + } +} + +static const ggml_backend_i ggml_backend_cuda_interface = { + /* .get_name = */ ggml_backend_cuda_get_name, + /* .free = */ ggml_backend_cuda_free, + /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type, + /* .set_tensor_async = */ ggml_backend_cuda_set_tensor_async, + /* .get_tensor_async = */ ggml_backend_cuda_get_tensor_async, + /* .cpy_tensor_async = */ ggml_backend_cuda_cpy_tensor_async, + /* .synchronize = */ ggml_backend_cuda_synchronize, + /* .graph_plan_create = */ NULL, + /* .graph_plan_free = */ NULL, + /* .graph_plan_update = */ NULL, + /* .graph_plan_compute = */ NULL, + /* .graph_compute = */ ggml_backend_cuda_graph_compute, + /* .supports_op = */ NULL, // moved to device + /* .supports_buft = */ NULL, // moved to device + /* .offload_op = */ NULL, // moved to device + /* .event_record = */ ggml_backend_cuda_event_record, + /* .event_wait = */ ggml_backend_cuda_event_wait, +}; + +static ggml_guid_t ggml_backend_cuda_guid() { + static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 }; + return &guid; +} + +bool ggml_backend_is_cuda(ggml_backend_t backend) { + return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid()); +} + +int ggml_backend_cuda_get_device_count() { + return ggml_cuda_info().device_count; +} + +void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) { + cudaDeviceProp prop; + CUDA_CHECK(cudaGetDeviceProperties(&prop, device)); + snprintf(description, description_size, "%s", prop.name); +} + +void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) { + ggml_cuda_set_device(device); + + CUDA_CHECK(cudaMemGetInfo(free, total)); +} + +bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) { + if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) { + return false; + } + +#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA) + cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly); + if (err != cudaSuccess) { + // clear the error + cudaGetLastError(); + + GGML_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__, + size / 1024.0 / 1024.0, cudaGetErrorString(err)); + return false; + } + return true; +#else + return false; +#endif +} + +void ggml_backend_cuda_unregister_host_buffer(void * buffer) { + if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) { + return; + } + + cudaError_t err = cudaHostUnregister(buffer); + if (err != cudaSuccess) { + // clear the error + cudaGetLastError(); + } +} + + +// backend device + +struct ggml_backend_cuda_device_context { + int device; + std::string name; + std::string description; +}; + +static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) { + ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context; + return ctx->name.c_str(); +} + +static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t dev) { + ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context; + return ctx->description.c_str(); +} + +static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { + ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context; + ggml_cuda_set_device(ctx->device); + CUDA_CHECK(cudaMemGetInfo(free, total)); +} + +static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) { + GGML_UNUSED(dev); + return GGML_BACKEND_DEVICE_TYPE_GPU_FULL; +} + +static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) { + props->name = ggml_backend_cuda_device_get_name(dev); + props->description = ggml_backend_cuda_device_get_description(dev); + props->type = ggml_backend_cuda_device_get_type(dev); + ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total); + + bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr; +#ifdef GGML_CUDA_NO_PEER_COPY + bool events = false; +#else + bool events = true; +#endif + + props->caps = { + /* .async = */ true, + /* .host_buffer = */ host_buffer, + /* .buffer_from_host_ptr = */ false, + /* .events = */ events, + }; +} + +static ggml_backend_t ggml_backend_cuda_device_init(ggml_backend_dev_t dev, const char * params) { + GGML_UNUSED(params); + ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context; + return ggml_backend_cuda_init(ctx->device); +} + +static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_buffer_type(ggml_backend_dev_t dev) { + ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context; + return ggml_backend_cuda_buffer_type(ctx->device); +} + +static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_host_buffer_type(ggml_backend_dev_t dev) { + GGML_UNUSED(dev); + return ggml_backend_cuda_host_buffer_type(); +} + +static ggml_backend_buffer_t ggml_backend_cuda_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) { + GGML_UNUSED(dev); + GGML_UNUSED(ptr); + GGML_UNUSED(size); + GGML_UNUSED(max_tensor_size); + return nullptr; +} + +// TODO: move these functions here +static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) { + ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context; + switch (op->op) { case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { @@ -2929,6 +3084,15 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons return false; } break; case GGML_OP_DUP: + { + ggml_type src0_type = op->src[0]->type; + return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16; + } break; + case GGML_OP_ARGMAX: + case GGML_OP_COUNT_EQUAL: + { + return true; + } break; case GGML_OP_REPEAT: { ggml_type src0_type = op->src[0]->type; @@ -3004,7 +3168,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) { return true; } - const int cc = ggml_cuda_info().devices[cuda_ctx->device].cc; + const int cc = ggml_cuda_info().devices[dev_ctx->device].cc; return cc >= CC_VOLTA && cc < CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16; } case GGML_OP_CROSS_ENTROPY_LOSS: @@ -3014,205 +3178,181 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons default: return false; } - - GGML_UNUSED(backend); } -GGML_CALL static bool ggml_backend_cuda_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { +static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) { if (ggml_backend_buft_is_cuda_split(buft)) { return true; } if (ggml_backend_buft_is_cuda(buft)) { - ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; + ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context; ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; - return buft_ctx->device == cuda_ctx->device; + return buft_ctx->device == dev_ctx->device; } return false; } -GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const ggml_tensor * op) { +static bool ggml_backend_cuda_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) { const int min_batch_size = 32; return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) || (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID); - GGML_UNUSED(backend); + GGML_UNUSED(dev); } -static ggml_backend_event_t ggml_backend_cuda_event_new(ggml_backend_t backend) { +static ggml_backend_event_t ggml_backend_cuda_device_event_new(ggml_backend_dev_t dev) { #ifdef GGML_CUDA_NO_PEER_COPY return nullptr; #else - ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; + ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context; - ggml_cuda_set_device(cuda_ctx->device); + ggml_cuda_set_device(dev_ctx->device); cudaEvent_t event; CUDA_CHECK(cudaEventCreateWithFlags(&event, cudaEventDisableTiming)); return new ggml_backend_event { - /* .backend = */ backend, + /* .device = */ dev, /* .context = */ event, }; #endif } -static void ggml_backend_cuda_event_free(ggml_backend_event_t event) { - CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context)); +static void ggml_backend_cuda_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) { + GGML_UNUSED(dev); + CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context)); delete event; } -static void ggml_backend_cuda_event_record(ggml_backend_event_t event) { - ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)event->backend->context; - - CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, cuda_ctx->stream())); -} - -static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) { - ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; - - if (ggml_backend_is_cuda(event->backend)) { - CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), (cudaEvent_t)event->context, 0)); - } else { -#if 0 - // untested - auto wait_fn = [](void * user_data) { - ggml_backend_event_t event = (ggml_backend_event_t)user_data; - ggml_backend_event_synchronize(event); - }; - - CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event)); -#endif - GGML_ABORT("fatal error"); - } -} - -static void ggml_backend_cuda_event_synchronize(ggml_backend_event_t event) { +static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) { + GGML_UNUSED(dev); CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context)); } -static ggml_backend_i ggml_backend_cuda_interface = { - /* .get_name = */ ggml_backend_cuda_name, - /* .free = */ ggml_backend_cuda_free, - /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type, - /* .set_tensor_async = */ ggml_backend_cuda_set_tensor_async, - /* .get_tensor_async = */ ggml_backend_cuda_get_tensor_async, - /* .cpy_tensor_async = */ ggml_backend_cuda_cpy_tensor_async, - /* .synchronize = */ ggml_backend_cuda_synchronize, - /* .graph_plan_create = */ NULL, - /* .graph_plan_free = */ NULL, - /* .graph_plan_update = */ NULL, - /* .graph_plan_compute = */ NULL, - /* .graph_compute = */ ggml_backend_cuda_graph_compute, - /* .supports_op = */ ggml_backend_cuda_supports_op, - /* .supports_buft = */ ggml_backend_cuda_supports_buft, - /* .offload_op = */ ggml_backend_cuda_offload_op, - /* .event_new = */ ggml_backend_cuda_event_new, - /* .event_free = */ ggml_backend_cuda_event_free, - /* .event_record = */ ggml_backend_cuda_event_record, - /* .event_wait = */ ggml_backend_cuda_event_wait, - /* .event_synchronize = */ ggml_backend_cuda_event_synchronize, +static const ggml_backend_device_i ggml_backend_cuda_device_interface = { + /* .get_name = */ ggml_backend_cuda_device_get_name, + /* .get_description = */ ggml_backend_cuda_device_get_description, + /* .get_memory = */ ggml_backend_cuda_device_get_memory, + /* .get_type = */ ggml_backend_cuda_device_get_type, + /* .get_props = */ ggml_backend_cuda_device_get_props, + /* .init_backend = */ ggml_backend_cuda_device_init, + /* .get_buffer_type = */ ggml_backend_cuda_device_get_buffer_type, + /* .get_host_buffer_type = */ ggml_backend_cuda_device_get_host_buffer_type, + /* .buffer_from_host_ptr = */ ggml_backend_cuda_device_buffer_from_host_ptr, + /* .supports_op = */ ggml_backend_cuda_device_supports_op, + /* .supports_buft = */ ggml_backend_cuda_device_supports_buft, + /* .offload_op = */ ggml_backend_cuda_device_offload_op, + /* .event_new = */ ggml_backend_cuda_device_event_new, + /* .event_free = */ ggml_backend_cuda_device_event_free, + /* .event_synchronize = */ ggml_backend_cuda_device_event_synchronize, }; -static ggml_guid_t ggml_backend_cuda_guid() { - static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 }; - return &guid; +// backend reg + +struct ggml_backend_cuda_reg_context { + std::vector devices; +}; + +static const char * ggml_backend_cuda_reg_get_name(ggml_backend_reg_t reg) { + GGML_UNUSED(reg); + return GGML_CUDA_NAME; } -GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) { +static size_t ggml_backend_cuda_reg_get_device_count(ggml_backend_reg_t reg) { + ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *)reg->context; + return ctx->devices.size(); +} + +static ggml_backend_dev_t ggml_backend_cuda_reg_get_device(ggml_backend_reg_t reg, size_t index) { + ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *)reg->context; + GGML_ASSERT(index < ctx->devices.size()); + return ctx->devices[index]; +} + +static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) { + GGML_UNUSED(reg); + if (strcmp(name, "ggml_backend_split_buffer_type") == 0) { + return (void *)ggml_backend_cuda_split_buffer_type; + } + if (strcmp(name, "ggml_backend_register_host_buffer") == 0) { + return (void *)ggml_backend_cuda_register_host_buffer; + } + if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) { + return (void *)ggml_backend_cuda_unregister_host_buffer; + } + return nullptr; +} + +static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = { + /* .get_name = */ ggml_backend_cuda_reg_get_name, + /* .get_device_count = */ ggml_backend_cuda_reg_get_device_count, + /* .get_device_get = */ ggml_backend_cuda_reg_get_device, + /* .get_proc_address = */ ggml_backend_cuda_reg_get_proc_address, +}; + +// backend registry +ggml_backend_reg_t ggml_backend_cuda_reg() { + static ggml_backend_reg reg; + static bool initialized = false; + + { + static std::mutex mutex; + std::lock_guard lock(mutex); + if (!initialized) { + ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context; + + for (int i = 0; i < ggml_cuda_info().device_count; i++) { + ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context; + dev_ctx->device = i; + dev_ctx->name = GGML_CUDA_NAME + std::to_string(i); + + ggml_cuda_set_device(i); + cudaDeviceProp prop; + CUDA_CHECK(cudaGetDeviceProperties(&prop, i)); + dev_ctx->description = prop.name; + + ggml_backend_dev_t dev = new ggml_backend_device { + /* .interface = */ ggml_backend_cuda_device_interface, + /* .reg = */ ®, + /* .context = */ dev_ctx + }; + ctx->devices.push_back(dev); + } + + reg = ggml_backend_reg { + /* .interface = */ ggml_backend_cuda_reg_interface, + /* .context = */ ctx + }; + } + + initialized = true; + } + + return ® +} + +ggml_backend_t ggml_backend_cuda_init(int device) { if (device < 0 || device >= ggml_backend_cuda_get_device_count()) { - GGML_CUDA_LOG_ERROR("%s: invalid device %d\n", __func__, device); + GGML_LOG_ERROR("%s: invalid device %d\n", __func__, device); return nullptr; } ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device); if (ctx == nullptr) { - GGML_CUDA_LOG_ERROR("%s: failed to allocate context\n", __func__); + GGML_LOG_ERROR("%s: failed to allocate context\n", __func__); return nullptr; } ggml_backend_t cuda_backend = new ggml_backend { /* .guid = */ ggml_backend_cuda_guid(), /* .interface = */ ggml_backend_cuda_interface, - /* .context = */ ctx + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), device), + /* .context = */ ctx, }; return cuda_backend; } - -GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend) { - return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid()); -} - -GGML_CALL int ggml_backend_cuda_get_device_count() { - return ggml_cuda_info().device_count; -} - -GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) { - cudaDeviceProp prop; - CUDA_CHECK(cudaGetDeviceProperties(&prop, device)); - snprintf(description, description_size, "%s", prop.name); -} - -GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) { - ggml_cuda_set_device(device); - - CUDA_CHECK(cudaMemGetInfo(free, total)); -} - -GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) { - if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) { - return false; - } - -#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA) - cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly); - if (err != cudaSuccess) { - // clear the error - cudaGetLastError(); - - GGML_CUDA_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__, - size / 1024.0 / 1024.0, cudaGetErrorString(err)); - return false; - } - return true; -#else - return false; -#endif -} - -GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer) { - if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) { - return; - } - - cudaError_t err = cudaHostUnregister(buffer); - if (err != cudaSuccess) { - // clear the error - cudaGetLastError(); - } -} - -// backend registry -GGML_CALL static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, void * user_data) { - ggml_backend_t cuda_backend = ggml_backend_cuda_init((int) (intptr_t) user_data); - return cuda_backend; - - GGML_UNUSED(params); -} - -extern "C" GGML_CALL int ggml_backend_cuda_reg_devices(); - -GGML_CALL int ggml_backend_cuda_reg_devices() { - int device_count = ggml_backend_cuda_get_device_count(); - //int device_count = 1; // DEBUG: some tools require delaying CUDA initialization - for (int i = 0; i < device_count; i++) { - char name[128]; - snprintf(name, sizeof(name), "%s%d", GGML_CUDA_NAME, i); - ggml_backend_register(name, ggml_backend_reg_cuda_init, ggml_backend_cuda_buffer_type(i), (void *) (intptr_t) i); - } - return device_count; -} diff --git a/ggml/src/ggml-cuda/argmax.cu b/ggml/src/ggml-cuda/argmax.cu new file mode 100644 index 000000000..aab04eca7 --- /dev/null +++ b/ggml/src/ggml-cuda/argmax.cu @@ -0,0 +1,79 @@ +#include "common.cuh" +#include "argmax.cuh" +#include "sum.cuh" + +#include + +static __global__ void argmax_f32( + const float * x, int32_t * dst, const int64_t ncols, const int64_t nrows) { + + int argmax_thread = 0; + const int64_t row0 = (int64_t)blockIdx.x*WARP_SIZE; + +#pragma unroll + for (int64_t row1 = 0; row1 < WARP_SIZE; ++row1) { + const int64_t row = row0 + row1; + + if (row >= nrows) { + break; + } + + float maxval = -FLT_MAX; + int argmax = -1; + + for (int32_t col = threadIdx.x; col < ncols; col += WARP_SIZE) { + const float val = x[row*ncols + col]; + const int bigger = val > maxval; + const int not_bigger = bigger ^ 0x00000001; + + maxval = maxval*not_bigger + val*bigger; + argmax = argmax*not_bigger + col*bigger; + } + +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, mask, WARP_SIZE); + const int col = __shfl_xor_sync(0xFFFFFFFF, argmax, mask, WARP_SIZE); + const int bigger = val > maxval; + const int not_bigger = bigger ^ 0x00000001; + + maxval = maxval*not_bigger + val*bigger; + argmax = argmax*not_bigger + col*bigger; + } + + const int store = row1 == threadIdx.x; + argmax_thread += store*argmax; + } + + const int row = row0 + threadIdx.x; + + if (row >= nrows) { + return; + } + + dst[row] = argmax_thread; +} + +void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_I32); + + GGML_ASSERT(ggml_is_contiguous(src0)); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + const float * src0_d = (const float *) src0->data; + int32_t * dst_d = (int32_t *) dst->data; + + cudaStream_t stream = ctx.stream(); + + const int64_t num_blocks = (nrows + WARP_SIZE - 1) / WARP_SIZE; + + const dim3 blocks_dim(WARP_SIZE, 1, 1); + const dim3 blocks_num(num_blocks, 1, 1); + + argmax_f32<<>>(src0_d, dst_d, ne00, nrows); +} diff --git a/ggml/src/ggml-cuda/argmax.cuh b/ggml/src/ggml-cuda/argmax.cuh new file mode 100644 index 000000000..5b7223adc --- /dev/null +++ b/ggml/src/ggml-cuda/argmax.cuh @@ -0,0 +1,3 @@ +#include "common.cuh" + +void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 6a4bcdba0..dd203fcde 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -175,6 +175,18 @@ static __device__ void no_device_code( #define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.") #endif // __CUDA_ARCH__ +static __device__ __forceinline__ int warp_reduce_sum(int x) { +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE + return __reduce_add_sync(0xffffffff, x); +#else +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + x += __shfl_xor_sync(0xffffffff, x, mask, 32); + } + return x; +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE +} + static __device__ __forceinline__ float warp_reduce_sum(float x) { #pragma unroll for (int mask = 16; mask > 0; mask >>= 1) { diff --git a/ggml/src/ggml-cuda/count-equal.cu b/ggml/src/ggml-cuda/count-equal.cu new file mode 100644 index 000000000..ffb053b10 --- /dev/null +++ b/ggml/src/ggml-cuda/count-equal.cu @@ -0,0 +1,64 @@ +#include "common.cuh" +#include "count-equal.cuh" + +#include + +template +static __global__ void count_equal(const T * __restrict__ x, const T * __restrict__ y, int64_t * __restrict__ dst, const int64_t dk, const int64_t k) { + const int64_t i0 = (int64_t) blockIdx.x*dk; + const int64_t i1 = min(i0 + dk, k); + + int nequal = 0; + + for (int64_t i = i0 + threadIdx.x; i < i1; i += WARP_SIZE) { + const T xi = x[i]; + const T yi = y[i]; + nequal += xi == yi; + } + + nequal = warp_reduce_sum(nequal); + + if (threadIdx.x != 0) { + return; + } + + atomicAdd((int *) dst, nequal); +} + +void ggml_cuda_count_equal(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == src1->type); + GGML_ASSERT( dst->type == GGML_TYPE_I64); + + GGML_ASSERT(ggml_are_same_shape(src0, src1)); + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(ggml_is_contiguous(dst)); + + int64_t * dst_d = (int64_t *) dst->data; + + cudaStream_t stream = ctx.stream(); + const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm; + + const int64_t ne = ggml_nelements(src0); + GGML_ASSERT(ne < (1 << 30) && "atomicAdd implementation only supports int"); + const int64_t dne = GGML_PAD(ne / (4*nsm), CUDA_COUNT_EQUAL_CHUNK_SIZE); + + CUDA_CHECK(cudaMemsetAsync(dst_d, 0, ggml_nbytes(dst), stream)); + + const dim3 blocks_dim(WARP_SIZE, 1, 1); + const dim3 blocks_num(std::min((int64_t)4*nsm, (ne + CUDA_COUNT_EQUAL_CHUNK_SIZE - 1)/CUDA_COUNT_EQUAL_CHUNK_SIZE), 1, 1); + + switch (src0->type) { + case GGML_TYPE_I32: { + const int * src0_d = (const int *) src0->data; + const int * src1_d = (const int *) src1->data; + count_equal<<>>(src0_d, src1_d, dst_d, dne, ne); + } break; + default: + GGML_ASSERT(false); + break; + } +} diff --git a/ggml/src/ggml-cuda/count-equal.cuh b/ggml/src/ggml-cuda/count-equal.cuh new file mode 100644 index 000000000..8467da79e --- /dev/null +++ b/ggml/src/ggml-cuda/count-equal.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_COUNT_EQUAL_CHUNK_SIZE 128 + +void ggml_cuda_count_equal(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu index 342f2eb66..5af02c7ec 100644 --- a/ggml/src/ggml-cuda/fattn-tile-f16.cu +++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu @@ -259,7 +259,7 @@ static __global__ void flash_attn_tile_ext_f16( } half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]); - kqsum_j = warp_reduce_sum(kqsum_j); + kqsum_j = warp_reduce_sum((float)kqsum_j); #pragma unroll for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) { diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh index 448a9a905..2ed6509ac 100644 --- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh +++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh @@ -196,7 +196,7 @@ static __global__ void flash_attn_vec_ext_f16( #pragma unroll for (int j = 0; j < ncols; ++j) { half sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_h2[j], Q_i32[j], Q_ds[j]); - sum = warp_reduce_sum(sum); + sum = warp_reduce_sum((float)sum); if (use_logit_softcap) { sum = logit_softcap*tanhf(sum); @@ -265,7 +265,7 @@ static __global__ void flash_attn_vec_ext_f16( #pragma unroll for (int j = 0; j < ncols; ++j) { - kqsum[j] = warp_reduce_sum(kqsum[j]); + kqsum[j] = warp_reduce_sum((float)kqsum[j]); if (threadIdx.x == 0) { kqsum_shared[j][threadIdx.y] = kqsum[j]; } @@ -280,7 +280,7 @@ static __global__ void flash_attn_vec_ext_f16( } kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x]; - kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]); + kqsum[j_VKQ] = warp_reduce_sum((float)kqsum[j_VKQ]); half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ])); if (parallel_blocks == 1) { diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h index 833984190..d3f4bad8c 100644 --- a/ggml/src/ggml-impl.h +++ b/ggml/src/ggml-impl.h @@ -33,6 +33,21 @@ extern "C" { #endif #endif +// +// logging +// + +GGML_ATTRIBUTE_FORMAT(2, 3) +void ggml_log_internal (enum ggml_log_level level, const char * format, ...); +void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data); + +#define GGML_LOG(...) ggml_log_internal(GGML_LOG_LEVEL_NONE , __VA_ARGS__) +#define GGML_LOG_INFO(...) ggml_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__) +#define GGML_LOG_WARN(...) ggml_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__) +#define GGML_LOG_ERROR(...) ggml_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) +#define GGML_LOG_DEBUG(...) ggml_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__) +#define GGML_LOG_CONT(...) ggml_log_internal(GGML_LOG_LEVEL_CONT , __VA_ARGS__) + // bitset typedef uint32_t ggml_bitset_t; diff --git a/ggml/src/ggml-kompute.cpp b/ggml/src/ggml-kompute.cpp index 9cbc57a64..2c926aaee 100644 --- a/ggml/src/ggml-kompute.cpp +++ b/ggml/src/ggml-kompute.cpp @@ -1921,6 +1921,7 @@ ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) { for (const auto & dev : devices) { vec.push_back({ /* .iface = */ ggml_backend_kompute_buffer_type_interface, + /* .device = */ nullptr, /* .context = */ new ggml_backend_kompute_buffer_type_context(dev.index, dev.bufferAlignment, dev.maxAlloc) }); } @@ -1989,11 +1990,8 @@ static struct ggml_backend_i kompute_backend_i = { /* .supports_op = */ ggml_backend_kompute_supports_op, /* .supports_buft = */ ggml_backend_kompute_supports_buft, /* .offload_op = */ NULL, - /* .event_new = */ NULL, - /* .event_free = */ NULL, /* .event_record = */ NULL, /* .event_wait = */ NULL, - /* .event_synchronize = */ NULL, }; static ggml_guid_t ggml_backend_kompute_guid() { @@ -2008,6 +2006,7 @@ ggml_backend_t ggml_backend_kompute_init(int device) { ggml_backend_t kompute_backend = new ggml_backend { /* .guid = */ ggml_backend_kompute_guid(), /* .interface = */ kompute_backend_i, + /* .device = */ nullptr, /* .context = */ s_kompute_context, }; @@ -2017,23 +2016,3 @@ ggml_backend_t ggml_backend_kompute_init(int device) { bool ggml_backend_is_kompute(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid()); } - -static ggml_backend_t ggml_backend_reg_kompute_init(const char * params, void * user_data) { - GGML_UNUSED(params); - return ggml_backend_kompute_init(intptr_t(user_data)); -} - -extern "C" int ggml_backend_kompute_reg_devices(); - -int ggml_backend_kompute_reg_devices() { - auto devices = ggml_vk_available_devices_internal(0); - for (const auto & device : devices) { - ggml_backend_register( - ggml_kompute_format_name(device.index).c_str(), - ggml_backend_reg_kompute_init, - ggml_backend_kompute_buffer_type(device.index), - reinterpret_cast(intptr_t(device.index)) - ); - } - return devices.size(); -} diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m index ef3b7f0e8..172a0f925 100644 --- a/ggml/src/ggml-metal.m +++ b/ggml/src/ggml-metal.m @@ -12,21 +12,77 @@ #define MIN(a, b) ((a) < (b) ? (a) : (b)) #define MAX(a, b) ((a) > (b) ? (a) : (b)) -#ifdef GGML_METAL_NDEBUG -#define GGML_METAL_LOG(...) -#define GGML_METAL_LOG_INFO(...) -#define GGML_METAL_LOG_WARN(...) -#define GGML_METAL_LOG_ERROR(...) -#else -#define GGML_METAL_LOG(...) ggml_metal_log(GGML_LOG_LEVEL_NONE, __VA_ARGS__) -#define GGML_METAL_LOG_INFO(...) ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__) -#define GGML_METAL_LOG_WARN(...) ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__) -#define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) -#define GGML_METAL_LOG_DEBUG(...) ggml_metal_log(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__) -#endif +// max memory buffers that can be mapped to the device +#define GGML_METAL_MAX_BUFFERS 64 + +// max number of MTLCommandBuffer used to submit a graph for processing +#define GGML_METAL_MAX_COMMAND_BUFFERS 8 #define UNUSED(x) (void)(x) +// globals + +// overload of MTLGPUFamilyMetal3 (not available in some environments) +static const NSInteger MTLGPUFamilyMetal3_GGML = 5001; + +// initialized in ggml_backend_metal_reg +static struct ggml_backend_reg g_ggml_backend_metal_reg; +static struct ggml_backend_device g_ggml_backend_metal_device; + +// information about a Metal device +// note: assumes single GPU device - the default one +// TODO: support multiple GPU devices +static struct ggml_backend_metal_device_context { + id mtl_device; + int mtl_device_ref_count; + + bool support_simdgroup_reduction; + bool support_simdgroup_mm; + + char name[128]; +} g_ggml_ctx_dev_main = { + /*.mtl_device =*/ nil, + /*.mtl_device_ref_count =*/ 0, + /*.support_simdgroup_reduction =*/ false, + /*.support_simdgroup_mm =*/ false, + /*.name =*/ "", +}; + +// acquire +static id ggml_backend_metal_device_acq(struct ggml_backend_metal_device_context * ctx) { + assert(ctx != NULL); + + if (ctx->mtl_device == nil) { + ctx->mtl_device = MTLCreateSystemDefaultDevice(); + + ctx->support_simdgroup_reduction = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7]; + ctx->support_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML]; + + ctx->support_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7]; + + strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1); + } + + ctx->mtl_device_ref_count++; + + return ctx->mtl_device; +} + +// release +static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_context * ctx) { + assert(ctx != NULL); + assert(ctx->mtl_device_ref_count > 0); + + ctx->mtl_device_ref_count--; + + if (ctx->mtl_device_ref_count == 0) { + [ctx->mtl_device release]; + ctx->mtl_device = nil; + } +} + +// kernels + struct ggml_metal_kernel { id pipeline; }; @@ -221,19 +277,31 @@ enum ggml_metal_kernel_type { }; struct ggml_backend_metal_context { - int n_cb; - - id device; id queue; dispatch_queue_t d_queue; struct ggml_metal_kernel kernels[GGML_METAL_KERNEL_TYPE_COUNT]; - bool support_simdgroup_reduction; - bool support_simdgroup_mm; + // capture state + bool capture_next_compute; + bool capture_started; - bool should_capture_next_compute; + id capture_scope; + + // command buffer state + int n_cb; // number of extra threads used to submit the command buffers + int n_nodes_0; // number of nodes submitted by the main thread + int n_nodes_1; // remaining number of nodes submitted by the n_cb threads + int n_nodes_per_cb; + + struct ggml_cgraph * gf; + + // the callback given to the thread pool + void (^encode_async)(size_t ith); + + // n_cb command buffers + 1 used by the main thread + id command_buffers[GGML_METAL_MAX_COMMAND_BUFFERS + 1]; // abort ggml_metal_graph_compute if callback returns true ggml_abort_callback abort_callback; @@ -251,51 +319,19 @@ struct ggml_backend_metal_context { @implementation GGMLMetalClass @end -static void ggml_metal_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) { - fprintf(stderr, "%s", msg); - - UNUSED(level); - UNUSED(user_data); -} - -ggml_log_callback ggml_metal_log_callback = ggml_metal_default_log_callback; -void * ggml_metal_log_user_data = NULL; - -GGML_ATTRIBUTE_FORMAT(2, 3) -static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){ - if (ggml_metal_log_callback != NULL) { - va_list args; - va_start(args, format); - char buffer[128]; - int len = vsnprintf(buffer, 128, format, args); - if (len < 128) { - ggml_metal_log_callback(level, buffer, ggml_metal_log_user_data); - } else { - char* buffer2 = malloc(len+1); - va_end(args); - va_start(args, format); - vsnprintf(buffer2, len+1, format, args); - buffer2[len] = 0; - ggml_metal_log_callback(level, buffer2, ggml_metal_log_user_data); - free(buffer2); - } - va_end(args); - } -} - static void * ggml_metal_host_malloc(size_t n) { void * data = NULL; #if TARGET_OS_OSX kern_return_t err = vm_allocate((vm_map_t) mach_task_self(), (void *) &data, n, VM_FLAGS_ANYWHERE); if (err != KERN_SUCCESS) { - GGML_METAL_LOG_ERROR("%s: error: vm_allocate failed\n", __func__); + GGML_LOG_ERROR("%s: error: vm_allocate failed\n", __func__); return NULL; } #else const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n); if (result != 0) { - GGML_METAL_LOG_ERROR("%s: error: posix_memalign failed\n", __func__); + GGML_LOG_ERROR("%s: error: posix_memalign failed\n", __func__); return NULL; } #endif @@ -303,27 +339,26 @@ static void * ggml_metal_host_malloc(size_t n) { return data; } -static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { - GGML_METAL_LOG_INFO("%s: allocating\n", __func__); +static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t dev) { + GGML_LOG_INFO("%s: allocating\n", __func__); #if TARGET_OS_OSX && !GGML_METAL_NDEBUG // Show all the Metal device instances in the system NSArray * devices = MTLCopyAllDevices(); for (id device in devices) { - GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]); + GGML_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]); } [devices release]; // since it was created by a *Copy* C method #endif - // Pick and show default Metal device - id device = MTLCreateSystemDefaultDevice(); - GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]); - - // Configure context + // init context struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context)); - ctx->device = device; - ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS); - ctx->queue = [ctx->device newCommandQueue]; + struct ggml_backend_metal_device_context * ctx_dev = dev->context; + + id device = ggml_backend_metal_device_acq(ctx_dev); + GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]); + + ctx->queue = [device newCommandQueue]; ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); id metal_library; @@ -354,28 +389,28 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { if (try_metallib && path_lib != nil) { // pre-compiled library found NSURL * libURL = [NSURL fileURLWithPath:path_lib]; - GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]); + GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]); - metal_library = [ctx->device newLibraryWithURL:libURL error:&error]; + metal_library = [device newLibraryWithURL:libURL error:&error]; if (error) { - GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; } } else { #if GGML_METAL_EMBED_LIBRARY - GGML_METAL_LOG_INFO("%s: using embedded metal library\n", __func__); + GGML_LOG_INFO("%s: using embedded metal library\n", __func__); extern const char ggml_metallib_start[]; extern const char ggml_metallib_end[]; NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding]; #else - GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); + GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); NSString * path_source; NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"]; - GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil"); + GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil"); if (path_resource) { path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"]; @@ -384,15 +419,15 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { } if (path_source == nil) { - GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); + GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); path_source = @"ggml-metal.metal"; } - GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]); + GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]); NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error]; if (error) { - GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; } #endif // GGML_METAL_EMBED_LIBRARY @@ -406,9 +441,9 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { //[options setFastMathEnabled:false]; - metal_library = [ctx->device newLibraryWithSource:src options:options error:&error]; + metal_library = [device newLibraryWithSource:src options:options error:&error]; if (error) { - GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; } } @@ -416,56 +451,51 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { } // print MTL GPU family: - GGML_METAL_LOG_INFO("%s: GPU name: %s\n", __func__, [[ctx->device name] UTF8String]); - - const NSInteger MTLGPUFamilyMetal3 = 5001; + GGML_LOG_INFO("%s: GPU name: %s\n", __func__, [[device name] UTF8String]); // determine max supported GPU family // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf { for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) { - if ([ctx->device supportsFamily:i]) { - GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i); + if ([device supportsFamily:i]) { + GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i); break; } } for (int i = MTLGPUFamilyCommon1 + 5; i >= MTLGPUFamilyCommon1; --i) { - if ([ctx->device supportsFamily:i]) { - GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyCommon%d (%d)\n", __func__, i - (int) MTLGPUFamilyCommon1 + 1, i); + if ([device supportsFamily:i]) { + GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyCommon%d (%d)\n", __func__, i - (int) MTLGPUFamilyCommon1 + 1, i); break; } } - for (int i = MTLGPUFamilyMetal3 + 5; i >= MTLGPUFamilyMetal3; --i) { - if ([ctx->device supportsFamily:i]) { - GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyMetal%d (%d)\n", __func__, i - (int) MTLGPUFamilyMetal3 + 3, i); + for (int i = MTLGPUFamilyMetal3_GGML + 5; i >= MTLGPUFamilyMetal3_GGML; --i) { + if ([device supportsFamily:i]) { + GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyMetal%d (%d)\n", __func__, i - (int) MTLGPUFamilyMetal3_GGML + 3, i); break; } } } - ctx->support_simdgroup_reduction = [ctx->device supportsFamily:MTLGPUFamilyApple7]; - ctx->support_simdgroup_reduction |= [ctx->device supportsFamily:MTLGPUFamilyMetal3]; + GGML_LOG_INFO("%s: simdgroup reduction support = %s\n", __func__, ctx_dev->support_simdgroup_reduction ? "true" : "false"); + GGML_LOG_INFO("%s: simdgroup matrix mul. support = %s\n", __func__, ctx_dev->support_simdgroup_mm ? "true" : "false"); + GGML_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx_dev->mtl_device.hasUnifiedMemory ? "true" : "false"); - ctx->support_simdgroup_mm = [ctx->device supportsFamily:MTLGPUFamilyApple7]; + ctx->capture_next_compute = false; + ctx->capture_started = false; + ctx->capture_scope = nil; - GGML_METAL_LOG_INFO("%s: simdgroup reduction support = %s\n", __func__, ctx->support_simdgroup_reduction ? "true" : "false"); - GGML_METAL_LOG_INFO("%s: simdgroup matrix mul. support = %s\n", __func__, ctx->support_simdgroup_mm ? "true" : "false"); - GGML_METAL_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false"); - - ctx->should_capture_next_compute = false; + ctx->gf = nil; + ctx->encode_async = nil; + for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) { + ctx->command_buffers[i] = nil; + } #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15) if (@available(macOS 10.12, iOS 16.0, *)) { - GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1e6); - } -#elif TARGET_OS_OSX - if (ctx->device.maxTransferRate != 0) { - GGML_METAL_LOG_INFO("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1e6); - } else { - GGML_METAL_LOG_INFO("%s: maxTransferRate = built-in GPU\n", __func__); + GGML_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, device.recommendedMaxWorkingSetSize / 1e6); } #endif @@ -478,7 +508,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { } /* - GGML_METAL_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \ + GGML_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \ (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \ (int) kernel->pipeline.threadExecutionWidth); \ */ @@ -486,17 +516,20 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { if (supported) { \ struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \ id metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \ - kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \ + kernel->pipeline = [device newComputePipelineStateWithFunction:metal_function error:&error]; \ [metal_function release]; \ if (error) { \ - GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ + GGML_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ [metal_library release]; \ return NULL; \ } \ } else { \ - GGML_METAL_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \ + GGML_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \ } + const bool support_simdgroup_mm = ctx_dev->support_simdgroup_mm; + const bool support_simdgroup_reduction = ctx_dev->support_simdgroup_reduction; + // simd_sum and simd_max requires MTLGPUFamilyApple7 GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); @@ -523,10 +556,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true); @@ -551,101 +584,101 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, ssm_conv_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, ssm_scan_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, support_simdgroup_mm); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32, rope_norm_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16, rope_norm_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, rope_neox_f32, true); @@ -659,14 +692,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, ctx->support_simdgroup_mm); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, ctx->support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, support_simdgroup_mm); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); @@ -686,18 +719,20 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { } [metal_library release]; + return ctx; } static void ggml_metal_free(struct ggml_backend_metal_context * ctx) { - GGML_METAL_LOG_INFO("%s: deallocating\n", __func__); + GGML_LOG_INFO("%s: deallocating\n", __func__); for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) { [ctx->kernels[i].pipeline release]; } + Block_release(ctx->encode_async); + [ctx->queue release]; - [ctx->device release]; dispatch_release(ctx->d_queue); @@ -728,7 +763,7 @@ struct ggml_backend_metal_buffer_context { // Metal buffer based on the host memory pointer // static id ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs) { - //GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach); + //GGML_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach); const int64_t tsize = ggml_nbytes(t); @@ -740,28 +775,31 @@ static id ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs for (int i = 0; i < buf_ctx->n_buffers; ++i) { const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data; - //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size); + //GGML_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size); if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) { *offs = (size_t) ioffs; - //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs); + //GGML_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs); return buf_ctx->buffers[i].metal; } } - GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name); + GGML_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name); return nil; } -static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) { +static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_context * ctx_dev, const struct ggml_tensor * op) { for (size_t i = 0, n = 3; i < n; ++i) { if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) { return false; } } + const bool support_simdgroup_mm = ctx_dev->support_simdgroup_mm; + const bool support_simdgroup_reduction = ctx_dev->support_simdgroup_reduction; + switch (op->op) { case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { @@ -799,7 +837,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx case GGML_OP_SOFT_MAX: case GGML_OP_RMS_NORM: case GGML_OP_GROUP_NORM: - return ctx->support_simdgroup_reduction; + return support_simdgroup_reduction; case GGML_OP_NORM: case GGML_OP_ROPE: return true; @@ -825,13 +863,13 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx if (op->src[0]->ne[0] == 256) { return false; } - return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels + return support_simdgroup_mm; // TODO: over-restricted for vec-kernels case GGML_OP_SSM_CONV: case GGML_OP_SSM_SCAN: return true; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: - return ctx->support_simdgroup_reduction && + return support_simdgroup_reduction && (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32); case GGML_OP_CPY: case GGML_OP_DUP: @@ -874,875 +912,826 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx } } -static enum ggml_status ggml_metal_graph_compute( - struct ggml_backend_metal_context * ctx, - struct ggml_cgraph * gf) { +static void ggml_metal_encode_node( + ggml_backend_t backend, + int idx, + id encoder) { + struct ggml_backend_metal_context * ctx = backend->context; + struct ggml_backend_metal_device_context * ctx_dev = backend->device->context; - @autoreleasepool { - MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor; - edesc.dispatchType = MTLDispatchTypeSerial; + struct ggml_cgraph * gf = ctx->gf; - // create multiple command buffers and enqueue them - // then, we encode the graph into the command buffers in parallel + struct ggml_tensor * node = ggml_graph_node(gf, idx); - const int n_nodes = gf->n_nodes; - const int n_cb = ctx->n_cb; - const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb; + //GGML_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, idx, ggml_op_name(node->op)); - const bool should_capture = ctx->should_capture_next_compute; - if (should_capture) { - ctx->should_capture_next_compute = false; + struct ggml_tensor * src0 = node->src[0]; + struct ggml_tensor * src1 = node->src[1]; + struct ggml_tensor * src2 = node->src[2]; + struct ggml_tensor * dst = node; - MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new]; - descriptor.captureObject = ctx->queue; - - NSError * error = nil; - if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) { - GGML_METAL_LOG_ERROR("%s: error: unable to start capture '%s'\n", __func__, [[error localizedDescription] UTF8String]); - GGML_ABORT("capture failed"); - } + if (ggml_is_empty(dst)) { + return; } - id command_buffer_builder[n_cb]; - for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) { - id command_buffer = [ctx->queue commandBufferWithUnretainedReferences]; - command_buffer_builder[cb_idx] = command_buffer; - - // always enqueue the first two command buffers - // enqueue all of the command buffers if we don't need to abort - if (cb_idx < 2 || ctx->abort_callback == NULL) { - [command_buffer enqueue]; - } + switch (dst->op) { + case GGML_OP_NONE: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_TRANSPOSE: + case GGML_OP_PERMUTE: + { + // noop -> next node + } return; + default: + { + } break; } - const id *command_buffers = command_buffer_builder; + if (!ggml_metal_supports_op(ctx_dev, dst)) { + GGML_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst)); + GGML_ABORT("unsupported op"); + } - dispatch_apply(n_cb, ctx->d_queue, ^(size_t iter) { - const int cb_idx = iter; + const int64_t ne00 = src0 ? src0->ne[0] : 0; + const int64_t ne01 = src0 ? src0->ne[1] : 0; + const int64_t ne02 = src0 ? src0->ne[2] : 0; + const int64_t ne03 = src0 ? src0->ne[3] : 0; - size_t offs_src0 = 0; - size_t offs_src1 = 0; - size_t offs_src2 = 0; - size_t offs_dst = 0; + const uint64_t nb00 = src0 ? src0->nb[0] : 0; + const uint64_t nb01 = src0 ? src0->nb[1] : 0; + const uint64_t nb02 = src0 ? src0->nb[2] : 0; + const uint64_t nb03 = src0 ? src0->nb[3] : 0; - id command_buffer = command_buffers[cb_idx]; - id encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc]; + const int64_t ne10 = src1 ? src1->ne[0] : 0; + const int64_t ne11 = src1 ? src1->ne[1] : 0; + const int64_t ne12 = src1 ? src1->ne[2] : 0; + const int64_t ne13 = src1 ? src1->ne[3] : 0; - const int node_start = (cb_idx + 0) * n_nodes_per_cb; - const int node_end = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes); + const uint64_t nb10 = src1 ? src1->nb[0] : 0; + const uint64_t nb11 = src1 ? src1->nb[1] : 0; + const uint64_t nb12 = src1 ? src1->nb[2] : 0; + const uint64_t nb13 = src1 ? src1->nb[3] : 0; - for (int i = node_start; i < node_end; ++i) { - if (i == -1) { - [encoder memoryBarrierWithScope:MTLBarrierScopeBuffers]; - continue; - } + const int64_t ne20 = src2 ? src2->ne[0] : 0; + const int64_t ne21 = src2 ? src2->ne[1] : 0; + const int64_t ne22 = src2 ? src2->ne[2] : 0; GGML_UNUSED(ne22); + const int64_t ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23); - //GGML_METAL_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op)); + const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20); + const uint64_t nb21 = src2 ? src2->nb[1] : 0; + const uint64_t nb22 = src2 ? src2->nb[2] : 0; + const uint64_t nb23 = src2 ? src2->nb[3] : 0; - struct ggml_tensor * src0 = gf->nodes[i]->src[0]; - struct ggml_tensor * src1 = gf->nodes[i]->src[1]; - struct ggml_tensor * src2 = gf->nodes[i]->src[2]; - struct ggml_tensor * dst = gf->nodes[i]; + const int64_t ne0 = dst ? dst->ne[0] : 0; + const int64_t ne1 = dst ? dst->ne[1] : 0; + const int64_t ne2 = dst ? dst->ne[2] : 0; + const int64_t ne3 = dst ? dst->ne[3] : 0; - if (ggml_is_empty(dst)) { - continue; - } + const uint64_t nb0 = dst ? dst->nb[0] : 0; + const uint64_t nb1 = dst ? dst->nb[1] : 0; + const uint64_t nb2 = dst ? dst->nb[2] : 0; + const uint64_t nb3 = dst ? dst->nb[3] : 0; - switch (dst->op) { - case GGML_OP_NONE: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_TRANSPOSE: - case GGML_OP_PERMUTE: - { - // noop -> next node - } continue; + const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT; + const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT; + const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT; + + size_t offs_src0 = 0; + size_t offs_src1 = 0; + size_t offs_src2 = 0; + size_t offs_dst = 0; + + id id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil; + id id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil; + id id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil; + id id_dst = dst ? ggml_metal_get_buffer(dst, &offs_dst) : nil; + + //GGML_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op)); + //if (src0) { + // GGML_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02, + // ggml_is_contiguous(src0), src0->name); + //} + //if (src1) { + // GGML_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12, + // ggml_is_contiguous(src1), src1->name); + //} + //if (dst) { + // GGML_LOG_INFO("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2, + // dst->name); + //} + + id device = ctx_dev->mtl_device; + + switch (dst->op) { + case GGML_OP_CONCAT: + { + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline; + + const int32_t dim = ((const int32_t *) dst->op_params)[0]; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9]; + [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10]; + [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11]; + [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12]; + [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13]; + [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17]; + [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21]; + [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25]; + [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26]; + [encoder setBytes:&dim length:sizeof(dim) atIndex:27]; + + const int nth = MIN(1024, ne0); + + [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_ADD: + case GGML_OP_SUB: + case GGML_OP_MUL: + case GGML_OP_DIV: + { + GGML_ASSERT(src0t == GGML_TYPE_F32); + GGML_ASSERT(src1t == GGML_TYPE_F32); + + const size_t offs = 0; + + bool bcast_row = false; + + int64_t nb = ne00; // used by the "row" kernels + + id pipeline = nil; + + if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) { + GGML_ASSERT(ggml_is_contiguous(src0)); + + // src1 is a row + GGML_ASSERT(ne11 == 1); + + nb = ne00 / 4; + switch (dst->op) { + case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break; + case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break; + case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break; + case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break; + default: GGML_ABORT("fatal error"); + } + + bcast_row = true; + } else { + switch (dst->op) { + case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break; + case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB].pipeline; break; + case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break; + case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break; + default: GGML_ABORT("fatal error"); + } + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9]; + [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10]; + [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11]; + [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12]; + [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13]; + [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17]; + [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21]; + [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25]; + [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26]; + [encoder setBytes:&offs length:sizeof(offs) atIndex:27]; + [encoder setBytes:&nb length:sizeof(nb) atIndex:28]; + + if (bcast_row) { + const int64_t n = ggml_nelements(dst)/4; + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } else { + const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); + + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } + } break; + case GGML_OP_REPEAT: + { + id pipeline; + + switch (src0t) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break; + case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break; + case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break; + default: GGML_ABORT("fatal error"); + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; + [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; + [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; + [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; + + const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); + + [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_ACC: + { + GGML_ASSERT(src0t == GGML_TYPE_F32); + GGML_ASSERT(src1t == GGML_TYPE_F32); + GGML_ASSERT(dstt == GGML_TYPE_F32); + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + + const size_t pnb1 = ((const int32_t *) dst->op_params)[0]; + const size_t pnb2 = ((const int32_t *) dst->op_params)[1]; + const size_t pnb3 = ((const int32_t *) dst->op_params)[2]; + const size_t offs = ((const int32_t *) dst->op_params)[3]; + + const bool inplace = (bool) ((const int32_t *) dst->op_params)[4]; + + if (!inplace) { + // run a separete kernel to cpy src->dst + // not sure how to avoid this + // TODO: make a simpler cpy_bytes kernel + + const id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12]; + [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14]; + [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15]; + [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16]; + [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17]; + + const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00); + + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } + + const id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; + [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8]; + [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9]; + [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10]; + [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11]; + [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12]; + [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13]; + [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17]; + [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21]; + [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23]; + [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24]; + [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25]; + [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26]; + [encoder setBytes:&offs length:sizeof(offs) atIndex:27]; + + const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00); + + [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_SCALE: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + + float scale; + memcpy(&scale, dst->op_params, sizeof(scale)); + + int64_t n = ggml_nelements(dst); + + id pipeline = nil; + + if (n % 4 == 0) { + n /= 4; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE_4].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE].pipeline; + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&scale length:sizeof(scale) atIndex:2]; + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_CLAMP: + { + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline; + + float min; + float max; + memcpy(&min, ((const int32_t *) dst->op_params) + 0, sizeof(float)); + memcpy(&max, ((const int32_t *) dst->op_params) + 1, sizeof(float)); + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&min length:sizeof(min) atIndex:2]; + [encoder setBytes:&max length:sizeof(max) atIndex:3]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_UNARY: + switch (ggml_get_unary_op(node)) { + // we are not taking into account the strides, so for now require contiguous tensors + GGML_ASSERT(ggml_is_contiguous(src0)); + + case GGML_UNARY_OP_TANH: + { + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TANH].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_UNARY_OP_RELU: + { + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RELU].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_UNARY_OP_SIGMOID: + { + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIGMOID].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_UNARY_OP_GELU: + { + int64_t n = ggml_nelements(dst); + + id pipeline = nil; + + if (n % 4 == 0) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_4].pipeline; + n /= 4; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline; + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_UNARY_OP_GELU_QUICK: + { + int64_t n = ggml_nelements(dst); + + id pipeline = nil; + + if (n % 4 == 0) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK_4].pipeline; + n /= 4; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline; + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_UNARY_OP_SILU: + { + int64_t n = ggml_nelements(dst); + + id pipeline = nil; + + if (n % 4 == 0) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU_4].pipeline; + n /= 4; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline; + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; default: - { - } break; - } - - if (!ggml_metal_supports_op(ctx, dst)) { - GGML_METAL_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst)); - GGML_ABORT("unsupported op"); - } - - if (should_capture) { - [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(dst) encoding:NSUTF8StringEncoding]]; - } - - const int64_t ne00 = src0 ? src0->ne[0] : 0; - const int64_t ne01 = src0 ? src0->ne[1] : 0; - const int64_t ne02 = src0 ? src0->ne[2] : 0; - const int64_t ne03 = src0 ? src0->ne[3] : 0; - - const uint64_t nb00 = src0 ? src0->nb[0] : 0; - const uint64_t nb01 = src0 ? src0->nb[1] : 0; - const uint64_t nb02 = src0 ? src0->nb[2] : 0; - const uint64_t nb03 = src0 ? src0->nb[3] : 0; - - const int64_t ne10 = src1 ? src1->ne[0] : 0; - const int64_t ne11 = src1 ? src1->ne[1] : 0; - const int64_t ne12 = src1 ? src1->ne[2] : 0; - const int64_t ne13 = src1 ? src1->ne[3] : 0; - - const uint64_t nb10 = src1 ? src1->nb[0] : 0; - const uint64_t nb11 = src1 ? src1->nb[1] : 0; - const uint64_t nb12 = src1 ? src1->nb[2] : 0; - const uint64_t nb13 = src1 ? src1->nb[3] : 0; - - const int64_t ne20 = src2 ? src2->ne[0] : 0; - const int64_t ne21 = src2 ? src2->ne[1] : 0; - const int64_t ne22 = src2 ? src2->ne[2] : 0; GGML_UNUSED(ne22); - const int64_t ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23); - - const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20); - const uint64_t nb21 = src2 ? src2->nb[1] : 0; - const uint64_t nb22 = src2 ? src2->nb[2] : 0; - const uint64_t nb23 = src2 ? src2->nb[3] : 0; - - const int64_t ne0 = dst ? dst->ne[0] : 0; - const int64_t ne1 = dst ? dst->ne[1] : 0; - const int64_t ne2 = dst ? dst->ne[2] : 0; - const int64_t ne3 = dst ? dst->ne[3] : 0; - - const uint64_t nb0 = dst ? dst->nb[0] : 0; - const uint64_t nb1 = dst ? dst->nb[1] : 0; - const uint64_t nb2 = dst ? dst->nb[2] : 0; - const uint64_t nb3 = dst ? dst->nb[3] : 0; - - const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT; - const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT; - const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT; - - id id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil; - id id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil; - id id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil; - id id_dst = dst ? ggml_metal_get_buffer(dst, &offs_dst) : nil; - - //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op)); - //if (src0) { - // GGML_METAL_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02, - // ggml_is_contiguous(src0), src0->name); - //} - //if (src1) { - // GGML_METAL_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12, - // ggml_is_contiguous(src1), src1->name); - //} - //if (dst) { - // GGML_METAL_LOG_INFO("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2, - // dst->name); - //} - - switch (dst->op) { - case GGML_OP_CONCAT: - { - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline; - - const int32_t dim = ((int32_t *) dst->op_params)[0]; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13]; - [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26]; - [encoder setBytes:&dim length:sizeof(dim) atIndex:27]; - - const int nth = MIN(1024, ne0); - - [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_ADD: - case GGML_OP_SUB: - case GGML_OP_MUL: - case GGML_OP_DIV: - { - GGML_ASSERT(src0t == GGML_TYPE_F32); - GGML_ASSERT(src1t == GGML_TYPE_F32); - - const size_t offs = 0; - - bool bcast_row = false; - - int64_t nb = ne00; // used by the "row" kernels - - id pipeline = nil; - - if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) { - GGML_ASSERT(ggml_is_contiguous(src0)); - - // src1 is a row - GGML_ASSERT(ne11 == 1); - - nb = ne00 / 4; - switch (dst->op) { - case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break; - case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break; - case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break; - case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break; - default: GGML_ABORT("fatal error"); - } - - bcast_row = true; - } else { - switch (dst->op) { - case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break; - case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB].pipeline; break; - case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break; - case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break; - default: GGML_ABORT("fatal error"); - } - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13]; - [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26]; - [encoder setBytes:&offs length:sizeof(offs) atIndex:27]; - [encoder setBytes:&nb length:sizeof(nb) atIndex:28]; - - if (bcast_row) { - const int64_t n = ggml_nelements(dst)/4; - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } else { - const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); - - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } - } break; - case GGML_OP_REPEAT: - { - id pipeline; - - switch (src0t) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break; - case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break; - case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break; - default: GGML_ABORT("fatal error"); - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; - - const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); - - [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_ACC: - { - GGML_ASSERT(src0t == GGML_TYPE_F32); - GGML_ASSERT(src1t == GGML_TYPE_F32); - GGML_ASSERT(dstt == GGML_TYPE_F32); - - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(src1)); - - const size_t pnb1 = ((int32_t *) dst->op_params)[0]; - const size_t pnb2 = ((int32_t *) dst->op_params)[1]; - const size_t pnb3 = ((int32_t *) dst->op_params)[2]; - const size_t offs = ((int32_t *) dst->op_params)[3]; - - const bool inplace = (bool) ((int32_t *) dst->op_params)[4]; - - if (!inplace) { - // run a separete kernel to cpy src->dst - // not sure how to avoid this - // TODO: make a simpler cpy_bytes kernel - - const id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17]; - - const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00); - - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } - - const id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; - [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8]; - [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9]; - [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13]; - [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23]; - [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24]; - [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25]; - [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26]; - [encoder setBytes:&offs length:sizeof(offs) atIndex:27]; - - const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00); - - [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_SCALE: - { - GGML_ASSERT(ggml_is_contiguous(src0)); - - float scale; - memcpy(&scale, dst->op_params, sizeof(scale)); - - int64_t n = ggml_nelements(dst); - - id pipeline = nil; - - if (n % 4 == 0) { - n /= 4; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE_4].pipeline; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE].pipeline; - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&scale length:sizeof(scale) atIndex:2]; - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_CLAMP: - { - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline; - - float min; - float max; - memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float)); - memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float)); - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&min length:sizeof(min) atIndex:2]; - [encoder setBytes:&max length:sizeof(max) atIndex:3]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_UNARY: - switch (ggml_get_unary_op(gf->nodes[i])) { - // we are not taking into account the strides, so for now require contiguous tensors - GGML_ASSERT(ggml_is_contiguous(src0)); - - case GGML_UNARY_OP_TANH: - { - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TANH].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_UNARY_OP_RELU: - { - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RELU].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_UNARY_OP_SIGMOID: - { - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIGMOID].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_UNARY_OP_GELU: - { - int64_t n = ggml_nelements(dst); - - id pipeline = nil; - - if (n % 4 == 0) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_4].pipeline; - n /= 4; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline; - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_UNARY_OP_GELU_QUICK: - { - int64_t n = ggml_nelements(dst); - - id pipeline = nil; - - if (n % 4 == 0) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK_4].pipeline; - n /= 4; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline; - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_UNARY_OP_SILU: - { - int64_t n = ggml_nelements(dst); - - id pipeline = nil; - - if (n % 4 == 0) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU_4].pipeline; - n /= 4; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline; - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - default: - { - GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op)); - GGML_ABORT("fatal error"); - } - } break; - case GGML_OP_SQR: - { - GGML_ASSERT(ggml_is_contiguous(src0)); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQR].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_SQRT: - { - GGML_ASSERT(ggml_is_contiguous(src0)); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQRT].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_SIN: - { - GGML_ASSERT(ggml_is_contiguous(src0)); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIN].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_COS: - { - GGML_ASSERT(ggml_is_contiguous(src0)); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_COS].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_SUM_ROWS: - { - GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12]; - [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:18]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:19]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:20]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:21]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:22]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:23]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:24]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:25]; - - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_SOFT_MAX: - { - GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); - - int nth = 32; // SIMD width - - id pipeline = nil; - - const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16); - - if (ne00%4 == 0) { - while (nth < ne00/4 && nth*ne01*ne02*ne03 < 256) { - nth *= 2; - } - if (use_f16) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4].pipeline; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline; - } - } else { - while (nth < ne00 && nth*ne01*ne02*ne03 < 256) { - nth *= 2; - } - if (use_f16) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16].pipeline; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32].pipeline; - } - } - - float scale; - float max_bias; - - memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale)); - memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias)); - - const int64_t nrows_x = ggml_nrows(src0); - const int64_t nrows_y = src0->ne[1]; - - const uint32_t n_head = nrows_x/nrows_y; - const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head)); - - const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); - const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - if (id_src1) { - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - } else { - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; - } - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&scale length:sizeof(scale) atIndex:6]; - [encoder setBytes:&max_bias length:sizeof(max_bias) atIndex:7]; - [encoder setBytes:&m0 length:sizeof(m0) atIndex:8]; - [encoder setBytes:&m1 length:sizeof(m1) atIndex:9]; - [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10]; - [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; - - [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_DIAG_MASK_INF: - { - const int n_past = ((int32_t *)(dst->op_params))[0]; - - id pipeline = nil; - - if (ne00%8 == 0) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8].pipeline; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline; - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&n_past length:sizeof(int) atIndex:4]; - - if (ne00%8 == 0) { - [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } - else { - [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } - } break; - case GGML_OP_SSM_CONV: - { - GGML_ASSERT(src0t == GGML_TYPE_F32); - GGML_ASSERT(src1t == GGML_TYPE_F32); - - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(src1)); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:11]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:12]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:15]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:16]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:17]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:18]; - - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_SSM_SCAN: - { - struct ggml_tensor * src3 = gf->nodes[i]->src[3]; - struct ggml_tensor * src4 = gf->nodes[i]->src[4]; - struct ggml_tensor * src5 = gf->nodes[i]->src[5]; - - GGML_ASSERT(src3); - GGML_ASSERT(src4); - GGML_ASSERT(src5); - - size_t offs_src3 = 0; - size_t offs_src4 = 0; - size_t offs_src5 = 0; - - id id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil; - id id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil; - id id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil; - - const int64_t ne30 = src3->ne[0]; GGML_UNUSED(ne30); - const int64_t ne31 = src3->ne[1]; GGML_UNUSED(ne31); - - const uint64_t nb30 = src3->nb[0]; - const uint64_t nb31 = src3->nb[1]; - - const int64_t ne40 = src4->ne[0]; GGML_UNUSED(ne40); - const int64_t ne41 = src4->ne[1]; GGML_UNUSED(ne41); - const int64_t ne42 = src4->ne[2]; GGML_UNUSED(ne42); - - const uint64_t nb40 = src4->nb[0]; - const uint64_t nb41 = src4->nb[1]; - const uint64_t nb42 = src4->nb[2]; - - const int64_t ne50 = src5->ne[0]; GGML_UNUSED(ne50); - const int64_t ne51 = src5->ne[1]; GGML_UNUSED(ne51); - const int64_t ne52 = src5->ne[2]; GGML_UNUSED(ne52); - - const uint64_t nb50 = src5->nb[0]; - const uint64_t nb51 = src5->nb[1]; - const uint64_t nb52 = src5->nb[2]; - - const int64_t d_state = ne00; - const int64_t d_inner = ne01; - const int64_t n_seq_tokens = ne11; - const int64_t n_seqs = ne02; - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; - [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; - [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4]; - [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:6]; - - [encoder setBytes:&d_state length:sizeof(d_state) atIndex:7]; - [encoder setBytes:&d_inner length:sizeof(d_inner) atIndex:8]; - [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9]; - [encoder setBytes:&n_seqs length:sizeof(n_seqs) atIndex:10]; - - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; - [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18]; - [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19]; - [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20]; - [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21]; - [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22]; - [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23]; - [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24]; - [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25]; - [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26]; - [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27]; - [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28]; - - [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_MUL_MAT: - { - GGML_ASSERT(ne00 == ne10); - - GGML_ASSERT(ne12 % ne02 == 0); - GGML_ASSERT(ne13 % ne03 == 0); - - const uint r2 = ne12/ne02; - const uint r3 = ne13/ne03; - - // find the break-even point where the matrix-matrix kernel becomes more efficient compared - // to the matrix-vector kernel - int ne11_mm_min = 1; + { + GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op)); + GGML_ABORT("fatal error"); + } + } break; + case GGML_OP_SQR: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQR].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_SQRT: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQRT].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_SIN: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIN].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_COS: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_COS].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_SUM_ROWS: + { + GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; + [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; + [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10]; + [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11]; + [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12]; + [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; + [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:18]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:19]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:20]; + [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:21]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:22]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:23]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:24]; + [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:25]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_SOFT_MAX: + { + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); + + int nth = 32; // SIMD width + + id pipeline = nil; + + const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16); + + if (ne00%4 == 0) { + while (nth < ne00/4 && nth*ne01*ne02*ne03 < 256) { + nth *= 2; + } + if (use_f16) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline; + } + } else { + while (nth < ne00 && nth*ne01*ne02*ne03 < 256) { + nth *= 2; + } + if (use_f16) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32].pipeline; + } + } + + float scale; + float max_bias; + + memcpy(&scale, ((const int32_t *) dst->op_params) + 0, sizeof(scale)); + memcpy(&max_bias, ((const int32_t *) dst->op_params) + 1, sizeof(max_bias)); + + const int64_t nrows_x = ggml_nrows(src0); + const int64_t nrows_y = src0->ne[1]; + + const uint32_t n_head = nrows_x/nrows_y; + const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + if (id_src1) { + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + } else { + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; + } + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; + [encoder setBytes:&scale length:sizeof(scale) atIndex:6]; + [encoder setBytes:&max_bias length:sizeof(max_bias) atIndex:7]; + [encoder setBytes:&m0 length:sizeof(m0) atIndex:8]; + [encoder setBytes:&m1 length:sizeof(m1) atIndex:9]; + [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10]; + [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_DIAG_MASK_INF: + { + const int n_past = ((const int32_t *)(dst->op_params))[0]; + + id pipeline = nil; + + if (ne00%8 == 0) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline; + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; + [encoder setBytes:&n_past length:sizeof(int) atIndex:4]; + + if (ne00%8 == 0) { + [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } + else { + [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } + } break; + case GGML_OP_SSM_CONV: + { + GGML_ASSERT(src0t == GGML_TYPE_F32); + GGML_ASSERT(src1t == GGML_TYPE_F32); + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; + [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9]; + [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:11]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:12]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:15]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:16]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:17]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:18]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_SSM_SCAN: + { + struct ggml_tensor * src3 = node->src[3]; + struct ggml_tensor * src4 = node->src[4]; + struct ggml_tensor * src5 = node->src[5]; + + GGML_ASSERT(src3); + GGML_ASSERT(src4); + GGML_ASSERT(src5); + + size_t offs_src3 = 0; + size_t offs_src4 = 0; + size_t offs_src5 = 0; + + id id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil; + id id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil; + id id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil; + + const int64_t ne30 = src3->ne[0]; GGML_UNUSED(ne30); + const int64_t ne31 = src3->ne[1]; GGML_UNUSED(ne31); + + const uint64_t nb30 = src3->nb[0]; + const uint64_t nb31 = src3->nb[1]; + + const int64_t ne40 = src4->ne[0]; GGML_UNUSED(ne40); + const int64_t ne41 = src4->ne[1]; GGML_UNUSED(ne41); + const int64_t ne42 = src4->ne[2]; GGML_UNUSED(ne42); + + const uint64_t nb40 = src4->nb[0]; + const uint64_t nb41 = src4->nb[1]; + const uint64_t nb42 = src4->nb[2]; + + const int64_t ne50 = src5->ne[0]; GGML_UNUSED(ne50); + const int64_t ne51 = src5->ne[1]; GGML_UNUSED(ne51); + const int64_t ne52 = src5->ne[2]; GGML_UNUSED(ne52); + + const uint64_t nb50 = src5->nb[0]; + const uint64_t nb51 = src5->nb[1]; + const uint64_t nb52 = src5->nb[2]; + + const int64_t d_state = ne00; + const int64_t d_inner = ne01; + const int64_t n_seq_tokens = ne11; + const int64_t n_seqs = ne02; + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; + [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4]; + [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:6]; + + [encoder setBytes:&d_state length:sizeof(d_state) atIndex:7]; + [encoder setBytes:&d_inner length:sizeof(d_inner) atIndex:8]; + [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9]; + [encoder setBytes:&n_seqs length:sizeof(n_seqs) atIndex:10]; + + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; + [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; + [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18]; + [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19]; + [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20]; + [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21]; + [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22]; + [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23]; + [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24]; + [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25]; + [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26]; + [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27]; + [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28]; + + [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_MUL_MAT: + { + GGML_ASSERT(ne00 == ne10); + + GGML_ASSERT(ne12 % ne02 == 0); + GGML_ASSERT(ne13 % ne03 == 0); + + const uint r2 = ne12/ne02; + const uint r3 = ne13/ne03; + + // find the break-even point where the matrix-matrix kernel becomes more efficient compared + // to the matrix-vector kernel + int ne11_mm_min = 1; #if 0 - // the numbers below are measured on M2 Ultra for 7B and 13B models - // these numbers do not translate to other devices or model sizes - // TODO: need to find a better approach - if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) { + // the numbers below are measured on M2 Ultra for 7B and 13B models + // these numbers do not translate to other devices or model sizes + // TODO: need to find a better approach + if ([device.name isEqualToString:@"Apple M2 Ultra"]) { switch (src0t) { case GGML_TYPE_F16: ne11_mm_min = 2; break; case GGML_TYPE_Q8_0: ne11_mm_min = 7; break; @@ -1762,12 +1751,12 @@ static enum ggml_status ggml_metal_graph_compute( // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel - if ([ctx->device supportsFamily:MTLGPUFamilyApple7] && - !ggml_is_transposed(src0) && - !ggml_is_transposed(src1) && - src1t == GGML_TYPE_F32 && - ne00 % 32 == 0 && ne00 >= 64 && - (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) { + if ([device supportsFamily:MTLGPUFamilyApple7] && + !ggml_is_transposed(src0) && + !ggml_is_transposed(src1) && + src1t == GGML_TYPE_F32 && + ne00 % 32 == 0 && ne00 >= 64 && + (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) { //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); // some Metal matrix data types require aligned pointers @@ -1974,7 +1963,7 @@ static enum ggml_status ggml_metal_graph_compute( } break; default: { - GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t); + GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t); GGML_ABORT("not implemented"); } }; @@ -2001,8 +1990,8 @@ static enum ggml_status ggml_metal_graph_compute( [encoder setBytes:&r3 length:sizeof(r3) atIndex:18]; if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 || - src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || - src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) { + src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || + src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) { @@ -2036,1041 +2025,1120 @@ static enum ggml_status ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } } - } break; - case GGML_OP_MUL_MAT_ID: - { - const int n_as = src0->ne[2]; - - // src2 = ids - const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t); - - GGML_ASSERT(src2t == GGML_TYPE_I32); - - GGML_ASSERT(!ggml_is_transposed(src0)); - GGML_ASSERT(!ggml_is_transposed(src1)); - - GGML_ASSERT(src1t == GGML_TYPE_F32); - - // find the break-even point where the matrix-matrix kernel becomes more efficient compared - // to the matrix-vector kernel - // ne20 = n_used_experts - // ne21 = n_rows - const int dst_rows = ne20*ne21; - const int dst_rows_min = n_as; - const int dst_rows_max = (ctx->device.maxThreadgroupMemoryLength - 32 - 8192)/4; - - // max size of the rowids array in the kernel shared buffer - GGML_ASSERT(dst_rows <= dst_rows_max); - - // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs - // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel - // !!! - // TODO: for now, always use mat-vec kernels until we figure out how to improve the - // indirect matrix multiplication - // !!! - if ([ctx->device supportsFamily:MTLGPUFamilyApple7] && - ne00 % 32 == 0 && ne00 >= 64 && - dst_rows > dst_rows_min) { - - // some Metal matrix data types require aligned pointers - // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5) - switch (src0->type) { - case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break; - case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break; - default: break; - } - - id pipeline = nil; - - switch (src0->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32 ].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32 ].pipeline; break; - case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32 ].pipeline; break; - case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32 ].pipeline; break; - case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32 ].pipeline; break; - case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32 ].pipeline; break; - case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32 ].pipeline; break; - case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32 ].pipeline; break; - case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32 ].pipeline; break; - case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32 ].pipeline; break; - case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32 ].pipeline; break; - case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32 ].pipeline; break; - case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break; - case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break; - case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break; - case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32 ].pipeline; break; - case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32 ].pipeline; break; - case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break; - case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32 ].pipeline; break; - case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break; - case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break; - default: GGML_ABORT("MUL_MAT_ID not implemented"); - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3]; - [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4]; - [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5]; - [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:8]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:9]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:10]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12]; - [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:17]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:18]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:19]; - - [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0]; - - [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; - } else { - int nth0 = 32; - int nth1 = 1; - int nrows = 1; - //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); - - id pipeline = nil; - - // use custom matrix x vector kernel - switch (src0t) { - case GGML_TYPE_F32: - { - GGML_ASSERT(src1t == GGML_TYPE_F32); - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32].pipeline; - } break; - case GGML_TYPE_F16: - { - GGML_ASSERT(src1t == GGML_TYPE_F32); - nth0 = 32; - nth1 = 1; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32].pipeline; - } break; - case GGML_TYPE_Q4_0: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32].pipeline; - } break; - case GGML_TYPE_Q4_1: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32].pipeline; - } break; - case GGML_TYPE_Q5_0: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32].pipeline; - } break; - case GGML_TYPE_Q5_1: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32].pipeline; - } break; - case GGML_TYPE_Q8_0: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline; - } break; - case GGML_TYPE_Q2_K: - { - nth0 = 2; - nth1 = 32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32].pipeline; - } break; - case GGML_TYPE_Q3_K: - { - nth0 = 2; - nth1 = 32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32].pipeline; - } break; - case GGML_TYPE_Q4_K: - { - nth0 = 4; //1; - nth1 = 8; //32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32].pipeline; - } break; - case GGML_TYPE_Q5_K: - { - nth0 = 2; - nth1 = 32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32].pipeline; - } break; - case GGML_TYPE_Q6_K: - { - nth0 = 2; - nth1 = 32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32].pipeline; - } break; - case GGML_TYPE_IQ2_XXS: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32].pipeline; - } break; - case GGML_TYPE_IQ2_XS: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32].pipeline; - } break; - case GGML_TYPE_IQ3_XXS: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline; - } break; - case GGML_TYPE_IQ3_S: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline; - } break; - case GGML_TYPE_IQ2_S: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline; - } break; - case GGML_TYPE_IQ1_S: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline; - } break; - case GGML_TYPE_IQ1_M: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline; - } break; - case GGML_TYPE_IQ4_NL: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline; - } break; - case GGML_TYPE_IQ4_XS: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline; - } break; - default: - { - GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t); - GGML_ABORT("not implemented"); - } - }; - - if (ggml_is_quantized(src0t)) { - GGML_ASSERT(ne00 >= nth0*nth1); - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3]; - [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4]; - [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5]; - [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15]; - [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:20]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:21]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:22]; - - const int64_t _ne1 = 1; - const int tgz = dst_rows; - - if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 || - src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || - src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) { - const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) { - const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) { - const int mem_size = 32*sizeof(float); - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q4_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q3_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q5_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q6_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } else { - const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1 - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - } - } break; - case GGML_OP_GET_ROWS: - { - id pipeline = nil; - - switch (src0->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F32 ].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F16 ].pipeline; break; - case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0 ].pipeline; break; - case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1 ].pipeline; break; - case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0 ].pipeline; break; - case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1 ].pipeline; break; - case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0 ].pipeline; break; - case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K ].pipeline; break; - case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K ].pipeline; break; - case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K ].pipeline; break; - case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K ].pipeline; break; - case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K ].pipeline; break; - case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break; - case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break; - case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break; - case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S ].pipeline; break; - case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S ].pipeline; break; - case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break; - case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M ].pipeline; break; - case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break; - case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break; - case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break; - default: GGML_ABORT("not implemented"); - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5]; - [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6]; - [encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7]; - [encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:10]; - - [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)]; - } break; - case GGML_OP_RMS_NORM: - { - GGML_ASSERT(ne00 % 4 == 0); - GGML_ASSERT(ggml_is_contiguous_1(src0)); - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - int nth = 32; // SIMD width - - while (nth < ne00/4 && nth < 1024) { - nth *= 2; - } - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; - [encoder setBytes:&eps length:sizeof( float) atIndex:4]; - [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; - - const int64_t nrows = ggml_nrows(src0); - - [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_GROUP_NORM: - { - GGML_ASSERT(ne00 % 4 == 0); - GGML_ASSERT(ggml_is_contiguous(src0)); - - float eps; - memcpy(&eps, dst->op_params + 1, sizeof(float)); - - const int32_t n_groups = ((int32_t *) dst->op_params)[0]; - - int nth = 32; // SIMD width - - //while (nth < ne00/4 && nth < 1024) { - // nth *= 2; - //} - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; - [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:5]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:6]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8]; - [encoder setBytes:&eps length:sizeof( float) atIndex:9]; - [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; - - [encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_NORM: - { - GGML_ASSERT(ggml_is_contiguous_1(src0)); - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - const int nth = MIN(256, ne00); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NORM].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; - [encoder setBytes:&eps length:sizeof( float) atIndex:4]; - [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0]; - - const int64_t nrows = ggml_nrows(src0); - - [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_ROPE: - { - GGML_ASSERT(ne10 == ne02); - - const int nth = MIN(1024, ne00); - - const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal - const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; - - float freq_base; - float freq_scale; - float ext_factor; - float attn_factor; - float beta_fast; - float beta_slow; - - memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); - memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); - memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); - memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); - - const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; - - id pipeline = nil; - - if (!is_neox) { - switch (src0->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break; - default: GGML_ABORT("fatal error"); - }; - } else { - switch (src0->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break; - default: GGML_ABORT("fatal error"); - }; - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - if (id_src2 != nil) { - [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; - } else { - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:2]; - } - [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:4]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6]; - [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7]; - [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:8]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:10]; - [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:11]; - [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:12]; - [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:13]; - [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:14]; - [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:15]; - [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:16]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:17]; - [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:18]; - [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:19]; - [encoder setBytes:&n_past length:sizeof( int) atIndex:20]; - [encoder setBytes:&n_dims length:sizeof( int) atIndex:21]; - [encoder setBytes:&n_ctx_orig length:sizeof( int) atIndex:22]; - [encoder setBytes:&freq_base length:sizeof( float) atIndex:23]; - [encoder setBytes:&freq_scale length:sizeof( float) atIndex:24]; - [encoder setBytes:&ext_factor length:sizeof( float) atIndex:25]; - [encoder setBytes:&attn_factor length:sizeof( float) atIndex:26]; - [encoder setBytes:&beta_fast length:sizeof( float) atIndex:27]; - [encoder setBytes:&beta_slow length:sizeof( float) atIndex:28]; - - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_IM2COL: - { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32); - - const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; - - const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; - - const int32_t N = src1->ne[is_2D ? 3 : 2]; - const int32_t IC = src1->ne[is_2D ? 2 : 1]; - const int32_t IH = is_2D ? src1->ne[1] : 1; - const int32_t IW = src1->ne[0]; - - const int32_t KH = is_2D ? src0->ne[1] : 1; - const int32_t KW = src0->ne[0]; - - const int32_t OH = is_2D ? dst->ne[2] : 1; - const int32_t OW = dst->ne[1]; - - const int32_t CHW = IC * KH * KW; - - const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; - const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; - - id pipeline = nil; - - switch (dst->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break; - default: GGML_ABORT("fatal error"); - }; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ofs0 length:sizeof( int32_t) atIndex:2]; - [encoder setBytes:&ofs1 length:sizeof( int32_t) atIndex:3]; - [encoder setBytes:&IW length:sizeof( int32_t) atIndex:4]; - [encoder setBytes:&IH length:sizeof( int32_t) atIndex:5]; - [encoder setBytes:&CHW length:sizeof( int32_t) atIndex:6]; - [encoder setBytes:&s0 length:sizeof( int32_t) atIndex:7]; - [encoder setBytes:&s1 length:sizeof( int32_t) atIndex:8]; - [encoder setBytes:&p0 length:sizeof( int32_t) atIndex:9]; - [encoder setBytes:&p1 length:sizeof( int32_t) atIndex:10]; - [encoder setBytes:&d0 length:sizeof( int32_t) atIndex:11]; - [encoder setBytes:&d1 length:sizeof( int32_t) atIndex:12]; - - [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)]; - } break; - case GGML_OP_UPSCALE: - { - GGML_ASSERT(src0->type == GGML_TYPE_F32); - - const float sf0 = (float)ne0/src0->ne[0]; - const float sf1 = (float)ne1/src0->ne[1]; - const float sf2 = (float)ne2/src0->ne[2]; - const float sf3 = (float)ne3/src0->ne[3]; - - const id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; - [encoder setBytes:&sf0 length:sizeof(sf0) atIndex:18]; - [encoder setBytes:&sf1 length:sizeof(sf1) atIndex:19]; - [encoder setBytes:&sf2 length:sizeof(sf2) atIndex:20]; - [encoder setBytes:&sf3 length:sizeof(sf3) atIndex:21]; - - const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); - - [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_PAD: - { - GGML_ASSERT(src0->type == GGML_TYPE_F32); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; - - const int nth = MIN(1024, ne0); - - [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_ARANGE: - { - GGML_ASSERT(dst->type == GGML_TYPE_F32); - - float start; - float step; - - memcpy(&start, ((int32_t *) dst->op_params) + 0, sizeof(float)); - memcpy(&step, ((int32_t *) dst->op_params) + 2, sizeof(float)); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:0]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:1]; - [encoder setBytes:&start length:sizeof(start) atIndex:2]; - [encoder setBytes:&step length:sizeof(step) atIndex:3]; - - const int nth = MIN(1024, ne0); - - [encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_TIMESTEP_EMBEDDING: - { - GGML_ASSERT(src0->type == GGML_TYPE_F32); - - const int dim = dst->op_params[0]; - const int max_period = dst->op_params[1]; - - const int half = dim / 2; - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:2]; - [encoder setBytes:&dim length:sizeof(dim) atIndex:3]; - [encoder setBytes:&max_period length:sizeof(max_period) atIndex:4]; - - const int nth = MIN(1024, half); - - [encoder dispatchThreadgroups:MTLSizeMake(ne00, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - case GGML_OP_ARGSORT: - { - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_I32); - - const int nrows = ggml_nrows(src0); - - enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0]; - - // bitonic sort requires the number of elements to be power of 2 - int64_t ne00_padded = 1; - while (ne00_padded < ne00) { - ne00_padded *= 2; - } - - // Metal kernels require the buffer size to be multiple of 16 bytes - // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength - const int mem_size = GGML_PAD(ne00_padded*sizeof(int32_t), 16); - - id pipeline = nil; - - switch (order) { - case GGML_SORT_ORDER_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break; - case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break; - default: GGML_ABORT("fatal error"); - }; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne00_padded length:sizeof( int64_t) atIndex:3]; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - - [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00_padded, 1, 1)]; - } break; - case GGML_OP_LEAKY_RELU: - { - GGML_ASSERT(src0->type == GGML_TYPE_F32); - - float slope; - memcpy(&slope, dst->op_params, sizeof(float)); - - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline; - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&slope length:sizeof(slope) atIndex:2]; - - const int64_t n = ggml_nelements(dst); - - [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; - } break; - case GGML_OP_FLASH_ATTN_EXT: - { - GGML_ASSERT(ne00 % 4 == 0); - GGML_ASSERT(ne11 % 32 == 0); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - - GGML_ASSERT(ggml_are_same_shape (src1, src2)); - - struct ggml_tensor * src3 = gf->nodes[i]->src[3]; - - size_t offs_src3 = 0; - - id id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil; - - GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16); - GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) && - "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big"); - - const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30); - //const int64_t ne31 = src3 ? src3->ne[1] : 0; - const int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32); - const int64_t ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33); - - const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30); - const uint64_t nb31 = src3 ? src3->nb[1] : 0; - const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32); - const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33); - - const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t); - - float scale; - float max_bias; - float logit_softcap; - memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale)); - memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias)); - memcpy(&logit_softcap, ((int32_t *) dst->op_params) + 2, sizeof(logit_softcap)); - - if (logit_softcap != 0.0f) { - scale /= logit_softcap; - } - - const uint32_t n_head = src0->ne[2]; - const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head)); - - const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); - const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - - id pipeline = nil; - - bool use_vec_kernel = false; - - if (ne01 >= 4 || (ne00%128 != 0)) { - switch (ne00) { - case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break; - case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break; - case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break; - case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break; - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break; - //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break; - default: - { - GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_METAL_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } - } - } else { - use_vec_kernel = true; - - switch (ne00) { - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break; - //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break; - default: - { - GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_METAL_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } - } - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; - if (id_src3) { - [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; - } else { - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:3]; - } - [encoder setBuffer:id_dst offset:offs_dst atIndex:4]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6]; - [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10]; - [encoder setBytes:&ne11 length:sizeof( int64_t) atIndex:11]; - [encoder setBytes:&ne12 length:sizeof( int64_t) atIndex:12]; - [encoder setBytes:&ne13 length:sizeof( int64_t) atIndex:13]; - [encoder setBytes:&nb11 length:sizeof(uint64_t) atIndex:14]; - [encoder setBytes:&nb12 length:sizeof(uint64_t) atIndex:15]; - [encoder setBytes:&nb13 length:sizeof(uint64_t) atIndex:16]; - [encoder setBytes:&nb21 length:sizeof(uint64_t) atIndex:17]; - [encoder setBytes:&nb22 length:sizeof(uint64_t) atIndex:18]; - [encoder setBytes:&nb23 length:sizeof(uint64_t) atIndex:19]; - [encoder setBytes:&nb31 length:sizeof(uint64_t) atIndex:20]; - [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:21]; - [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:22]; - [encoder setBytes:&scale length:sizeof( float) atIndex:23]; - [encoder setBytes:&max_bias length:sizeof( float) atIndex:24]; - [encoder setBytes:&m0 length:sizeof(m0) atIndex:25]; - [encoder setBytes:&m1 length:sizeof(m1) atIndex:26]; - [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27]; - [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:28]; - - if (!use_vec_kernel) { - // half8x8 kernel - const int64_t nqptg = 8; // queries per threadgroup !! sync with kernel template arguments !! - const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !! - - GGML_ASSERT(nqptg <= 32); - GGML_ASSERT(nqptg % 8 == 0); - GGML_ASSERT(ncpsg % 32 == 0); - - int64_t nsgmax = 2; - - while (true) { - const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2); - if (smem > ctx->device.maxThreadgroupMemoryLength) { - break; - } - nsgmax *= 2; - } - nsgmax /= 2; - - // simdgroups per threadgroup (a.k.a. warps) - const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4; - - const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2); - - //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength); - GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength); - - [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; - - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; - } else { - // half1x4 kernel - const int64_t nqptg = 1; // queries per threadgroup !! sync with kernel template arguments !! - const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !! - - GGML_ASSERT(nqptg <= 32); - GGML_ASSERT(nqptg % 1 == 0); - GGML_ASSERT(ncpsg % 32 == 0); - - // simdgroups per threadgroup (a.k.a. warps) - const int64_t nsgt = MAX(2, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)); - - int64_t nsg = 1; - while (nsg <= nsgt) { - nsg *= 2; - } - nsg /= 2; - - const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2); - - //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength); - GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength); - [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; - - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; - } - } break; - case GGML_OP_DUP: - case GGML_OP_CPY: - case GGML_OP_CONT: - { - GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0); - - int nth = MIN(1024, ne00/ggml_blck_size(src0->type)); - - id pipeline = nil; - - switch (src0t) { - case GGML_TYPE_F32: - { - GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0); - - switch (dstt) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline; break; - case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break; - case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break; - case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break; - case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break; - case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break; - case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break; - default: GGML_ABORT("not implemented"); - }; - } break; - case GGML_TYPE_F16: - { - switch (dstt) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F32].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline; break; - default: GGML_ABORT("not implemented"); - }; - } break; - default: GGML_ABORT("not implemented"); - } - - [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17]; - - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - } break; - default: - { - GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op)); - GGML_ABORT("fatal error"); + } break; + case GGML_OP_MUL_MAT_ID: + { + const int n_as = src0->ne[2]; + + // src2 = ids + const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t); + + GGML_ASSERT(src2t == GGML_TYPE_I32); + + GGML_ASSERT(!ggml_is_transposed(src0)); + GGML_ASSERT(!ggml_is_transposed(src1)); + + GGML_ASSERT(src1t == GGML_TYPE_F32); + + // find the break-even point where the matrix-matrix kernel becomes more efficient compared + // to the matrix-vector kernel + // ne20 = n_used_experts + // ne21 = n_rows + const int dst_rows = ne20*ne21; + const int dst_rows_min = n_as; + const int dst_rows_max = (device.maxThreadgroupMemoryLength - 32 - 8192)/4; + + // max size of the rowids array in the kernel shared buffer + GGML_ASSERT(dst_rows <= dst_rows_max); + + // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs + // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel + // !!! + // TODO: for now, always use mat-vec kernels until we figure out how to improve the + // indirect matrix multiplication + // !!! + if ([device supportsFamily:MTLGPUFamilyApple7] && + ne00 % 32 == 0 && ne00 >= 64 && + dst_rows > dst_rows_min) { + + // some Metal matrix data types require aligned pointers + // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5) + switch (src0->type) { + case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break; + case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break; + default: break; } - } - if (should_capture) { - [encoder popDebugGroup]; + id pipeline = nil; + + switch (src0->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32 ].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32 ].pipeline; break; + case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32 ].pipeline; break; + case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32 ].pipeline; break; + case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32 ].pipeline; break; + case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32 ].pipeline; break; + case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32 ].pipeline; break; + case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32 ].pipeline; break; + case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32 ].pipeline; break; + case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32 ].pipeline; break; + case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32 ].pipeline; break; + case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32 ].pipeline; break; + case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break; + case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break; + case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break; + case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32 ].pipeline; break; + case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32 ].pipeline; break; + case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break; + case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32 ].pipeline; break; + case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break; + case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break; + default: GGML_ABORT("MUL_MAT_ID not implemented"); + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3]; + [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4]; + [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5]; + [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:8]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:9]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:10]; + [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11]; + [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12]; + [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:17]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:18]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:19]; + + [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; + } else { + int nth0 = 32; + int nth1 = 1; + int nrows = 1; + //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); + + id pipeline = nil; + + // use custom matrix x vector kernel + switch (src0t) { + case GGML_TYPE_F32: + { + GGML_ASSERT(src1t == GGML_TYPE_F32); + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32].pipeline; + } break; + case GGML_TYPE_F16: + { + GGML_ASSERT(src1t == GGML_TYPE_F32); + nth0 = 32; + nth1 = 1; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32].pipeline; + } break; + case GGML_TYPE_Q4_0: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32].pipeline; + } break; + case GGML_TYPE_Q4_1: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32].pipeline; + } break; + case GGML_TYPE_Q5_0: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32].pipeline; + } break; + case GGML_TYPE_Q5_1: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32].pipeline; + } break; + case GGML_TYPE_Q8_0: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline; + } break; + case GGML_TYPE_Q2_K: + { + nth0 = 2; + nth1 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32].pipeline; + } break; + case GGML_TYPE_Q3_K: + { + nth0 = 2; + nth1 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32].pipeline; + } break; + case GGML_TYPE_Q4_K: + { + nth0 = 4; //1; + nth1 = 8; //32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32].pipeline; + } break; + case GGML_TYPE_Q5_K: + { + nth0 = 2; + nth1 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32].pipeline; + } break; + case GGML_TYPE_Q6_K: + { + nth0 = 2; + nth1 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32].pipeline; + } break; + case GGML_TYPE_IQ2_XXS: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32].pipeline; + } break; + case GGML_TYPE_IQ2_XS: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32].pipeline; + } break; + case GGML_TYPE_IQ3_XXS: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline; + } break; + case GGML_TYPE_IQ3_S: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline; + } break; + case GGML_TYPE_IQ2_S: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline; + } break; + case GGML_TYPE_IQ1_S: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline; + } break; + case GGML_TYPE_IQ1_M: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline; + } break; + case GGML_TYPE_IQ4_NL: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline; + } break; + case GGML_TYPE_IQ4_XS: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline; + } break; + default: + { + GGML_LOG_ERROR("Asserting on type %d\n", (int)src2t); + GGML_ABORT("not implemented"); + } + }; + + if (ggml_is_quantized(src0t)) { + GGML_ASSERT(ne00 >= nth0*nth1); + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3]; + [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4]; + [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5]; + [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12]; + [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13]; + [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14]; + [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15]; + [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:20]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:21]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:22]; + + const int64_t _ne1 = 1; + const int tgz = dst_rows; + + if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 || + src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || + src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) { + const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128; + [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) { + const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4; + [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) { + const int mem_size = 32*sizeof(float); + [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q4_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q3_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q5_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q6_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } else { + const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1 + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + } + } break; + case GGML_OP_GET_ROWS: + { + id pipeline = nil; + + switch (src0->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F32 ].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F16 ].pipeline; break; + case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0 ].pipeline; break; + case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1 ].pipeline; break; + case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0 ].pipeline; break; + case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1 ].pipeline; break; + case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0 ].pipeline; break; + case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K ].pipeline; break; + case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K ].pipeline; break; + case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K ].pipeline; break; + case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K ].pipeline; break; + case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K ].pipeline; break; + case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break; + case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break; + case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break; + case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S ].pipeline; break; + case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S ].pipeline; break; + case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break; + case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M ].pipeline; break; + case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break; + case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break; + case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break; + default: GGML_ABORT("not implemented"); + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5]; + [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6]; + [encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7]; + [encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8]; + [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:10]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)]; + } break; + case GGML_OP_RMS_NORM: + { + GGML_ASSERT(ne00 % 4 == 0); + GGML_ASSERT(ggml_is_contiguous_1(src0)); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + int nth = 32; // SIMD width + + while (nth < ne00/4 && nth < 1024) { + nth *= 2; + } + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; + [encoder setBytes:&eps length:sizeof( float) atIndex:4]; + [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; + + const int64_t nrows = ggml_nrows(src0); + + [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_GROUP_NORM: + { + GGML_ASSERT(ne00 % 4 == 0); + GGML_ASSERT(ggml_is_contiguous(src0)); + + float eps; + memcpy(&eps, dst->op_params + 1, sizeof(float)); + + const int32_t n_groups = ((const int32_t *) dst->op_params)[0]; + + int nth = 32; // SIMD width + + //while (nth < ne00/4 && nth < 1024) { + // nth *= 2; + //} + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; + [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:5]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:6]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:7]; + [encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8]; + [encoder setBytes:&eps length:sizeof( float) atIndex:9]; + [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_NORM: + { + GGML_ASSERT(ggml_is_contiguous_1(src0)); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + const int nth = MIN(256, ne00); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NORM].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; + [encoder setBytes:&eps length:sizeof( float) atIndex:4]; + [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0]; + + const int64_t nrows = ggml_nrows(src0); + + [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_ROPE: + { + GGML_ASSERT(ne10 == ne02); + + const int nth = MIN(1024, ne00); + + const int n_past = ((const int32_t *) dst->op_params)[0]; + const int n_dims = ((const int32_t *) dst->op_params)[1]; + const int mode = ((const int32_t *) dst->op_params)[2]; + // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal + const int n_ctx_orig = ((const int32_t *) dst->op_params)[4]; + + float freq_base; + float freq_scale; + float ext_factor; + float attn_factor; + float beta_fast; + float beta_slow; + + memcpy(&freq_base, (const int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (const int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (const int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (const int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (const int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (const int32_t *) dst->op_params + 10, sizeof(float)); + + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + + id pipeline = nil; + + if (!is_neox) { + switch (src0->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break; + default: GGML_ABORT("fatal error"); + }; + } else { + switch (src0->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break; + default: GGML_ABORT("fatal error"); + }; + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + if (id_src2 != nil) { + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + } else { + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:2]; + } + [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:4]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7]; + [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:10]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:11]; + [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:12]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:14]; + [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:15]; + [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:16]; + [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:17]; + [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:18]; + [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:19]; + [encoder setBytes:&n_past length:sizeof( int) atIndex:20]; + [encoder setBytes:&n_dims length:sizeof( int) atIndex:21]; + [encoder setBytes:&n_ctx_orig length:sizeof( int) atIndex:22]; + [encoder setBytes:&freq_base length:sizeof( float) atIndex:23]; + [encoder setBytes:&freq_scale length:sizeof( float) atIndex:24]; + [encoder setBytes:&ext_factor length:sizeof( float) atIndex:25]; + [encoder setBytes:&attn_factor length:sizeof( float) atIndex:26]; + [encoder setBytes:&beta_fast length:sizeof( float) atIndex:27]; + [encoder setBytes:&beta_slow length:sizeof( float) atIndex:28]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_IM2COL: + { + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32); + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int32_t N = src1->ne[is_2D ? 3 : 2]; + const int32_t IC = src1->ne[is_2D ? 2 : 1]; + const int32_t IH = is_2D ? src1->ne[1] : 1; + const int32_t IW = src1->ne[0]; + + const int32_t KH = is_2D ? src0->ne[1] : 1; + const int32_t KW = src0->ne[0]; + + const int32_t OH = is_2D ? dst->ne[2] : 1; + const int32_t OW = dst->ne[1]; + + const int32_t CHW = IC * KH * KW; + + const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; + const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; + + id pipeline = nil; + + switch (dst->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break; + default: GGML_ABORT("fatal error"); + }; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ofs0 length:sizeof( int32_t) atIndex:2]; + [encoder setBytes:&ofs1 length:sizeof( int32_t) atIndex:3]; + [encoder setBytes:&IW length:sizeof( int32_t) atIndex:4]; + [encoder setBytes:&IH length:sizeof( int32_t) atIndex:5]; + [encoder setBytes:&CHW length:sizeof( int32_t) atIndex:6]; + [encoder setBytes:&s0 length:sizeof( int32_t) atIndex:7]; + [encoder setBytes:&s1 length:sizeof( int32_t) atIndex:8]; + [encoder setBytes:&p0 length:sizeof( int32_t) atIndex:9]; + [encoder setBytes:&p1 length:sizeof( int32_t) atIndex:10]; + [encoder setBytes:&d0 length:sizeof( int32_t) atIndex:11]; + [encoder setBytes:&d1 length:sizeof( int32_t) atIndex:12]; + + [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)]; + } break; + case GGML_OP_UPSCALE: + { + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + const float sf0 = (float)ne0/src0->ne[0]; + const float sf1 = (float)ne1/src0->ne[1]; + const float sf2 = (float)ne2/src0->ne[2]; + const float sf3 = (float)ne3/src0->ne[3]; + + const id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; + [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; + [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; + [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; + [encoder setBytes:&sf0 length:sizeof(sf0) atIndex:18]; + [encoder setBytes:&sf1 length:sizeof(sf1) atIndex:19]; + [encoder setBytes:&sf2 length:sizeof(sf2) atIndex:20]; + [encoder setBytes:&sf3 length:sizeof(sf3) atIndex:21]; + + const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); + + [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_PAD: + { + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; + [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; + [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; + [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; + + const int nth = MIN(1024, ne0); + + [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_ARANGE: + { + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + float start; + float step; + + memcpy(&start, ((const int32_t *) dst->op_params) + 0, sizeof(float)); + memcpy(&step, ((const int32_t *) dst->op_params) + 2, sizeof(float)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:0]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:1]; + [encoder setBytes:&start length:sizeof(start) atIndex:2]; + [encoder setBytes:&step length:sizeof(step) atIndex:3]; + + const int nth = MIN(1024, ne0); + + [encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_TIMESTEP_EMBEDDING: + { + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + const int dim = dst->op_params[0]; + const int max_period = dst->op_params[1]; + + const int half = dim / 2; + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:2]; + [encoder setBytes:&dim length:sizeof(dim) atIndex:3]; + [encoder setBytes:&max_period length:sizeof(max_period) atIndex:4]; + + const int nth = MIN(1024, half); + + [encoder dispatchThreadgroups:MTLSizeMake(ne00, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_ARGSORT: + { + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_I32); + + const int nrows = ggml_nrows(src0); + + enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0]; + + // bitonic sort requires the number of elements to be power of 2 + int64_t ne00_padded = 1; + while (ne00_padded < ne00) { + ne00_padded *= 2; + } + + // Metal kernels require the buffer size to be multiple of 16 bytes + // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength + const int mem_size = GGML_PAD(ne00_padded*sizeof(int32_t), 16); + + id pipeline = nil; + + switch (order) { + case GGML_SORT_ORDER_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break; + case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break; + default: GGML_ABORT("fatal error"); + }; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&ne00_padded length:sizeof( int64_t) atIndex:3]; + [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00_padded, 1, 1)]; + } break; + case GGML_OP_LEAKY_RELU: + { + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + float slope; + memcpy(&slope, dst->op_params, sizeof(float)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&slope length:sizeof(slope) atIndex:2]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_FLASH_ATTN_EXT: + { + GGML_ASSERT(ne00 % 4 == 0); + GGML_ASSERT(ne11 % 32 == 0); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + GGML_ASSERT(ggml_are_same_shape (src1, src2)); + + struct ggml_tensor * src3 = node->src[3]; + + size_t offs_src3 = 0; + + id id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil; + + GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16); + GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) && + "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big"); + + const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30); + //const int64_t ne31 = src3 ? src3->ne[1] : 0; + const int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32); + const int64_t ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33); + + const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30); + const uint64_t nb31 = src3 ? src3->nb[1] : 0; + const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32); + const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33); + + const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t); + + float scale; + float max_bias; + float logit_softcap; + memcpy(&scale, ((const int32_t *) dst->op_params) + 0, sizeof(scale)); + memcpy(&max_bias, ((const int32_t *) dst->op_params) + 1, sizeof(max_bias)); + memcpy(&logit_softcap, ((const int32_t *) dst->op_params) + 2, sizeof(logit_softcap)); + + if (logit_softcap != 0.0f) { + scale /= logit_softcap; + } + + const uint32_t n_head = src0->ne[2]; + const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + id pipeline = nil; + + bool use_vec_kernel = false; + + if (ne01 >= 4 || (ne00%128 != 0)) { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break; + //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } + } else { + use_vec_kernel = true; + + switch (ne00) { + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break; + //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + if (id_src3) { + [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; + } else { + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:3]; + } + [encoder setBuffer:id_dst offset:offs_dst atIndex:4]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10]; + [encoder setBytes:&ne11 length:sizeof( int64_t) atIndex:11]; + [encoder setBytes:&ne12 length:sizeof( int64_t) atIndex:12]; + [encoder setBytes:&ne13 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&nb11 length:sizeof(uint64_t) atIndex:14]; + [encoder setBytes:&nb12 length:sizeof(uint64_t) atIndex:15]; + [encoder setBytes:&nb13 length:sizeof(uint64_t) atIndex:16]; + [encoder setBytes:&nb21 length:sizeof(uint64_t) atIndex:17]; + [encoder setBytes:&nb22 length:sizeof(uint64_t) atIndex:18]; + [encoder setBytes:&nb23 length:sizeof(uint64_t) atIndex:19]; + [encoder setBytes:&nb31 length:sizeof(uint64_t) atIndex:20]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:21]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:22]; + [encoder setBytes:&scale length:sizeof( float) atIndex:23]; + [encoder setBytes:&max_bias length:sizeof( float) atIndex:24]; + [encoder setBytes:&m0 length:sizeof(m0) atIndex:25]; + [encoder setBytes:&m1 length:sizeof(m1) atIndex:26]; + [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27]; + [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:28]; + + if (!use_vec_kernel) { + // half8x8 kernel + const int64_t nqptg = 8; // queries per threadgroup !! sync with kernel template arguments !! + const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !! + + GGML_ASSERT(nqptg <= 32); + GGML_ASSERT(nqptg % 8 == 0); + GGML_ASSERT(ncpsg % 32 == 0); + + int64_t nsgmax = 2; + + while (true) { + const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2); + if (smem > device.maxThreadgroupMemoryLength) { + break; + } + nsgmax *= 2; + } + nsgmax /= 2; + + // simdgroups per threadgroup (a.k.a. warps) + const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4; + + const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2); + + //printf("smem: %zu, max: %zu\n", smem, device.maxThreadgroupMemoryLength); + GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength); + + [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; + } else { + // half1x4 kernel + const int64_t nqptg = 1; // queries per threadgroup !! sync with kernel template arguments !! + const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !! + + GGML_ASSERT(nqptg <= 32); + GGML_ASSERT(nqptg % 1 == 0); + GGML_ASSERT(ncpsg % 32 == 0); + + // simdgroups per threadgroup (a.k.a. warps) + const int64_t nsgt = MAX(2, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)); + + int64_t nsg = 1; + while (nsg <= nsgt) { + nsg *= 2; + } + nsg /= 2; + + const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2); + + //printf("smem: %zu, max: %zu\n", smem, device.maxThreadgroupMemoryLength); + GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength); + [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; + } + } break; + case GGML_OP_DUP: + case GGML_OP_CPY: + case GGML_OP_CONT: + { + GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0); + + int nth = MIN(1024, ne00/ggml_blck_size(src0->type)); + + id pipeline = nil; + + switch (src0t) { + case GGML_TYPE_F32: + { + GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0); + + switch (dstt) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline; break; + case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break; + case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break; + case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break; + case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break; + case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break; + case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break; + default: GGML_ABORT("not implemented"); + }; + } break; + case GGML_TYPE_F16: + { + switch (dstt) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline; break; + default: GGML_ABORT("not implemented"); + }; + } break; + default: GGML_ABORT("not implemented"); + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12]; + [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14]; + [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15]; + [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16]; + [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + default: + { + GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op)); + GGML_ABORT("fatal error"); + } + } +} + +static enum ggml_status ggml_metal_graph_compute( + ggml_backend_t backend, + struct ggml_cgraph * gf) { + struct ggml_backend_metal_context * ctx = backend->context; + struct ggml_backend_metal_device_context * ctx_dev = backend->device->context; + + // number of nodes encoded by the main thread (empirically determined) + const int n_main = 128; + + // number of threads in addition to the main thread + const int n_cb = ctx->n_cb; + + // submit the ggml compute graph to the GPU by creating command buffers and encoding the ops in them + // the first n_nodes_0 are encoded and submitted for processing directly by the calling thread + // while these nodes are processing, we start n_cb threads to enqueue the rest of the nodes + // each thread creates it's own command buffer and enqueues the ops in parallel + // + // tests on M1 Pro and M2 Ultra using LLaMA models, show that optimal values for n_cb are 1 or 2 + + @autoreleasepool { + ctx->gf = gf; + + ctx->n_nodes_0 = MIN(n_main, gf->n_nodes); + ctx->n_nodes_1 = gf->n_nodes - ctx->n_nodes_0; + + ctx->n_nodes_per_cb = (ctx->n_nodes_1 + ctx->n_cb - 1) / ctx->n_cb; + + const bool should_capture = ctx->capture_next_compute; + if (should_capture) { + ctx->capture_next_compute = false; + + if (!ctx->capture_started) { + // create capture scope + ctx->capture_scope = [[MTLCaptureManager sharedCaptureManager] newCaptureScopeWithDevice:ctx_dev->mtl_device]; + + MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new]; + descriptor.captureObject = ctx->capture_scope; + descriptor.destination = MTLCaptureDestinationGPUTraceDocument; + descriptor.outputURL = [NSURL fileURLWithPath:[NSString stringWithFormat:@"/tmp/perf-metal.gputrace"]]; + + NSError * error = nil; + if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) { + GGML_LOG_ERROR("%s: error: unable to start capture '%s'\n", __func__, [[error localizedDescription] UTF8String]); + } else { + [ctx->capture_scope beginScope]; + ctx->capture_started = true; + } } } - [encoder endEncoding]; + // the main thread commits the first few commands immediately + // command_buffer[n_cb] + { + id command_buffer = [ctx->queue commandBufferWithUnretainedReferences]; + ctx->command_buffers[n_cb] = command_buffer; - if (cb_idx < 2 || ctx->abort_callback == NULL) { - [command_buffer commit]; + [command_buffer enqueue]; + ctx->encode_async(n_cb); } - }); - // Wait for completion and check status of each command buffer - // needed to detect if the device ran out-of-memory for example (#1881) + // prepare the rest of the command buffers asynchronously + // command_buffer[0.. n_cb) + for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) { + id command_buffer = [ctx->queue commandBufferWithUnretainedReferences]; + ctx->command_buffers[cb_idx] = command_buffer; - for (int i = 0; i < n_cb; ++i) { - id command_buffer = command_buffers[i]; - [command_buffer waitUntilCompleted]; + // always enqueue the first two command buffers + // enqueue all of the command buffers if we don't need to abort + if (cb_idx < 2 || ctx->abort_callback == NULL) { + [command_buffer enqueue]; + } + } - MTLCommandBufferStatus status = [command_buffer status]; - if (status != MTLCommandBufferStatusCompleted) { - GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status); - if (status == MTLCommandBufferStatusError) { - GGML_METAL_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]); + dispatch_apply(n_cb, ctx->d_queue, ctx->encode_async); + + // wait for completion and check status of each command buffer + // needed to detect if the device ran out-of-memory for example (#1881) + { + id command_buffer = ctx->command_buffers[n_cb]; + [command_buffer waitUntilCompleted]; + + MTLCommandBufferStatus status = [command_buffer status]; + if (status != MTLCommandBufferStatusCompleted) { + GGML_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, n_cb, status); + if (status == MTLCommandBufferStatusError) { + GGML_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]); + } + + return GGML_STATUS_FAILED; + } + } + + for (int i = 0; i < n_cb; ++i) { + id command_buffer = ctx->command_buffers[i]; + [command_buffer waitUntilCompleted]; + + MTLCommandBufferStatus status = [command_buffer status]; + if (status != MTLCommandBufferStatusCompleted) { + GGML_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status); + if (status == MTLCommandBufferStatusError) { + GGML_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]); + } + + return GGML_STATUS_FAILED; } - return GGML_STATUS_FAILED; + id next_buffer = (i + 1 < n_cb ? ctx->command_buffers[i + 1] : nil); + if (!next_buffer) { + continue; + } + + const bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued); + if (next_queued) { + continue; + } + + if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) { + GGML_LOG_INFO("%s: command buffer %d aborted", __func__, i); + return GGML_STATUS_ABORTED; + } + + [next_buffer commit]; } - id next_buffer = (i + 1 < n_cb ? command_buffers[i + 1] : nil); - if (!next_buffer) { - continue; + if (!should_capture && ctx->capture_started) { + [ctx->capture_scope endScope]; + [[MTLCaptureManager sharedCaptureManager] stopCapture]; } - - bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued); - if (next_queued) { - continue; - } - - if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) { - GGML_METAL_LOG_INFO("%s: command buffer %d aborted", __func__, i); - return GGML_STATUS_ABORTED; - } - - [next_buffer commit]; } - if (should_capture) { - [[MTLCaptureManager sharedCaptureManager] stopCapture]; - } - - } return GGML_STATUS_SUCCESS; } @@ -3078,44 +3146,19 @@ static enum ggml_status ggml_metal_graph_compute( // backend interface -// default buffer -static id g_backend_device = nil; -static int g_backend_device_ref_count = 0; - -static id ggml_backend_metal_get_device(void) { - if (g_backend_device == nil) { - g_backend_device = MTLCreateSystemDefaultDevice(); - } - - g_backend_device_ref_count++; - - return g_backend_device; -} - -static void ggml_backend_metal_free_device(void) { - assert(g_backend_device_ref_count > 0); - - g_backend_device_ref_count--; - - if (g_backend_device_ref_count == 0) { - [g_backend_device release]; - g_backend_device = nil; - } -} - -GGML_CALL static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) { return "Metal"; UNUSED(buffer); } -GGML_CALL static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) { struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context; for (int i = 0; i < ctx->n_buffers; i++) { [ctx->buffers[i].metal release]; } - ggml_backend_metal_free_device(); + ggml_backend_metal_device_rel(buffer->buft->device->context); if (ctx->owned) { #if TARGET_OS_OSX @@ -3128,25 +3171,25 @@ GGML_CALL static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_ free(ctx); } -GGML_CALL static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) { +static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) { struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context; return ctx->all_data; } -GGML_CALL static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { memcpy((char *)tensor->data + offset, data, size); UNUSED(buffer); } -GGML_CALL static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { +static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { memcpy(data, (const char *)tensor->data + offset, size); UNUSED(buffer); } -GGML_CALL static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { +static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { if (ggml_backend_buffer_is_host(src->buffer)) { memcpy(dst->data, src->data, ggml_nbytes(src)); return true; @@ -3156,7 +3199,7 @@ GGML_CALL static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t UNUSED(buffer); } -GGML_CALL static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { +static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context; memset(ctx->all_data, value, ctx->all_size); @@ -3177,7 +3220,7 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = { // default buffer type -GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) { return "Metal"; UNUSED(buft); @@ -3187,17 +3230,17 @@ static void ggml_backend_metal_log_allocated_size(id device, size_t s #ifndef GGML_METAL_NDEBUG #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15) if (@available(macOS 10.12, iOS 16.0, *)) { - GGML_METAL_LOG_DEBUG("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)\n", + GGML_LOG_DEBUG("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)\n", __func__, size_aligned / 1024.0 / 1024.0, device.currentAllocatedSize / 1024.0 / 1024.0, device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) { - GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__); + GGML_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__); } } else { - GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n", + GGML_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n", __func__, size_aligned / 1024.0 / 1024.0, device.currentAllocatedSize / 1024.0 / 1024.0); @@ -3208,8 +3251,8 @@ static void ggml_backend_metal_log_allocated_size(id device, size_t s UNUSED(size_aligned); } -GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context)); +static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context)); const size_t size_page = sysconf(_SC_PAGESIZE); @@ -3218,7 +3261,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff size_aligned += (size_page - (size_aligned % size_page)); } - id device = ggml_backend_metal_get_device(); + id device = ggml_backend_metal_device_acq(buft->device->context); ctx->all_data = ggml_metal_host_malloc(size_aligned); ctx->all_size = size_aligned; @@ -3232,16 +3275,16 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff if (size_aligned > 0) { ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data - length:size_aligned - options:MTLResourceStorageModeShared - deallocator:nil]; + length:size_aligned + options:MTLResourceStorageModeShared + deallocator:nil]; } } if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) { - GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0); + GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0); free(ctx); - ggml_backend_metal_free_device(); + ggml_backend_metal_device_rel(buft->device->context); return NULL; } @@ -3250,28 +3293,28 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size); } -GGML_CALL static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return 32; UNUSED(buft); } -GGML_CALL static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { - id device = ggml_backend_metal_get_device(); - size_t max_size = device.maxBufferLength; - ggml_backend_metal_free_device(); +static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { + id device = ggml_backend_metal_device_acq(buft->device->context); + const size_t max_size = device.maxBufferLength; + ggml_backend_metal_device_rel(buft->device->context); return max_size; UNUSED(buft); } -GGML_CALL static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) { +static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) { return true; UNUSED(buft); } -GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) { +ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) { static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = { /* .iface = */ { /* .get_name = */ ggml_backend_metal_buffer_type_get_name, @@ -3281,16 +3324,16 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) { /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes /* .is_host = */ ggml_backend_metal_buffer_type_is_host, }, + /* .device = */ &g_ggml_backend_metal_device, /* .context = */ NULL, }; return &ggml_backend_buffer_type_metal; } -// buffer from ptr - -GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) { - struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context)); +// TODO: obsoleted by ggml_backend_metal_device_buffer_from_ptr +ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) { + struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context)); ctx->all_data = data; ctx->all_size = size; @@ -3311,7 +3354,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_aligned += (size_page - (size_aligned % size_page)); } - id device = ggml_backend_metal_get_device(); + id device = ggml_backend_metal_device_acq(&g_ggml_ctx_dev_main); // the buffer fits into the max buffer size allowed by the device if (size_aligned <= device.maxBufferLength) { @@ -3323,7 +3366,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil]; if (ctx->buffers[ctx->n_buffers].metal == nil) { - GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0); + GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0); return false; } } @@ -3349,7 +3392,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil]; if (ctx->buffers[ctx->n_buffers].metal == nil) { - GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0); + GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0); return false; } } @@ -3357,7 +3400,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, ggml_backend_metal_log_allocated_size(device, size_step_aligned); if (i + size_step < size) { - GGML_METAL_LOG_INFO("\n"); + GGML_LOG_INFO("\n"); } ++ctx->n_buffers; @@ -3369,40 +3412,89 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, // backend -GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) { +static const char * ggml_backend_metal_name(ggml_backend_t backend) { return "Metal"; UNUSED(backend); } -GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) { - struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; +static void ggml_backend_metal_free(ggml_backend_t backend) { + struct ggml_backend_metal_context * ctx = backend->context; + struct ggml_backend_metal_device_context * ctx_dev = backend->device->context; + + ggml_backend_metal_device_rel(ctx_dev); ggml_metal_free(ctx); + free(backend); } -GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) { +static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) { return ggml_backend_metal_buffer_type(); UNUSED(backend); } -GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { - struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context; - - return ggml_metal_graph_compute(metal_ctx, cgraph); +static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { + return ggml_metal_graph_compute(backend, cgraph); } -GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { - struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context; +static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) { + GGML_ASSERT(ggml_backend_is_metal(backend)); - return ggml_metal_supports_op(metal_ctx, op); -} + struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; -GGML_CALL static bool ggml_backend_metal_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { - return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name; + if (ctx->n_cb != n_cb) { + ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_COMMAND_BUFFERS); - UNUSED(backend); + if (ctx->n_cb > 2) { + GGML_LOG_WARN("%s: n_cb = %d, using n_cb > 2 is not recommended and can degrade the performance in some cases\n", __func__, n_cb); + } + } + + if (ctx->encode_async) { + Block_release(ctx->encode_async); + } + + ctx->encode_async = Block_copy(^(size_t iter) { + const int cb_idx = iter; + const int n_cb_l = ctx->n_cb; + + const int n_nodes_0 = ctx->n_nodes_0; + const int n_nodes_1 = ctx->n_nodes_1; + + const int n_nodes_per_cb = ctx->n_nodes_per_cb; + + id command_buffer = ctx->command_buffers[cb_idx]; + id encoder = [command_buffer computeCommandEncoder]; + + int node_start = 0; + int node_end = n_nodes_0; + + if (cb_idx < n_cb_l) { + node_start = n_nodes_0 + ( (cb_idx + 0) * n_nodes_per_cb); + node_end = n_nodes_0 + (MIN((cb_idx == n_cb_l - 1) ? n_nodes_1 : (cb_idx + 1) * n_nodes_per_cb, n_nodes_1)); + } + + const bool should_capture = ctx->capture_next_compute; + + for (int idx = node_start; idx < node_end; ++idx) { + if (should_capture) { + [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(ggml_graph_node(ctx->gf, idx)) encoding:NSUTF8StringEncoding]]; + } + + ggml_metal_encode_node(backend, idx, encoder); + + if (should_capture) { + [encoder popDebugGroup]; + } + } + + [encoder endEncoding]; + + if (cb_idx < 2 || ctx->abort_callback == NULL) { + [command_buffer commit]; + } + }); } static struct ggml_backend_i ggml_backend_metal_i = { @@ -3418,56 +3510,46 @@ static struct ggml_backend_i ggml_backend_metal_i = { /* .graph_plan_update = */ NULL, /* .graph_plan_compute = */ NULL, /* .graph_compute = */ ggml_backend_metal_graph_compute, - /* .supports_op = */ ggml_backend_metal_supports_op, - /* .supports_buft = */ ggml_backend_metal_supports_buft, + /* .supports_op = */ NULL, + /* .supports_buft = */ NULL, /* .offload_op = */ NULL, - /* .event_new = */ NULL, - /* .event_free = */ NULL, /* .event_record = */ NULL, /* .event_wait = */ NULL, - /* .event_synchronize = */ NULL, }; -void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) { - ggml_metal_log_callback = log_callback; - ggml_metal_log_user_data = user_data; -} - static ggml_guid_t ggml_backend_metal_guid(void) { static ggml_guid guid = { 0x81, 0xa1, 0x8b, 0x1e, 0x71, 0xec, 0x79, 0xed, 0x2b, 0x85, 0xdc, 0x8a, 0x61, 0x98, 0x30, 0xe6 }; return &guid; } +// TODO: remove in the future ggml_backend_t ggml_backend_metal_init(void) { - struct ggml_backend_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS); + ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_metal_reg(), 0); + + struct ggml_backend_metal_context * ctx = ggml_metal_init(dev); if (ctx == NULL) { - GGML_METAL_LOG_ERROR("%s: error: failed to allocate context\n", __func__); + GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__); return NULL; } - ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend)); + ggml_backend_t backend = malloc(sizeof(struct ggml_backend)); - *metal_backend = (struct ggml_backend) { + *backend = (struct ggml_backend) { /* .guid = */ ggml_backend_metal_guid(), /* .interface = */ ggml_backend_metal_i, + /* .device = */ dev, /* .context = */ ctx, }; - return metal_backend; + ggml_backend_metal_set_n_cb(backend, 1); + + return backend; } bool ggml_backend_is_metal(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_metal_guid()); } -void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) { - GGML_ASSERT(ggml_backend_is_metal(backend)); - - struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; - - ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS); -} - void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data) { GGML_ASSERT(ggml_backend_is_metal(backend)); @@ -3480,23 +3562,258 @@ void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_ca bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) { GGML_ASSERT(ggml_backend_is_metal(backend)); - struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; + struct ggml_backend_metal_device_context * ctx_dev = backend->device->context; - return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)]; + return [ctx_dev->mtl_device supportsFamily:(MTLGPUFamilyApple1 + family - 1)]; } void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) { GGML_ASSERT(ggml_backend_is_metal(backend)); struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; - ctx->should_capture_next_compute = true; + ctx->capture_next_compute = true; } -GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning +// backend device -GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) { - return ggml_backend_metal_init(); +static const char * ggml_backend_metal_device_get_name(ggml_backend_dev_t dev) { + return "Metal"; + + GGML_UNUSED(dev); +} + +static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t dev) { + // acq/rel just to populate ctx->name in case it hasn't been done yet + struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; + ggml_backend_metal_device_acq(ctx_dev); + ggml_backend_metal_device_rel(ctx_dev); + + return ctx_dev->name; +} + +static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { + if (@available(macOS 10.12, iOS 16.0, *)) { + struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; + id device = ggml_backend_metal_device_acq(ctx_dev); + + *total = device.recommendedMaxWorkingSetSize; + *free = *total - device.currentAllocatedSize; + + ggml_backend_metal_device_rel(ctx_dev); + } else { + *free = 1; + *total = 1; + } +} + +static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backend_dev_t dev) { + return GGML_BACKEND_DEVICE_TYPE_GPU_FULL; + + GGML_UNUSED(dev); +} + +static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) { + props->name = ggml_backend_metal_device_get_name(dev); + props->description = ggml_backend_metal_device_get_description(dev); + props->type = ggml_backend_metal_device_get_type(dev); + ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total); + props->caps = (struct ggml_backend_dev_caps) { + /* .async = */ false, + /* .host_buffer = */ false, + /* .buffer_from_host_ptr = */ true, + /* .events = */ false, + }; +} + +static ggml_backend_t ggml_backend_metal_device_init(ggml_backend_dev_t dev, const char * params) { + struct ggml_backend_metal_context * ctx = ggml_metal_init(dev); + if (ctx == NULL) { + GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__); + return NULL; + } + + ggml_backend_t backend = malloc(sizeof(struct ggml_backend)); + + *backend = (struct ggml_backend) { + /* .guid = */ ggml_backend_metal_guid(), + /* .interface = */ ggml_backend_metal_i, + /* .device = */ dev, + /* .context = */ ctx, + }; + + ggml_backend_metal_set_n_cb(backend, 1); + + return backend; GGML_UNUSED(params); - GGML_UNUSED(user_data); +} + +static ggml_backend_buffer_type_t ggml_backend_metal_device_get_buffer_type(ggml_backend_dev_t dev) { + return ggml_backend_metal_buffer_type(); + + GGML_UNUSED(dev); +} + +static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) { + struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context)); + + ctx->all_data = ptr; + ctx->all_size = size; + ctx->owned = false; + ctx->n_buffers = 0; + + const size_t size_page = sysconf(_SC_PAGESIZE); + + // page-align the data ptr + { + const uintptr_t offs = (uintptr_t) ptr % size_page; + ptr = (void *) ((char *) ptr - offs); + size += offs; + } + + size_t size_aligned = size; + if ((size_aligned % size_page) != 0) { + size_aligned += (size_page - (size_aligned % size_page)); + } + + struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context; + id device = ggml_backend_metal_device_acq(ctx_dev); + + // the buffer fits into the max buffer size allowed by the device + if (size_aligned <= device.maxBufferLength) { + ctx->buffers[ctx->n_buffers].data = ptr; + ctx->buffers[ctx->n_buffers].size = size; + ctx->buffers[ctx->n_buffers].metal = nil; + + if (size_aligned > 0) { + ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:ptr length:size_aligned options:MTLResourceStorageModeShared deallocator:nil]; + + if (ctx->buffers[ctx->n_buffers].metal == nil) { + GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0); + return false; + } + } + + ggml_backend_metal_log_allocated_size(device, size_aligned); + + ++ctx->n_buffers; + } else { + // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into + // one of the views + const size_t size_ovlp = ((max_tensor_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case + const size_t size_step = device.maxBufferLength - size_ovlp; + const size_t size_view = device.maxBufferLength; + + for (size_t i = 0; i < size; i += size_step) { + const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i); + + ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) ptr + i); + ctx->buffers[ctx->n_buffers].size = size_step_aligned; + ctx->buffers[ctx->n_buffers].metal = nil; + + if (size_step_aligned > 0) { + ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) ptr + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil]; + + if (ctx->buffers[ctx->n_buffers].metal == nil) { + GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0); + return false; + } + } + + ggml_backend_metal_log_allocated_size(device, size_step_aligned); + + if (i + size_step < size) { + GGML_LOG_INFO("\n"); + } + + ++ctx->n_buffers; + } + } + + return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size); +} + +static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) { + struct ggml_backend_metal_device_context * ctx_dev = dev->context; + + return ggml_metal_supports_op(ctx_dev, op); +} + +static bool ggml_backend_metal_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) { + return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name; + + UNUSED(dev); +} + +static bool ggml_backend_metal_device_offload_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) { + return false; + + GGML_UNUSED(dev); + GGML_UNUSED(op); +} + +static struct ggml_backend_device_i ggml_backend_metal_device_i = { + /* .get_name = */ ggml_backend_metal_device_get_name, + /* .get_description = */ ggml_backend_metal_device_get_description, + /* .get_memory = */ ggml_backend_metal_device_get_memory, + /* .get_type = */ ggml_backend_metal_device_get_type, + /* .get_props = */ ggml_backend_metal_device_get_props, + /* .init_backend = */ ggml_backend_metal_device_init, + /* .get_buffer_type = */ ggml_backend_metal_device_get_buffer_type, + /* .get_host_buffer_type = */ NULL, + /* .buffer_from_host_ptr = */ ggml_backend_metal_device_buffer_from_ptr, + /* .supports_op = */ ggml_backend_metal_device_supports_op, + /* .supports_buft = */ ggml_backend_metal_device_supports_buft, + /* .offload_op = */ ggml_backend_metal_device_offload_op, + /* .event_new = */ NULL, + /* .event_free = */ NULL, + /* .event_synchronize = */ NULL, +}; + +// backend registry + +static const char * ggml_backend_metal_reg_get_name(ggml_backend_reg_t reg) { + return "Metal"; + + GGML_UNUSED(reg); +} + +static size_t ggml_backend_metal_reg_device_count(ggml_backend_reg_t reg) { + return 1; + + GGML_UNUSED(reg); +} + +static ggml_backend_dev_t ggml_backend_metal_reg_device_get(ggml_backend_reg_t reg, size_t index) { + GGML_ASSERT(index == 0); + + return &g_ggml_backend_metal_device; + + GGML_UNUSED(reg); + GGML_UNUSED(index); +} + +static struct ggml_backend_reg_i ggml_backend_metal_reg_i = { + /* .get_name = */ ggml_backend_metal_reg_get_name, + /* .device_count = */ ggml_backend_metal_reg_device_count, + /* .device_get = */ ggml_backend_metal_reg_device_get, + /* .get_proc_address = */ NULL, +}; + +ggml_backend_reg_t ggml_backend_metal_reg(void) { + // TODO: make this thread-safe somehow? + { + g_ggml_backend_metal_reg = (struct ggml_backend_reg) { + /* .iface = */ ggml_backend_metal_reg_i, + /* .context = */ NULL, + }; + + g_ggml_backend_metal_device = (struct ggml_backend_device) { + /* .iface = */ ggml_backend_metal_device_i, + /* .reg = */ &g_ggml_backend_metal_reg, + /* .context = */ &g_ggml_ctx_dev_main, + }; + } + + return &g_ggml_backend_metal_reg; } diff --git a/ggml/src/ggml-rpc.cpp b/ggml/src/ggml-rpc.cpp index 49b3fa911..ab7298cba 100644 --- a/ggml/src/ggml-rpc.cpp +++ b/ggml/src/ggml-rpc.cpp @@ -319,12 +319,12 @@ static std::shared_ptr get_socket(const std::string & endpoint) { return sock; } -GGML_CALL static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; return ctx->name.c_str(); } -GGML_CALL static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; // input serialization format: | remote_ptr (8 bytes) | std::vector input(sizeof(uint64_t), 0); @@ -337,7 +337,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t delete ctx; } -GGML_CALL static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) { +static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; if (ctx->base_cache.find(buffer) != ctx->base_cache.end()) { return ctx->base_cache[buffer]; @@ -388,7 +388,7 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) { return result; } -GGML_CALL static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { UNUSED(buffer); if (ggml_is_quantized(tensor->type)) { // TODO: this check is due to MATRIX_ROW_PADDING in CUDA and should be generalized @@ -396,7 +396,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t } } -GGML_CALL static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) | size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size; @@ -410,7 +410,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t b GGML_ASSERT(status); } -GGML_CALL static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { +static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; // input serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) | int input_size = sizeof(rpc_tensor) + 2*sizeof(uint64_t); @@ -427,7 +427,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t b memcpy(data, output.data(), size); } -GGML_CALL static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { +static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { // check if src and dst are on the same server ggml_backend_buffer_t src_buffer = src->buffer; ggml_backend_rpc_buffer_context * src_ctx = (ggml_backend_rpc_buffer_context *)src_buffer->context; @@ -452,7 +452,7 @@ GGML_CALL static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t b return output[0]; } -GGML_CALL static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { +static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; // serialization format: | bufptr (8 bytes) | value (1 byte) | int input_size = sizeof(uint64_t) + sizeof(uint8_t); @@ -477,12 +477,12 @@ static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = { /* .reset = */ NULL, }; -GGML_CALL static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t buft) { ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context; return buft_ctx->name.c_str(); } -GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context; // input serialization format: | size (8 bytes) | int input_size = sizeof(uint64_t); @@ -522,7 +522,7 @@ static size_t get_alignment(const std::shared_ptr & sock) { return alignment; } -GGML_CALL static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context; return buft_ctx->alignment; } @@ -540,12 +540,12 @@ static size_t get_max_size(const std::shared_ptr & sock) { return max_size; } -GGML_CALL static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) { ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context; return buft_ctx->max_size; } -GGML_CALL static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { +static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { UNUSED(buft); return ggml_nbytes(tensor); } @@ -559,24 +559,24 @@ static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = { /* .is_host = */ NULL, }; -GGML_CALL static const char * ggml_backend_rpc_name(ggml_backend_t backend) { +static const char * ggml_backend_rpc_name(ggml_backend_t backend) { ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; return rpc_ctx->name.c_str(); } -GGML_CALL static void ggml_backend_rpc_free(ggml_backend_t backend) { +static void ggml_backend_rpc_free(ggml_backend_t backend) { ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; delete rpc_ctx; delete backend; } -GGML_CALL static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) { +static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) { ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context; return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str()); } -GGML_CALL static void ggml_backend_rpc_synchronize(ggml_backend_t backend) { +static void ggml_backend_rpc_synchronize(ggml_backend_t backend) { UNUSED(backend); // this is no-op because we don't have any async operations } @@ -618,7 +618,7 @@ static void serialize_graph(const ggml_cgraph * cgraph, std::vector & o memcpy(out_tensors, tensors.data(), n_tensors * sizeof(rpc_tensor)); } -GGML_CALL static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { +static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; std::vector input; serialize_graph(cgraph, input); @@ -630,14 +630,14 @@ GGML_CALL static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t return (enum ggml_status)output[0]; } -GGML_CALL static bool ggml_backend_rpc_supports_op(ggml_backend_t backend, const ggml_tensor * op) { +static bool ggml_backend_rpc_supports_op(ggml_backend_t backend, const ggml_tensor * op) { UNUSED(backend); UNUSED(op); //TODO: call the remote backend and cache the results return true; } -GGML_CALL static bool ggml_backend_rpc_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { +static bool ggml_backend_rpc_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { if (!buft || buft->iface.get_name != ggml_backend_rpc_buffer_type_name) { return false; } @@ -662,14 +662,11 @@ static ggml_backend_i ggml_backend_rpc_interface = { /* .supports_op = */ ggml_backend_rpc_supports_op, /* .supports_buft = */ ggml_backend_rpc_supports_buft, /* .offload_op = */ NULL, - /* .event_new = */ NULL, - /* .event_free = */ NULL, /* .event_record = */ NULL, /* .event_wait = */ NULL, - /* .event_synchronize = */ NULL, }; -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) { +GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) { static std::mutex mutex; std::lock_guard lock(mutex); // NOTE: buffer types are allocated and never freed; this is by design @@ -694,13 +691,14 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const ggml_backend_buffer_type_t buft = new ggml_backend_buffer_type { /* .iface = */ ggml_backend_rpc_buffer_type_interface, + /* .device = */ nullptr, /* .context = */ buft_ctx }; buft_map[endpoint] = buft; return buft; } -GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) { +ggml_backend_t ggml_backend_rpc_init(const char * endpoint) { ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context { /* .endpoint = */ endpoint, /* .name = */ "RPC[" + std::string(endpoint) + "]", @@ -709,12 +707,13 @@ GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) { ggml_backend_t backend = new ggml_backend { /* .guid = */ ggml_backend_rpc_guid(), /* .interface = */ ggml_backend_rpc_interface, + /* .device = */ nullptr, /* .context = */ ctx }; return backend; } -GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend) { +GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid()); } @@ -734,7 +733,7 @@ static void get_device_memory(const std::shared_ptr & sock, size_t * f *total = total_mem; } -GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) { +GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) { auto sock = get_socket(endpoint); if (sock == nullptr) { *free = 0; diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index e306ba1c1..68f7d5b6e 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -4038,7 +4038,7 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens return true; } -GGML_API GGML_CALL void ggml_sycl_get_gpu_list(int *id_list, int max_len) try { +GGML_API void ggml_sycl_get_gpu_list(int *id_list, int max_len) try { GGML_SYCL_DEBUG("[SYCL] call ggml_sycl_get_gpu_list\n"); for(int i=0;icontext; return ctx->name.c_str(); } -GGML_CALL static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) { +static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) { return buffer->iface.get_name == ggml_backend_sycl_buffer_get_name; } @@ -4162,7 +4162,7 @@ static void * ggml_backend_sycl_buffer_get_base(ggml_backend_buffer_t buffer) { return ctx->dev_ptr; } -GGML_CALL static void +static void ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor *tensor) try { ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context; @@ -4237,7 +4237,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -GGML_CALL static bool +static bool ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor *src, ggml_tensor *dst) try { @@ -4339,12 +4339,12 @@ struct ggml_backend_sycl_buffer_type_context { queue_ptr stream = nullptr; }; -GGML_CALL static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_buffer_type_t buft) { ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; return ctx->name.c_str(); } -GGML_CALL static ggml_backend_buffer_t +static ggml_backend_buffer_t ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) try { ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; @@ -4368,7 +4368,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -GGML_CALL static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return 128; UNUSED(buft); } @@ -4379,7 +4379,7 @@ static size_t ggml_backend_sycl_buffer_type_get_max_size(ggml_backend_buffer_typ UNUSED(buft); } -GGML_CALL static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { +static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { size_t size = ggml_nbytes(tensor); int64_t ne0 = tensor->ne[0]; @@ -4424,6 +4424,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) { queue_ptr stream = &(device_i.default_queue()); ggml_backend_sycl_buffer_types[i] = { /* .iface = */ ggml_backend_sycl_buffer_type_interface, + /* .device = */ nullptr, /* .context = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i), stream}, }; } @@ -4449,6 +4450,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_conte for (int i = 0; i < ggml_sycl_info().device_count; i++) { ggml_backend_sycl_buffer_types[i] = { /* .iface = */ ggml_backend_sycl_buffer_type_interface, + /* .device = */ nullptr, /* .context = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i), ctx->stream(i, 0)}, }; } @@ -4513,7 +4515,7 @@ struct ggml_backend_sycl_split_buffer_context { std::vector streams; }; -GGML_CALL static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backend_buffer_t buffer) { return GGML_SYCL_NAME "_Split"; UNUSED(buffer); @@ -4523,19 +4525,19 @@ static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) { return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name; } -GGML_CALL static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context; delete ctx; } -GGML_CALL static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) { +static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) { // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced return (void *)0x1000; UNUSED(buffer); } -GGML_CALL static void +static void ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor *tensor) try { GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported @@ -4618,7 +4620,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -GGML_CALL static void +static void ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data, size_t offset, size_t size) try { @@ -4671,7 +4673,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -GGML_CALL static void +static void ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor *tensor, void *data, size_t offset, size_t size) try { @@ -4724,7 +4726,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -GGML_CALL static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { +static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { UNUSED(buffer); UNUSED(value); } @@ -4742,13 +4744,13 @@ static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = { /* .reset = */ NULL, }; -GGML_CALL static const char * ggml_backend_sycl_split_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_sycl_split_buffer_type_name(ggml_backend_buffer_type_t buft) { return GGML_SYCL_NAME "_Split"; UNUSED(buft); } -GGML_CALL static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point // instead, we allocate them for each tensor separately in init_tensor // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated, @@ -4758,12 +4760,12 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc return ggml_backend_buffer_init(buft, ggml_backend_sycl_split_buffer_interface, ctx, size); } -GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_sycl_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return 128; UNUSED(buft); } -GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { +static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { ggml_backend_sycl_split_buffer_type_context * ctx = (ggml_backend_sycl_split_buffer_type_context *)buft->context; size_t total_size = 0; @@ -4790,7 +4792,7 @@ GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_ return total_size; } -GGML_CALL static bool ggml_backend_sycl_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) { +static bool ggml_backend_sycl_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) { return false; UNUSED(buft); @@ -4805,7 +4807,7 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface /* .is_host = */ ggml_backend_sycl_split_buffer_type_is_host, }; -GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) { +ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) { static std::mutex mutex; std::lock_guard lock(mutex); @@ -4837,6 +4839,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const f struct ggml_backend_buffer_type buft { /* .iface = */ ggml_backend_sycl_split_buffer_type_interface, + /* .device = */ nullptr, /* .context = */ new ggml_backend_sycl_split_buffer_type_context{tensor_split_arr}, }; @@ -4846,13 +4849,13 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const f // host buffer type -GGML_CALL static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_type_t buft) { return GGML_SYCL_NAME "_Host"; UNUSED(buft); } -GGML_CALL static const char * ggml_backend_sycl_host_buffer_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_sycl_host_buffer_name(ggml_backend_buffer_t buffer) { return GGML_SYCL_NAME "_Host"; UNUSED(buffer); @@ -4890,6 +4893,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() { /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, }, + /* .device = */ nullptr, /* .context = */ nullptr, }; @@ -4898,14 +4902,14 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() { // backend -GGML_CALL static const char * ggml_backend_sycl_name(ggml_backend_t backend) { +static const char * ggml_backend_sycl_name(ggml_backend_t backend) { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; return sycl_ctx->name.c_str(); } -GGML_CALL static void ggml_backend_sycl_free(ggml_backend_t backend) { +static void ggml_backend_sycl_free(ggml_backend_t backend) { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; delete sycl_ctx; @@ -4913,12 +4917,12 @@ GGML_CALL static void ggml_backend_sycl_free(ggml_backend_t backend) { } -GGML_CALL static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) { +static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; return ggml_backend_sycl_buffer_type(sycl_ctx->device); } -GGML_CALL static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend, +static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend, ggml_tensor *tensor, const void *data, size_t offset, size_t size) try { @@ -4936,7 +4940,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -GGML_CALL static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend, +static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend, const ggml_tensor *tensor, void *data, size_t offset, size_t size) try { @@ -4954,9 +4958,9 @@ catch (sycl::exception const &exc) { std::exit(1); } -GGML_CALL static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend, - const ggml_tensor *src, - ggml_tensor *dst) try { +static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend, + const ggml_tensor *src, + ggml_tensor *dst) try { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; if (dst->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && ggml_backend_buffer_is_sycl(src->buffer)) { /* @@ -4991,7 +4995,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -GGML_CALL static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { +static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; ggml_sycl_set_main_device(sycl_ctx->device); @@ -5019,7 +5023,7 @@ GGML_CALL static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t back return GGML_STATUS_SUCCESS; } -GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) { +static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) { switch (op->op) { case GGML_OP_CONV_TRANSPOSE_1D: { @@ -5166,13 +5170,13 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons UNUSED(backend); } -GGML_CALL static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const ggml_tensor * op) { +static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const ggml_tensor * op) { const int min_batch_size = 32; return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS && op->op != GGML_OP_MUL_MAT_ID; GGML_UNUSED(backend); } -GGML_CALL static bool ggml_backend_sycl_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { +static bool ggml_backend_sycl_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { if (buft->iface.get_name != ggml_backend_sycl_buffer_type_name) { return false; } @@ -5257,11 +5261,8 @@ static ggml_backend_i ggml_backend_sycl_interface = { /* .supports_op = */ ggml_backend_sycl_supports_op, /* .supports_buft = */ ggml_backend_sycl_supports_buft, /* .offload_op = */ ggml_backend_sycl_offload_op, - /* .event_new = */ ggml_backend_sycl_event_new, - /* .event_free = */ ggml_backend_sycl_event_free, - /* .event_record = */ ggml_backend_sycl_event_record, + /* .event_record = */ NULL, /* .event_wait = */ NULL, - /* .event_synchronize = */ ggml_backend_sycl_event_synchronize, }; static ggml_guid_t ggml_backend_sycl_guid() { @@ -5269,7 +5270,7 @@ static ggml_guid_t ggml_backend_sycl_guid() { return &guid; } -GGML_CALL ggml_backend_t ggml_backend_sycl_init(int device) { +ggml_backend_t ggml_backend_sycl_init(int device) { GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_init\n"); ggml_check_sycl(); @@ -5284,6 +5285,7 @@ GGML_CALL ggml_backend_t ggml_backend_sycl_init(int device) { ggml_backend_t sycl_backend = new ggml_backend { /* .guid = */ ggml_backend_sycl_guid(), /* .interface = */ ggml_backend_sycl_interface, + /* .device = */ nullptr, /* .context = */ ctx }; @@ -5294,26 +5296,7 @@ bool ggml_backend_is_sycl(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_sycl_guid()); } -GGML_CALL int ggml_backend_sycl_get_device_count() { +int ggml_backend_sycl_get_device_count() { GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_count\n"); return ggml_sycl_info().device_count; } - -GGML_CALL static ggml_backend_t ggml_backend_reg_sycl_init(const char * params, void * user_data) { - ggml_backend_t sycl_backend = ggml_backend_sycl_init((int) (intptr_t) user_data); - return sycl_backend; - - UNUSED(params); -} - -extern "C" int ggml_backend_sycl_reg_devices(); - -int ggml_backend_sycl_reg_devices() { - assert(ggml_sycl_info().device_count>0); - for (int i = 0; i < ggml_sycl_info().device_count; i++) { - char name[128]; - snprintf(name, sizeof(name), "%s%d", GGML_SYCL_NAME, i); - ggml_backend_register(name, ggml_backend_reg_sycl_init, ggml_backend_sycl_buffer_type(i), (void *) (intptr_t) i); - } - return ggml_sycl_info().device_count; -} diff --git a/ggml/src/ggml-sycl/dequantize.hpp b/ggml/src/ggml-sycl/dequantize.hpp index 8f4041fff..b8304c3a2 100644 --- a/ggml/src/ggml-sycl/dequantize.hpp +++ b/ggml/src/ggml-sycl/dequantize.hpp @@ -55,12 +55,12 @@ static __dpct_inline__ void dequantize_q4_1(const void *vx, const int64_t ib, #ifdef GGML_SYCL_F16 // v = v * {d, d}; // v = v + {m, m}; - v.s0() = (v.s0() * d) + m; - v.s1() = (v.s1() * d) + m; + v.s0() = sycl::fma(v.s0(), d, m); + v.s1() = sycl::fma(v.s1(), d, m); #else - v.x() = (v.x() * d) + m; - v.y() = (v.y() * d) + m; + v.x() = sycl::fma(v.x(), d, m); + v.y() = sycl::fma(v.y(), d, m); #endif // GGML_SYCL_F16 } @@ -110,11 +110,11 @@ static __dpct_inline__ void dequantize_q5_1(const void *vx, const int64_t ib, #ifdef GGML_SYCL_F16 // v = v * {d, d}; // v = v + {m, m}; - v.s0() = (v.s0() * d) + m; - v.s1() = (v.s1() * d) + m; + v.s0() = sycl::fma(v.s0(), d, m); + v.s1() = sycl::fma(v.s1(), d, m); #else - v.x() = (v.x() * d) + m; - v.y() = (v.y() * d) + m; + v.x() = sycl::fma(v.x(), d, m); + v.y() = sycl::fma(v.y(), d, m); #endif // GGML_SYCL_F16 } diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp index c677a2728..30bd376da 100644 --- a/ggml/src/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan.cpp @@ -119,11 +119,11 @@ struct ggml_backend_vk_buffer_type_context { vk_device device; }; -GGML_CALL static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft); -GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size); -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft); -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft); -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor); +static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft); +static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size); +static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft); +static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft); +static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor); static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = { /* .get_name = */ ggml_backend_vk_buffer_type_name, /* .alloc_buffer = */ ggml_backend_vk_buffer_type_alloc_buffer, @@ -433,16 +433,6 @@ struct vk_context_struct { typedef std::shared_ptr vk_context; typedef std::weak_ptr vk_context_ref; -struct ggml_tensor_extra_gpu { - vk_buffer_ref buffer_gpu; - uint64_t offset; - - void reset() { - buffer_gpu.reset(); - offset = 0; - } -}; - struct ggml_vk_garbage_collector { std::vector tl_semaphores; std::vector semaphores; @@ -553,6 +543,31 @@ struct ggml_backend_vk_context { std::vector tensor_ctxs; }; +static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000; // NOLINT + +static uint64_t vk_tensor_offset(const ggml_tensor * tensor) { + if (tensor->view_src) { + return (uint8_t *) tensor->view_src->data - (uint8_t *) vk_ptr_base; + } + return (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base; +} + +struct ggml_backend_vk_buffer_context { + vk_device_ref device; + vk_buffer dev_buffer; + std::string name; + + ggml_backend_vk_buffer_context(vk_device_ref device, vk_buffer&& dev_buffer, std::string& name) : + device(device), + dev_buffer(dev_buffer), + name(name) { + } + + ~ggml_backend_vk_buffer_context() { + ggml_vk_destroy_buffer(dev_buffer); + } +}; + #ifdef GGML_VULKAN_MEMORY_DEBUG void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) { std::lock_guard guard(log_mutex); @@ -607,7 +622,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor); typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst); -GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend); +static void ggml_backend_vk_free(ggml_backend_t backend); // variables to track number of compiles in progress static uint32_t compile_count = 0; @@ -1055,10 +1070,25 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor try { buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index }); } catch (const vk::SystemError& e) { - // Out of Host/Device memory, clean up buffer - device->device.destroyBuffer(buf->buffer); - buf->size = 0; - throw e; + if (buf->memory_property_flags != fallback_flags) { + // Try again with fallback flags + memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags); + buf->memory_property_flags = fallback_flags; + + try { + buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index }); + } + catch (const vk::SystemError& e) { + device->device.destroyBuffer(buf->buffer); + buf->size = 0; + throw e; + } + } else { + // Out of Host/Device memory, clean up buffer + device->device.destroyBuffer(buf->buffer); + buf->size = 0; + throw e; + } } buf->ptr = nullptr; @@ -1164,11 +1194,11 @@ static void ggml_vk_load_shaders(vk_device& device) { // mulmat std::initializer_list warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; std::initializer_list warptile_m = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; - std::initializer_list warptile_s = { device->subgroup_size, 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size }; + std::initializer_list warptile_s = { std::max(device->subgroup_size, 16u), 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size }; std::initializer_list warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; std::initializer_list warptile_mmq_m = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; - std::initializer_list warptile_mmq_s = { device->subgroup_size, 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size }; + std::initializer_list warptile_mmq_s = { std::max(device->subgroup_size, 16u), 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size }; std::array l_wg_denoms = {128, 128, 1 }; std::array m_wg_denoms = { 64, 64, 1 }; @@ -1938,6 +1968,7 @@ static vk_device ggml_vk_get_device(size_t idx) { device->buffer_type = { /* .iface = */ ggml_backend_vk_buffer_type_interface, + /* .device = */ nullptr, /* .context = */ new ggml_backend_vk_buffer_type_context{ device->name, device }, }; @@ -3076,9 +3107,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub const uint64_t r2 = ne12 / ne02; const uint64_t r3 = ne13 / ne03; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; vk_buffer d_Qx; size_t qx_buf_offset = 0; @@ -3180,8 +3211,8 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub return; } - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); GGML_ASSERT(d_D->size >= d_buf_offset + d_sz * ne02 * ne03); vk_buffer d_X; @@ -3189,13 +3220,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub vk_buffer d_Y; uint64_t y_buf_offset = 0; if (!src0_uma) { - d_Qx = extra_src0->buffer_gpu.lock(); - qx_buf_offset = extra_src0->offset + src0->view_offs; + d_Qx = src0_buf_ctx->dev_buffer; + qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); } if (!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qy != nullptr); } if (qx_needs_dequant) { @@ -3276,9 +3307,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t r2 = ne12 / ne02; const uint64_t r3 = ne13 / ne03; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; vk_buffer d_Qx; size_t qx_buf_offset = 0; @@ -3357,21 +3388,21 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& return; } - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); vk_buffer d_X; uint64_t x_buf_offset = 0; vk_buffer d_Y; uint64_t y_buf_offset = 0; if(!src0_uma) { - d_Qx = extra_src0->buffer_gpu.lock(); - qx_buf_offset = extra_src0->offset + src0->view_offs; + d_Qx = src0_buf_ctx->dev_buffer; + qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); } if(!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qy != nullptr); } if (qx_needs_dequant) { @@ -3454,9 +3485,9 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c GGML_ASSERT(ne11 == 1); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; vk_buffer d_Qy; size_t qy_buf_offset = 0; @@ -3482,15 +3513,15 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c return; } - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); - vk_buffer d_Qx = extra_src0->buffer_gpu.lock(); - const uint64_t qx_buf_offset = extra_src0->offset + src0->view_offs; + vk_buffer d_Qx = src0_buf_ctx->dev_buffer; + const uint64_t qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); if (!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qx != nullptr); } @@ -3532,9 +3563,9 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con GGML_ASSERT(ne11 == 1); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; vk_buffer d_Qy = nullptr; size_t qy_buf_offset = 0; @@ -3561,15 +3592,15 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con return; } - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); - vk_buffer d_Qx = extra_src0->buffer_gpu.lock(); - const uint64_t qx_buf_offset = extra_src0->offset + src0->view_offs; + vk_buffer d_Qx = src0_buf_ctx->dev_buffer; + const uint64_t qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); if (!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qx != nullptr); } @@ -3631,10 +3662,10 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t n_as = ne02; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; - ggml_tensor_extra_gpu * extra_ids = (ggml_tensor_extra_gpu *) ids->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; + ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context; vk_buffer d_Qx; size_t qx_buf_offset = 0; @@ -3731,26 +3762,26 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& return; } - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); vk_buffer d_X; uint64_t x_buf_offset = 0; vk_buffer d_Y; uint64_t y_buf_offset = 0; if (!src0_uma) { - d_Qx = extra_src0->buffer_gpu.lock(); - qx_buf_offset = extra_src0->offset + src0->view_offs; + d_Qx = src0_buf_ctx->dev_buffer; + qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); } if (!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qy != nullptr); } if (!ids_uma) { - d_ids = extra_ids->buffer_gpu.lock(); - ids_buf_offset = extra_ids->offset + ids->view_offs; + d_ids = ids_buf_ctx->dev_buffer; + ids_buf_offset = vk_tensor_offset(ids) + ids->view_offs; GGML_ASSERT(d_ids != nullptr); } if (qx_needs_dequant) { @@ -3836,10 +3867,10 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte const uint64_t ne22 = dst->ne[2]; const uint64_t ne23 = dst->ne[3]; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra; - ggml_tensor_extra_gpu * extra_ids = (ggml_tensor_extra_gpu *) ids->extra; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; + ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context; vk_buffer d_Qx; size_t qx_buf_offset = 0; @@ -3924,26 +3955,26 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte return; } - vk_buffer d_D = extra->buffer_gpu.lock(); - const uint64_t d_buf_offset = extra->offset + dst->view_offs; + vk_buffer d_D = dst_buf_ctx->dev_buffer; + const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; GGML_ASSERT(d_D != nullptr); vk_buffer d_X; uint64_t x_buf_offset = 0; vk_buffer d_Y; uint64_t y_buf_offset = 0; if(!src0_uma) { - d_Qx = extra_src0->buffer_gpu.lock(); - qx_buf_offset = extra_src0->offset + src0->view_offs; + d_Qx = src0_buf_ctx->dev_buffer; + qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_Qx != nullptr); } if(!src1_uma) { - d_Qy = extra_src1->buffer_gpu.lock(); - qy_buf_offset = extra_src1->offset + src1->view_offs; + d_Qy = src1_buf_ctx->dev_buffer; + qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Qy != nullptr); } if(!ids_uma) { - d_ids = extra_ids->buffer_gpu.lock(); - ids_buf_offset = extra_ids->offset + ids->view_offs; + d_ids = ids_buf_ctx->dev_buffer; + ids_buf_offset = vk_tensor_offset(ids) + ids->view_offs; GGML_ASSERT(d_ids != nullptr); } if (qx_needs_dequant) { @@ -4250,7 +4281,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co std::cerr << "), " << ggml_op_name(op) << ", " << (dryrun ? "dryrun" : "") << ")"); GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type)))); // NOLINT GGML_ASSERT(ggml_vk_op_supports_incontiguous(op) || ggml_vk_dim01_contiguous(src0)); // NOLINT - GGML_ASSERT(dst->extra != nullptr); + GGML_ASSERT(dst->buffer != nullptr); const uint64_t ne00 = src0->ne[0]; const uint64_t ne01 = src0->ne[1]; const uint64_t ne02 = src0->ne[2]; @@ -4296,10 +4327,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co const bool op_supports_incontiguous = ggml_vk_op_supports_incontiguous(op); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * extra_src1 = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr; - ggml_tensor_extra_gpu * extra_src2 = use_src2 ? (ggml_tensor_extra_gpu *) src2->extra : nullptr; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; + ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + ggml_backend_vk_buffer_context * src1_buf_ctx = use_src1 ? (ggml_backend_vk_buffer_context *)src1->buffer->context : nullptr; + ggml_backend_vk_buffer_context * src2_buf_ctx = use_src2 ? (ggml_backend_vk_buffer_context *)src2->buffer->context : nullptr; vk_buffer d_X = nullptr; size_t x_buf_offset = 0; @@ -4330,7 +4361,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co uint64_t z_sz = use_src2 ? ggml_type_size(src2->type) * ne2 : 0; uint64_t d_sz = ggml_type_size(dst->type) * ned; - vk_buffer d_D = extra->buffer_gpu.lock(); + vk_buffer d_D = dst_buf_ctx->dev_buffer; // Workaround for tiny tensor inputs on ROPE if (op == GGML_OP_ROPE && use_src1 && y_sz > d_D->size) { @@ -4338,21 +4369,21 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co } GGML_ASSERT(d_D != nullptr); - uint64_t d_buf_offset = ((extra->offset + dst->view_offs) / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; - GGML_ASSERT(d_buf_offset == extra->offset || op == GGML_OP_CPY); // NOLINT + uint64_t d_buf_offset = ((vk_tensor_offset(dst) + dst->view_offs) / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; + GGML_ASSERT(d_buf_offset == vk_tensor_offset(dst) || op == GGML_OP_CPY); // NOLINT if(!src0_uma) { - d_X = extra_src0->buffer_gpu.lock(); - x_buf_offset = extra_src0->offset + src0->view_offs; + d_X = src0_buf_ctx->dev_buffer; + x_buf_offset = vk_tensor_offset(src0) + src0->view_offs; GGML_ASSERT(d_X != nullptr); } if (use_src1 && !src1_uma) { - d_Y = extra_src1->buffer_gpu.lock(); - y_buf_offset = extra_src1->offset + src1->view_offs; + d_Y = src1_buf_ctx->dev_buffer; + y_buf_offset = vk_tensor_offset(src1) + src1->view_offs; GGML_ASSERT(d_Y != nullptr); } if (use_src2 && !src2_uma) { - d_Z = extra_src2->buffer_gpu.lock(); - z_buf_offset = extra_src2->offset + src2->view_offs; + d_Z = src2_buf_ctx->dev_buffer; + z_buf_offset = vk_tensor_offset(src2) + src2->view_offs; GGML_ASSERT(d_Z != nullptr); } @@ -4531,11 +4562,10 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, } static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t src1_type_size = ggml_type_size(src1->type); const uint32_t dst_type_size = ggml_type_size(dst->type); - const uint32_t d_offset = ((extra->offset + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size; + const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size; int nb1 = dst->op_params[0] / 4; // 4 bytes of float32 int nb2 = dst->op_params[1] / 4; // 4 bytes of float32 @@ -4724,10 +4754,9 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co } static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); - const uint32_t d_offset = ((extra->offset + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size; + const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size; ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, { (uint32_t)ggml_nelements(src0), @@ -5535,14 +5564,6 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, } #endif -static ggml_tensor_extra_gpu * ggml_vk_tensor_create_extra(ggml_tensor * tensor) { - VK_LOG_DEBUG("ggml_vk_create_extra(" << tensor << " (" << tensor->name << ", " << ggml_op_name(tensor->op) << "))"); - ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu; - extra->reset(); - tensor->extra = extra; - return extra; -} - static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { #if defined(GGML_VULKAN_RUN_TESTS) ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32); @@ -5711,9 +5732,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_tensor* t // Returns true if node has enqueued work into the queue, false otherwise // If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution. static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool submit){ - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra; - - if (ggml_is_empty(node) || extra == nullptr) { + if (ggml_is_empty(node) || !node->buffer) { return false; } @@ -5965,7 +5984,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod } static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx, bool use_fence = true){ - ggml_tensor_extra_gpu * extra = nullptr; + ggml_backend_buffer * buf = nullptr; switch (tensor->op) { case GGML_OP_ADD: @@ -6001,7 +6020,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_LEAKY_RELU: case GGML_OP_REPEAT: - extra = (ggml_tensor_extra_gpu *) tensor->extra; + buf = tensor->buffer; break; case GGML_OP_UNARY: @@ -6011,7 +6030,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: - extra = (ggml_tensor_extra_gpu *) tensor->extra; + buf = tensor->buffer; break; default: return false; @@ -6019,14 +6038,14 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * break; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: - extra = (ggml_tensor_extra_gpu *) tensor->extra; + buf = tensor->buffer; break; default: return false; } - if (extra == nullptr) { + if (buf == nullptr) { return false; } @@ -6144,13 +6163,13 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) { ctx->device->device.destroyFence(ctx->fence); } -GGML_CALL static int ggml_vk_get_device_count() { +static int ggml_vk_get_device_count() { ggml_vk_instance_init(); return vk_instance.device_indices.size(); } -GGML_CALL static void ggml_vk_get_device_description(int device, char * description, size_t description_size) { +static void ggml_vk_get_device_description(int device, char * description, size_t description_size) { ggml_vk_instance_init(); std::vector devices = vk_instance.instance.enumeratePhysicalDevices(); @@ -6167,111 +6186,61 @@ GGML_CALL static void ggml_vk_get_device_description(int device, char * descript // device backend -static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000; // NOLINT - -struct ggml_backend_vk_buffer_context { - vk_device_ref device; - vk_buffer dev_buffer; - ggml_tensor_extra_gpu * temp_tensor_extras = nullptr; - size_t temp_tensor_extra_index = 0; - std::string name; - - ggml_backend_vk_buffer_context(vk_device_ref device, vk_buffer&& dev_buffer, std::string& name) : - device(device), - dev_buffer(dev_buffer), - name(name) { - } - - ~ggml_backend_vk_buffer_context() { - ggml_vk_destroy_buffer(dev_buffer); - if (temp_tensor_extras != nullptr) { - delete[] temp_tensor_extras; - } - } - - ggml_tensor_extra_gpu * ggml_vk_alloc_temp_tensor_extra() { - if (temp_tensor_extras == nullptr) { - temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_VK_MAX_NODES]; - } - - size_t alloc_index = temp_tensor_extra_index; - temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_VK_MAX_NODES; - ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index]; - extra->reset(); - - return extra; - } -}; - -GGML_CALL static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) { ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context; return ctx->name.c_str(); } -GGML_CALL static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) { +static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) { return buffer->iface.get_name == ggml_backend_vk_buffer_get_name; } -GGML_CALL static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) { VK_LOG_MEMORY("ggml_backend_vk_buffer_free_buffer()"); ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context; ggml_vk_destroy_buffer(ctx->dev_buffer); delete ctx; } -GGML_CALL static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) { +static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) { return vk_ptr_base; UNUSED(buffer); } -GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { VK_LOG_DEBUG("ggml_backend_vk_buffer_init_tensor(" << buffer << " (" << buffer->context << "), " << tensor << ")"); - ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context; - if (tensor->view_src != nullptr) { GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft); - GGML_ASSERT(tensor->view_src->extra != nullptr); - tensor->extra = tensor->view_src->extra; - } else { - ggml_tensor_extra_gpu * extra = ctx->ggml_vk_alloc_temp_tensor_extra(); - extra->buffer_gpu = ctx->dev_buffer; - extra->offset = (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base; - tensor->extra = extra; } } -GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { VK_LOG_DEBUG("ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context; + vk_buffer buf = buf_ctx->dev_buffer; - vk_buffer buf = extra->buffer_gpu.lock(); - - ggml_vk_buffer_write(buf, extra->offset + tensor->view_offs + offset, data, size); - - GGML_UNUSED(buffer); + ggml_vk_buffer_write(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size); } -GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { +static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { VK_LOG_DEBUG("ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context; - vk_buffer buf = extra->buffer_gpu.lock(); + vk_buffer buf = buf_ctx->dev_buffer; - ggml_vk_buffer_read(buf, extra->offset + tensor->view_offs + offset, data, size); - - GGML_UNUSED(buffer); + ggml_vk_buffer_read(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size); } -GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { +static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { if (ggml_backend_buffer_is_vk(src->buffer)) { - ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra; - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; - vk_buffer src_buf = src_extra->buffer_gpu.lock(); - vk_buffer dst_buf = dst_extra->buffer_gpu.lock(); + vk_buffer src_buf = src_buf_ctx->dev_buffer; + vk_buffer dst_buf = dst_buf_ctx->dev_buffer; - ggml_vk_buffer_copy(dst_buf, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src)); + ggml_vk_buffer_copy(dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src)); return true; } @@ -6280,7 +6249,7 @@ GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t bu UNUSED(buffer); } -GGML_CALL static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { +static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context; ggml_vk_buffer_memset(ctx->dev_buffer, 0, value, buffer->size); @@ -6300,13 +6269,13 @@ static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = { }; // vk buffer type -GGML_CALL static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft) { ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *)buft->context; return ctx->name.c_str(); } -GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { VK_LOG_MEMORY("ggml_backend_vk_buffer_type_alloc_buffer(" << size << ")"); ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context; @@ -6322,23 +6291,23 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer( return ggml_backend_buffer_init(buft, ggml_backend_vk_buffer_interface, bufctx, size); } -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context; return ctx->device->properties.limits.minStorageBufferOffsetAlignment; } -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context; return ctx->device->max_memory_allocation_size; } -GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { +static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { return ggml_nbytes(tensor); UNUSED(buft); } -GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num) { +ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num) { ggml_vk_instance_init(); VK_LOG_DEBUG("ggml_backend_vk_buffer_type(" << dev_num << ")"); @@ -6350,24 +6319,24 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num) // host buffer type -GGML_CALL static const char * ggml_backend_vk_host_buffer_type_name(ggml_backend_buffer_type_t buft) { +static const char * ggml_backend_vk_host_buffer_type_name(ggml_backend_buffer_type_t buft) { return GGML_VK_NAME "_Host"; UNUSED(buft); } -GGML_CALL static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffer) { +static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffer) { return GGML_VK_NAME "_Host"; UNUSED(buffer); } -GGML_CALL static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { +static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()"); ggml_vk_host_free(vk_instance.devices[0], buffer->context); } -GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { +static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { VK_LOG_MEMORY("ggml_backend_vk_host_buffer_type_alloc_buffer(" << size << ")"); size += 32; // Behave like the CPU buffer type @@ -6391,7 +6360,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_bu UNUSED(buft); } -GGML_CALL static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return vk_instance.devices[0]->properties.limits.minMemoryMapAlignment; UNUSED(buft); @@ -6399,7 +6368,7 @@ GGML_CALL static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_back // Should be changed to return device-specific host buffer type // but that probably requires changes in llama.cpp -GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() { +ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() { static struct ggml_backend_buffer_type ggml_backend_vk_buffer_type_host = { /* .iface = */ { /* .get_name = */ ggml_backend_vk_host_buffer_type_name, @@ -6409,6 +6378,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() { /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, }, + /* .device = */ nullptr, /* .context = */ nullptr, }; @@ -6422,13 +6392,13 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() { // backend -GGML_CALL static const char * ggml_backend_vk_name(ggml_backend_t backend) { +static const char * ggml_backend_vk_name(ggml_backend_t backend) { ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; return ctx->name.c_str(); } -GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend) { +static void ggml_backend_vk_free(ggml_backend_t backend) { ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; VK_LOG_DEBUG("ggml_backend_vk_free(" << ctx->name << ")"); @@ -6438,18 +6408,18 @@ GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend) { delete backend; } -GGML_CALL static ggml_backend_buffer_type_t ggml_backend_vk_get_default_buffer_type(ggml_backend_t backend) { +static ggml_backend_buffer_type_t ggml_backend_vk_get_default_buffer_type(ggml_backend_t backend) { ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; return &ctx->device->buffer_type; } -GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { VK_LOG_DEBUG("ggml_backend_vk_set_tensor_async(" << size << ")"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context; vk_context transfer_ctx; @@ -6462,17 +6432,17 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g transfer_ctx = ctx->transfer_ctx.lock(); } - vk_buffer buf = extra->buffer_gpu.lock(); + vk_buffer buf = buf_ctx->dev_buffer; - ggml_vk_buffer_write_async(transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size); + ggml_vk_buffer_write_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size); } -GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { +static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { VK_LOG_DEBUG("ggml_backend_vk_get_tensor_async(" << size << ")"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type"); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context; vk_context transfer_ctx; @@ -6485,17 +6455,17 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c transfer_ctx = ctx->transfer_ctx.lock(); } - vk_buffer buf = extra->buffer_gpu.lock(); + vk_buffer buf = buf_ctx->dev_buffer; - ggml_vk_buffer_read_async(transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size); + ggml_vk_buffer_read_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size); } -GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) { +static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) { VK_LOG_DEBUG("ggml_backend_vk_cpy_tensor_async()"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; if ((dst->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || dst->buffer->buft == ggml_backend_vk_host_buffer_type()) && ggml_backend_buffer_is_vk(src->buffer)) { - ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra; - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context; + ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; vk_context transfer_ctx; @@ -6508,17 +6478,17 @@ GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, c transfer_ctx = ctx->transfer_ctx.lock(); } - vk_buffer src_buf = src_extra->buffer_gpu.lock(); - vk_buffer dst_buf = dst_extra->buffer_gpu.lock(); + vk_buffer src_buf = src_buf_ctx->dev_buffer; + vk_buffer dst_buf = dst_buf_ctx->dev_buffer; - ggml_vk_buffer_copy_async(transfer_ctx, dst_buf, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src)); + ggml_vk_buffer_copy_async(transfer_ctx, dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src)); return true; } return false; } -GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) { +static void ggml_backend_vk_synchronize(ggml_backend_t backend) { VK_LOG_DEBUG("ggml_backend_vk_synchronize()"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; if(ctx->transfer_ctx.expired()) { @@ -6548,7 +6518,7 @@ static bool ggml_vk_is_empty(ggml_tensor * node) { return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE; } -GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { +static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; @@ -6611,7 +6581,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen UNUSED(backend); } -GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) { +static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) { // ggml_backend_vk_context * ctx = (ggml_backend_vk_context *) backend->context; switch (op->op) { @@ -6734,7 +6704,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const UNUSED(backend); } -GGML_CALL static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) { +static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) { const int min_batch_size = 32; return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) || @@ -6743,7 +6713,7 @@ GGML_CALL static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const g UNUSED(backend); } -GGML_CALL static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { +static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) { return false; } @@ -6771,11 +6741,8 @@ static ggml_backend_i ggml_backend_vk_interface = { /* .supports_op = */ ggml_backend_vk_supports_op, /* .supports_buft = */ ggml_backend_vk_supports_buft, /* .offload_op = */ ggml_backend_vk_offload_op, - /* .event_new = */ NULL, - /* .event_free = */ NULL, /* .event_record = */ NULL, /* .event_wait = */ NULL, - /* .event_synchronize = */ NULL, }; static ggml_guid_t ggml_backend_vk_guid() { @@ -6783,7 +6750,7 @@ static ggml_guid_t ggml_backend_vk_guid() { return &guid; } -GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num) { +ggml_backend_t ggml_backend_vk_init(size_t dev_num) { VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")"); ggml_backend_vk_context * ctx = new ggml_backend_vk_context; @@ -6792,25 +6759,26 @@ GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num) { ggml_backend_t vk_backend = new ggml_backend { /* .guid = */ ggml_backend_vk_guid(), /* .interface = */ ggml_backend_vk_interface, + /* .device = */ nullptr, /* .context = */ ctx, }; return vk_backend; } -GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend) { +bool ggml_backend_is_vk(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid()); } -GGML_CALL int ggml_backend_vk_get_device_count() { +int ggml_backend_vk_get_device_count() { return ggml_vk_get_device_count(); } -GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) { +void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) { ggml_vk_get_device_description(device, description, description_size); } -GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) { +void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) { GGML_ASSERT(device < (int) vk_instance.device_indices.size()); vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]]; @@ -6826,27 +6794,6 @@ GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size } } -// backend registry -GGML_CALL static ggml_backend_t ggml_backend_reg_vk_init(const char * params, void * user_data) { - ggml_backend_t vk_backend = ggml_backend_vk_init((int) (intptr_t) user_data); - return vk_backend; - - UNUSED(params); -} - -extern "C" GGML_CALL int ggml_backend_vk_reg_devices(); - -GGML_CALL int ggml_backend_vk_reg_devices() { - ggml_vk_instance_init(); - - for (size_t i = 0; i < vk_instance.device_indices.size(); i++) { - char name[128]; - snprintf(name, sizeof(name), "%s%ld", GGML_VK_NAME, i); - ggml_backend_register(name, ggml_backend_reg_vk_init, ggml_backend_vk_buffer_type(i), (void *) (intptr_t) i); // NOLINT - } - return vk_instance.device_indices.size(); -} - // Extension availability static bool ggml_vk_instance_validation_ext_available(const std::vector& instance_extensions) { #ifdef GGML_VULKAN_VALIDATE @@ -6949,10 +6896,10 @@ static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name) const size_t tensor_size = ggml_nbytes(tensor); tensor_data = malloc(tensor_size); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - ggml_vk_buffer_read(buffer_gpu, extra->offset + tensor->view_offs, tensor_data, tensor_size); + vk_buffer buffer_gpu = buf_ctx->dev_buffer; + ggml_vk_buffer_read(buffer_gpu, vk_tensor_offset(tensor) + tensor->view_offs, tensor_data, tensor_size); } std::cerr << "TENSOR CHECK " << name << " (" << tensor->name << "): " << ggml_op_name(tensor->op) << std::endl; @@ -7026,9 +6973,9 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { memcpy(src0_clone->data, src0->data, src0_size); memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS); } else if (ggml_backend_buffer_is_vk(src0->buffer)) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src0->extra; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - uint64_t offset = extra->offset + src0->view_offs; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + vk_buffer& buffer_gpu = buf_ctx->dev_buffer; + uint64_t offset = vk_tensor_offset(src0) + src0->view_offs; if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) { for (int i3 = 0; i3 < src0->ne[3]; i3++) { for (int i2 = 0; i2 < src0->ne[2]; i2++) { @@ -7068,9 +7015,9 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { memcpy(src1_clone->data, src1->data, src1_size); memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS); } else if (ggml_backend_buffer_is_vk(src1->buffer)) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src1->extra; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - uint64_t offset = extra->offset + src1->view_offs; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; + vk_buffer& buffer_gpu = buf_ctx->dev_buffer; + uint64_t offset = vk_tensor_offset(src1) + src1->view_offs; if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) { for (int i3 = 0; i3 < src1->ne[3]; i3++) { for (int i2 = 0; i2 < src1->ne[2]; i2++) { @@ -7110,9 +7057,9 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { memcpy(src2_clone->data, src2->data, src2_size); memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS); } else if (ggml_backend_buffer_is_vk(src2->buffer)) { - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src2->extra; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - uint64_t offset = extra->offset + src2->view_offs; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src2->buffer->context; + vk_buffer& buffer_gpu = buf_ctx->dev_buffer; + uint64_t offset = vk_tensor_offset(src2) + src2->view_offs; if (!ggml_is_contiguous(src2) && ggml_vk_dim01_contiguous(src2)) { for (int i3 = 0; i3 < src2->ne[3]; i3++) { for (int i2 = 0; i2 < src2->ne[2]; i2++) { @@ -7167,7 +7114,7 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { } else if (tensor->op == GGML_OP_PAD) { tensor_clone = ggml_pad(ggml_ctx, src0_clone, tensor->ne[0] - src0_clone->ne[0], tensor->ne[1] - src0_clone->ne[1], tensor->ne[2] - src0_clone->ne[2], tensor->ne[3] - src0_clone->ne[3]); } else if (tensor->op == GGML_OP_REPEAT) { - tensor_clone = ggml_repeat(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_repeat(ggml_ctx, src0_clone, tensor); } else if (tensor->op == GGML_OP_ADD) { tensor_clone = ggml_add(ggml_ctx, src0_clone, src1_clone); } else if (tensor->op == GGML_OP_ACC) { @@ -7312,14 +7259,15 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { size_t tensor_size = ggml_nbytes(tensor); tensor_data = malloc(tensor_size); - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context; - vk_buffer buffer_gpu = extra->buffer_gpu.lock(); - if (extra->offset + tensor->view_offs + tensor_size >= buffer_gpu->size) { - tensor_size = buffer_gpu->size - (extra->offset + tensor->view_offs); + vk_buffer& buffer_gpu = buf_ctx->dev_buffer; + uint64_t offset = vk_tensor_offset(tensor) + tensor->view_offs; + if (offset + tensor_size >= buffer_gpu->size) { + tensor_size = buffer_gpu->size - offset; } - ggml_vk_buffer_read(buffer_gpu, extra->offset + tensor->view_offs, tensor_data, tensor_size); + ggml_vk_buffer_read(buffer_gpu, offset, tensor_data, tensor_size); } float first_error_result = -1.0f; diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 81b651c6a..03b832d0f 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -319,26 +319,63 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) { // logging // +struct ggml_logger_state { + ggml_log_callback log_callback; + void * log_callback_user_data; +}; +static struct ggml_logger_state g_logger_state = {ggml_log_callback_default, NULL}; + +static void ggml_log_internal_v(enum ggml_log_level level, const char * format, va_list args) { + if (format == NULL) + return; + va_list args_copy; + va_copy(args_copy, args); + char buffer[128]; + int len = vsnprintf(buffer, 128, format, args); + if (len < 128) { + g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data); + } else { + char * buffer2 = (char *) calloc(len + 1, sizeof(char)); + vsnprintf(buffer2, len + 1, format, args_copy); + buffer2[len] = 0; + g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data); + free(buffer2); + } + va_end(args_copy); +} + +void ggml_log_internal(enum ggml_log_level level, const char * format, ...) { + va_list args; + va_start(args, format); + ggml_log_internal_v(level, format, args); + va_end(args); +} + +void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) { + (void) level; + (void) user_data; + fputs(text, stderr); + fflush(stderr); +} + #if (GGML_DEBUG >= 1) -#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__) +#define GGML_PRINT_DEBUG(...) GGML_LOG_DEBUG(__VA_ARGS__) #else #define GGML_PRINT_DEBUG(...) #endif #if (GGML_DEBUG >= 5) -#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__) +#define GGML_PRINT_DEBUG_5(...) GGML_LOG_DEBUG(__VA_ARGS__) #else #define GGML_PRINT_DEBUG_5(...) #endif #if (GGML_DEBUG >= 10) -#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__) +#define GGML_PRINT_DEBUG_10(...) GGML_LOG_DEBUG(__VA_ARGS__) #else #define GGML_PRINT_DEBUG_10(...) #endif -#define GGML_PRINT(...) printf(__VA_ARGS__) - // // end of logging block // @@ -355,7 +392,7 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) { #else inline static void * ggml_aligned_malloc(size_t size) { if (size == 0) { - GGML_PRINT("WARNING: Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n"); + GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n"); return NULL; } void * aligned_memory = NULL; @@ -377,7 +414,7 @@ inline static void * ggml_aligned_malloc(size_t size) { error_desc = "insufficient memory"; break; } - GGML_PRINT("%s: %s (attempted to allocate %6.2f MB)\n", __func__, error_desc, size/(1024.0*1024.0)); + GGML_LOG_ERROR("%s: %s (attempted to allocate %6.2f MB)\n", __func__, error_desc, size/(1024.0*1024.0)); GGML_ABORT("fatal error"); return NULL; } @@ -393,12 +430,12 @@ inline static void * ggml_aligned_malloc(size_t size) { inline static void * ggml_malloc(size_t size) { if (size == 0) { - GGML_PRINT("WARNING: Behavior may be unexpected when allocating 0 bytes for ggml_malloc!\n"); + GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_malloc!\n"); return NULL; } void * result = malloc(size); if (result == NULL) { - GGML_PRINT("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0)); + GGML_LOG_ERROR("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0)); GGML_ABORT("fatal error"); } return result; @@ -407,12 +444,12 @@ inline static void * ggml_malloc(size_t size) { // calloc inline static void * ggml_calloc(size_t num, size_t size) { if (num == 0 || size == 0) { - GGML_PRINT("WARNING: Behavior may be unexpected when allocating 0 bytes for ggml_calloc!\n"); + GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_calloc!\n"); return NULL; } void * result = calloc(num, size); if (result == NULL) { - GGML_PRINT("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0)); + GGML_LOG_ERROR("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0)); GGML_ABORT("fatal error"); } return result; @@ -461,7 +498,7 @@ struct ggml_arm_arch_features_type { } ggml_arm_arch_features = {-1, -1, -1, 0}; #endif -GGML_CALL const char * ggml_status_to_string(enum ggml_status status) { +const char * ggml_status_to_string(enum ggml_status status) { switch (status) { case GGML_STATUS_ALLOC_FAILED: return "GGML status: error (failed to allocate memory)"; case GGML_STATUS_FAILED: return "GGML status: error (operation failed)"; @@ -2957,6 +2994,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "SUM_ROWS", "MEAN", "ARGMAX", + "COUNT_EQUAL", "REPEAT", "REPEAT_BACK", "CONCAT", @@ -3030,7 +3068,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "OPT_STEP_ADAMW", }; -static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80"); +static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -3051,6 +3089,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "Ī£x_k", "Ī£x/n", "argmax(x)", + "count_equal(x)", "repeat(x)", "repeat_back(x)", "concat(x, y)", @@ -3124,7 +3163,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "adamw(x)", }; -static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80"); +static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -3347,7 +3386,7 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) { if (fptr != NULL) { char buf[42]; if (fgets(buf, sizeof(buf), fptr) && strncmp(buf, "0\n", sizeof(buf)) != 0) { - GGML_PRINT("WARNING: /proc/sys/kernel/numa_balancing is enabled, this has been observed to impair performance\n"); + GGML_LOG_WARN("/proc/sys/kernel/numa_balancing is enabled, this has been observed to impair performance\n"); } fclose(fptr); } @@ -3365,36 +3404,36 @@ bool ggml_is_numa(void) { //////////////////////////////////////////////////////////////////////////////// void ggml_print_object(const struct ggml_object * obj) { - GGML_PRINT(" - ggml_object: type = %d, offset = %zu, size = %zu, next = %p\n", + GGML_LOG_INFO(" - ggml_object: type = %d, offset = %zu, size = %zu, next = %p\n", obj->type, obj->offs, obj->size, (const void *) obj->next); } void ggml_print_objects(const struct ggml_context * ctx) { struct ggml_object * obj = ctx->objects_begin; - GGML_PRINT("%s: objects in context %p:\n", __func__, (const void *) ctx); + GGML_LOG_INFO("%s: objects in context %p:\n", __func__, (const void *) ctx); while (obj != NULL) { ggml_print_object(obj); obj = obj->next; } - GGML_PRINT("%s: --- end ---\n", __func__); + GGML_LOG_INFO("%s: --- end ---\n", __func__); } -GGML_CALL int64_t ggml_nelements(const struct ggml_tensor * tensor) { +int64_t ggml_nelements(const struct ggml_tensor * tensor) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); return tensor->ne[0]*tensor->ne[1]*tensor->ne[2]*tensor->ne[3]; } -GGML_CALL int64_t ggml_nrows(const struct ggml_tensor * tensor) { +int64_t ggml_nrows(const struct ggml_tensor * tensor) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); return tensor->ne[1]*tensor->ne[2]*tensor->ne[3]; } -GGML_CALL size_t ggml_nbytes(const struct ggml_tensor * tensor) { +size_t ggml_nbytes(const struct ggml_tensor * tensor) { size_t nbytes; size_t blck_size = ggml_blck_size(tensor->type); if (blck_size == 1) { @@ -3417,15 +3456,15 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) { return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN); } -GGML_CALL int64_t ggml_blck_size(enum ggml_type type) { +int64_t ggml_blck_size(enum ggml_type type) { return type_traits[type].blck_size; } -GGML_CALL size_t ggml_type_size(enum ggml_type type) { +size_t ggml_type_size(enum ggml_type type) { return type_traits[type].type_size; } -GGML_CALL size_t ggml_row_size(enum ggml_type type, int64_t ne) { +size_t ggml_row_size(enum ggml_type type, int64_t ne) { assert(ne % ggml_blck_size(type) == 0); return ggml_type_size(type)*ne/ggml_blck_size(type); } @@ -3434,15 +3473,15 @@ double ggml_type_sizef(enum ggml_type type) { return ((double)(type_traits[type].type_size))/type_traits[type].blck_size; } -GGML_CALL const char * ggml_type_name(enum ggml_type type) { +const char * ggml_type_name(enum ggml_type type) { return type < GGML_TYPE_COUNT ? type_traits[type].type_name : "NONE"; } -GGML_CALL bool ggml_is_quantized(enum ggml_type type) { +bool ggml_is_quantized(enum ggml_type type) { return type_traits[type].is_quantized; } -GGML_CALL const char * ggml_op_name(enum ggml_op op) { +const char * ggml_op_name(enum ggml_op op) { return GGML_OP_NAME[op]; } @@ -3454,7 +3493,7 @@ const char * ggml_unary_op_name(enum ggml_unary_op op) { return GGML_UNARY_OP_NAME[op]; } -GGML_CALL const char * ggml_op_desc(const struct ggml_tensor * t) { +const char * ggml_op_desc(const struct ggml_tensor * t) { if (t->op == GGML_OP_UNARY) { enum ggml_unary_op uop = ggml_get_unary_op(t); return ggml_unary_op_name(uop); @@ -3462,7 +3501,7 @@ GGML_CALL const char * ggml_op_desc(const struct ggml_tensor * t) { return ggml_op_name(t->op); } -GGML_CALL size_t ggml_element_size(const struct ggml_tensor * tensor) { +size_t ggml_element_size(const struct ggml_tensor * tensor) { return ggml_type_size(tensor->type); } @@ -3555,7 +3594,7 @@ size_t ggml_tensor_overhead(void) { return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE; } -GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor) { +bool ggml_is_transposed(const struct ggml_tensor * tensor) { return tensor->nb[0] > tensor->nb[1]; } @@ -3581,23 +3620,23 @@ static bool ggml_is_contiguous_n(const struct ggml_tensor * tensor, int n) { return true; } -GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor) { +bool ggml_is_contiguous(const struct ggml_tensor * tensor) { return ggml_is_contiguous_0(tensor); } -GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) { +bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) { return ggml_is_contiguous_n(tensor, 0); } -GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) { +bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) { return ggml_is_contiguous_n(tensor, 1); } -GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) { +bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) { return ggml_is_contiguous_n(tensor, 2); } -GGML_CALL bool ggml_is_permuted(const struct ggml_tensor * tensor) { +bool ggml_is_permuted(const struct ggml_tensor * tensor) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); return tensor->nb[0] > tensor->nb[1] || tensor->nb[1] > tensor->nb[2] || tensor->nb[2] > tensor->nb[3]; @@ -3612,7 +3651,7 @@ static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) { tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; } -GGML_CALL bool ggml_is_empty(const struct ggml_tensor * tensor) { +bool ggml_is_empty(const struct ggml_tensor * tensor) { for (int i = 0; i < GGML_MAX_DIMS; ++i) { if (tensor->ne[i] == 0) { // empty if any dimension has no elements @@ -3962,7 +4001,7 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end); if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) { - GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", + GGML_LOG_WARN("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size); assert(false); return NULL; @@ -4026,7 +4065,7 @@ static struct ggml_tensor * ggml_new_tensor_impl( if (ctx->scratch.data != NULL) { // allocate tensor data in the scratch buffer if (ctx->scratch.offs + data_size > ctx->scratch.size) { - GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n", + GGML_LOG_WARN("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n", __func__, ctx->scratch.offs + data_size, ctx->scratch.size); assert(false); return NULL; @@ -4195,9 +4234,13 @@ static void ggml_set_op_params_f32(struct ggml_tensor * tensor, uint32_t i, floa } struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) { + if (ggml_is_empty(tensor)) { + return tensor; + } if (tensor->buffer) { ggml_backend_tensor_memset(tensor, 0, 0, ggml_nbytes(tensor)); } else { + GGML_ASSERT(tensor->data); memset(tensor->data, 0, ggml_nbytes(tensor)); } return tensor; @@ -4628,7 +4671,7 @@ float * ggml_get_data_f32(const struct ggml_tensor * tensor) { return (float *)(tensor->data); } -GGML_CALL enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) { +enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) { GGML_ASSERT(tensor->op == GGML_OP_UNARY); return (enum ggml_unary_op) ggml_get_op_params_i32(tensor, 0); } @@ -4725,18 +4768,11 @@ struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * nam static struct ggml_tensor * ggml_dup_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - + struct ggml_tensor * a, + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_DUP; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_DUP; result->src[0] = a; return result; @@ -4744,13 +4780,13 @@ static struct ggml_tensor * ggml_dup_impl( struct ggml_tensor * ggml_dup( struct ggml_context * ctx, - struct ggml_tensor * a) { + struct ggml_tensor * a) { return ggml_dup_impl(ctx, a, false); } struct ggml_tensor * ggml_dup_inplace( struct ggml_context * ctx, - struct ggml_tensor * a) { + struct ggml_tensor * a) { return ggml_dup_impl(ctx, a, true); } @@ -4758,21 +4794,14 @@ struct ggml_tensor * ggml_dup_inplace( static struct ggml_tensor * ggml_add_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - bool inplace) { + struct ggml_tensor * a, + struct ggml_tensor * b, + bool inplace) { GGML_ASSERT(ggml_can_repeat(b, a)); - bool is_node = false; - - if (!inplace && (a->grad || b->grad)) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_ADD; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_ADD; result->src[0] = a; result->src[1] = b; @@ -4781,15 +4810,15 @@ static struct ggml_tensor * ggml_add_impl( struct ggml_tensor * ggml_add( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * a, + struct ggml_tensor * b) { return ggml_add_impl(ctx, a, b, false); } struct ggml_tensor * ggml_add_inplace( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * a, + struct ggml_tensor * b) { return ggml_add_impl(ctx, a, b, true); } @@ -4797,9 +4826,9 @@ struct ggml_tensor * ggml_add_inplace( static struct ggml_tensor * ggml_add_cast_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - enum ggml_type type) { + struct ggml_tensor * a, + struct ggml_tensor * b, + enum ggml_type type) { // TODO: support less-strict constraint // GGML_ASSERT(ggml_can_repeat(b, a)); GGML_ASSERT(ggml_can_repeat_rows(b, a)); @@ -4809,18 +4838,9 @@ static struct ggml_tensor * ggml_add_cast_impl( a->type == GGML_TYPE_F16 || a->type == GGML_TYPE_BF16); - bool is_node = false; - - if (a->grad || b->grad) { - // TODO: support backward pass for broadcasting - GGML_ASSERT(ggml_are_same_shape(a, b)); - is_node = true; - } - struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne); - result->op = GGML_OP_ADD; - result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne) : NULL; + result->op = GGML_OP_ADD; result->src[0] = a; result->src[1] = b; @@ -4829,9 +4849,9 @@ static struct ggml_tensor * ggml_add_cast_impl( struct ggml_tensor * ggml_add_cast( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - enum ggml_type type) { + struct ggml_tensor * a, + struct ggml_tensor * b, + enum ggml_type type) { return ggml_add_cast_impl(ctx, a, b, type); } @@ -4839,22 +4859,15 @@ struct ggml_tensor * ggml_add_cast( static struct ggml_tensor * ggml_add1_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - bool inplace) { + struct ggml_tensor * a, + struct ggml_tensor * b, + bool inplace) { GGML_ASSERT(ggml_is_scalar(b)); GGML_ASSERT(ggml_is_padded_1d(a)); - bool is_node = false; - - if (a->grad || b->grad) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_ADD1; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_ADD1; result->src[0] = a; result->src[1] = b; @@ -4863,15 +4876,15 @@ static struct ggml_tensor * ggml_add1_impl( struct ggml_tensor * ggml_add1( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * a, + struct ggml_tensor * b) { return ggml_add1_impl(ctx, a, b, false); } struct ggml_tensor * ggml_add1_inplace( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * a, + struct ggml_tensor * b) { return ggml_add1_impl(ctx, a, b, true); } @@ -4879,31 +4892,24 @@ struct ggml_tensor * ggml_add1_inplace( static struct ggml_tensor * ggml_acc_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t nb2, - size_t nb3, - size_t offset, - bool inplace) { + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset, + bool inplace) { GGML_ASSERT(ggml_nelements(b) <= ggml_nelements(a)); GGML_ASSERT(ggml_is_contiguous(a)); GGML_ASSERT(a->type == GGML_TYPE_F32); GGML_ASSERT(b->type == GGML_TYPE_F32); - bool is_node = false; - - if (!inplace && (a->grad || b->grad)) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_ACC; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_ACC; result->src[0] = a; result->src[1] = b; @@ -4912,23 +4918,23 @@ static struct ggml_tensor * ggml_acc_impl( struct ggml_tensor * ggml_acc( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t nb2, - size_t nb3, - size_t offset) { + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false); } struct ggml_tensor * ggml_acc_inplace( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - size_t nb1, - size_t nb2, - size_t nb3, - size_t offset) { + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, true); } @@ -4936,23 +4942,14 @@ struct ggml_tensor * ggml_acc_inplace( static struct ggml_tensor * ggml_sub_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - bool inplace) { + struct ggml_tensor * a, + struct ggml_tensor * b, + bool inplace) { GGML_ASSERT(ggml_can_repeat(b, a)); - bool is_node = false; - - if (!inplace && (a->grad || b->grad)) { - // TODO: support backward pass for broadcasting - GGML_ASSERT(ggml_are_same_shape(a, b)); - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_SUB; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SUB; result->src[0] = a; result->src[1] = b; @@ -4961,15 +4958,15 @@ static struct ggml_tensor * ggml_sub_impl( struct ggml_tensor * ggml_sub( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * a, + struct ggml_tensor * b) { return ggml_sub_impl(ctx, a, b, false); } struct ggml_tensor * ggml_sub_inplace( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * a, + struct ggml_tensor * b) { return ggml_sub_impl(ctx, a, b, true); } @@ -4977,27 +4974,14 @@ struct ggml_tensor * ggml_sub_inplace( static struct ggml_tensor * ggml_mul_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - bool inplace) { + struct ggml_tensor * a, + struct ggml_tensor * b, + bool inplace) { GGML_ASSERT(ggml_can_repeat(b, a)); - bool is_node = false; - - if (!inplace && (a->grad || b->grad)) { - // TODO: support backward pass for broadcasting - GGML_ASSERT(ggml_are_same_shape(a, b)); - is_node = true; - } - - if (inplace) { - GGML_ASSERT(!is_node); - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_MUL; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MUL; result->src[0] = a; result->src[1] = b; @@ -5022,25 +5006,14 @@ struct ggml_tensor * ggml_mul_inplace( static struct ggml_tensor * ggml_div_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - bool inplace) { + struct ggml_tensor * a, + struct ggml_tensor * b, + bool inplace) { GGML_ASSERT(ggml_can_repeat(b, a)); - bool is_node = false; - - if (!inplace && (a->grad || b->grad)) { - is_node = true; - } - - if (inplace) { - GGML_ASSERT(!is_node); - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_DIV; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_DIV; result->src[0] = a; result->src[1] = b; @@ -5065,18 +5038,11 @@ struct ggml_tensor * ggml_div_inplace( static struct ggml_tensor * ggml_sqr_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - + struct ggml_tensor * a, + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_SQR; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SQR; result->src[0] = a; return result; @@ -5098,18 +5064,11 @@ struct ggml_tensor * ggml_sqr_inplace( static struct ggml_tensor * ggml_sqrt_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - + struct ggml_tensor * a, + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_SQRT; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SQRT; result->src[0] = a; return result; @@ -5132,17 +5091,10 @@ struct ggml_tensor * ggml_sqrt_inplace( static struct ggml_tensor * ggml_log_impl( struct ggml_context * ctx, struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_LOG; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_LOG; result->src[0] = a; return result; @@ -5165,17 +5117,10 @@ struct ggml_tensor * ggml_log_inplace( static struct ggml_tensor * ggml_sin_impl( struct ggml_context * ctx, struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_SIN; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SIN; result->src[0] = a; return result; @@ -5198,17 +5143,10 @@ struct ggml_tensor * ggml_sin_inplace( static struct ggml_tensor * ggml_cos_impl( struct ggml_context * ctx, struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_COS; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_COS; result->src[0] = a; return result; @@ -5230,17 +5168,10 @@ struct ggml_tensor * ggml_cos_inplace( struct ggml_tensor * ggml_sum( struct ggml_context * ctx, - struct ggml_tensor * a) { - bool is_node = false; - - if (a->grad) { - is_node = true; - } - + struct ggml_tensor * a) { struct ggml_tensor * result = ggml_new_tensor_1d(ctx, a->type, 1); - result->op = GGML_OP_SUM; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SUM; result->src[0] = a; return result; @@ -5250,13 +5181,7 @@ struct ggml_tensor * ggml_sum( struct ggml_tensor * ggml_sum_rows( struct ggml_context * ctx, - struct ggml_tensor * a) { - bool is_node = false; - - if (a->grad) { - is_node = true; - } - + struct ggml_tensor * a) { int64_t ne[GGML_MAX_DIMS] = { 1 }; for (int i = 1; i < GGML_MAX_DIMS; ++i) { ne[i] = a->ne[i]; @@ -5264,8 +5189,7 @@ struct ggml_tensor * ggml_sum_rows( struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne); - result->op = GGML_OP_SUM_ROWS; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SUM_ROWS; result->src[0] = a; return result; @@ -5275,19 +5199,11 @@ struct ggml_tensor * ggml_sum_rows( struct ggml_tensor * ggml_mean( struct ggml_context * ctx, - struct ggml_tensor * a) { - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement - is_node = true; - } - + struct ggml_tensor * a) { int64_t ne[4] = { 1, a->ne[1], a->ne[2], a->ne[3] }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - result->op = GGML_OP_MEAN; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MEAN; result->src[0] = a; return result; @@ -5297,42 +5213,45 @@ struct ggml_tensor * ggml_mean( struct ggml_tensor * ggml_argmax( struct ggml_context * ctx, - struct ggml_tensor * a) { + struct ggml_tensor * a) { GGML_ASSERT(ggml_is_matrix(a)); - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); - is_node = true; - } struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]); - result->op = GGML_OP_ARGMAX; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_ARGMAX; result->src[0] = a; return result; } +// ggml_count_equal + +struct ggml_tensor * ggml_count_equal( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + GGML_ASSERT(ggml_are_same_shape(a, b)); + + struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, 1); + + result->op = GGML_OP_COUNT_EQUAL; + result->src[0] = a; + result->src[1] = b; + + return result; +} + // ggml_repeat struct ggml_tensor * ggml_repeat( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * a, + struct ggml_tensor * b) { GGML_ASSERT(ggml_can_repeat(a, b)); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne); - result->op = GGML_OP_REPEAT; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_REPEAT; result->src[0] = a; return result; @@ -5342,24 +5261,13 @@ struct ggml_tensor * ggml_repeat( struct ggml_tensor * ggml_repeat_back( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * a, + struct ggml_tensor * b) { GGML_ASSERT(ggml_can_repeat(b, a)); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - - if (ggml_are_same_shape(a, b) && !is_node) { - return a; - } - struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne); - result->op = GGML_OP_REPEAT_BACK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_REPEAT_BACK; result->src[0] = a; return result; @@ -5369,9 +5277,9 @@ struct ggml_tensor * ggml_repeat_back( struct ggml_tensor * ggml_concat( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int dim) { + struct ggml_tensor * a, + struct ggml_tensor * b, + int dim) { GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS); int64_t ne[GGML_MAX_DIMS]; @@ -5384,19 +5292,11 @@ struct ggml_tensor * ggml_concat( ne[d] = a->ne[d]; } - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ABORT("fatal error"); // TODO: implement - is_node = true; - } - struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne); ggml_set_op_params_i32(result, 0, dim); - result->op = GGML_OP_CONCAT; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CONCAT; result->src[0] = a; result->src[1] = b; @@ -5505,20 +5405,14 @@ struct ggml_tensor * ggml_relu_inplace( struct ggml_tensor * ggml_leaky_relu( struct ggml_context * ctx, - struct ggml_tensor * a, float negative_slope, bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - GGML_ABORT("fatal error"); // TODO: not implemented - is_node = true; - } - + struct ggml_tensor * a, + float negative_slope, + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, &negative_slope, sizeof(negative_slope)); - result->op = GGML_OP_LEAKY_RELU; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_LEAKY_RELU; result->src[0] = a; return result; @@ -5586,17 +5480,9 @@ struct ggml_tensor * ggml_silu_back( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) { - bool is_node = false; - - if (a->grad || b->grad) { - // TODO: implement backward - is_node = true; - } - struct ggml_tensor * result = ggml_dup_tensor(ctx, a); - result->op = GGML_OP_SILU_BACK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SILU_BACK; result->src[0] = a; result->src[1] = b; @@ -5604,6 +5490,7 @@ struct ggml_tensor * ggml_silu_back( } // ggml hardswish + struct ggml_tensor * ggml_hardswish( struct ggml_context * ctx, struct ggml_tensor * a) { @@ -5611,6 +5498,7 @@ struct ggml_tensor * ggml_hardswish( } // ggml hardsigmoid + struct ggml_tensor * ggml_hardsigmoid( struct ggml_context * ctx, struct ggml_tensor * a) { @@ -5618,6 +5506,7 @@ struct ggml_tensor * ggml_hardsigmoid( } // ggml exp + struct ggml_tensor * ggml_exp( struct ggml_context * ctx, struct ggml_tensor * a) { @@ -5635,21 +5524,13 @@ struct ggml_tensor * ggml_exp_inplace( static struct ggml_tensor * ggml_norm_impl( struct ggml_context * ctx, struct ggml_tensor * a, - float eps, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - + float eps, + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, &eps, sizeof(eps)); - result->op = GGML_OP_NORM; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_NORM; result->src[0] = a; return result; @@ -5658,14 +5539,14 @@ static struct ggml_tensor * ggml_norm_impl( struct ggml_tensor * ggml_norm( struct ggml_context * ctx, struct ggml_tensor * a, - float eps) { + float eps) { return ggml_norm_impl(ctx, a, eps, false); } struct ggml_tensor * ggml_norm_inplace( struct ggml_context * ctx, struct ggml_tensor * a, - float eps) { + float eps) { return ggml_norm_impl(ctx, a, eps, true); } @@ -5674,20 +5555,13 @@ struct ggml_tensor * ggml_norm_inplace( static struct ggml_tensor * ggml_rms_norm_impl( struct ggml_context * ctx, struct ggml_tensor * a, - float eps, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - + float eps, + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, &eps, sizeof(eps)); - result->op = GGML_OP_RMS_NORM; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_RMS_NORM; result->src[0] = a; return result; @@ -5696,14 +5570,14 @@ static struct ggml_tensor * ggml_rms_norm_impl( struct ggml_tensor * ggml_rms_norm( struct ggml_context * ctx, struct ggml_tensor * a, - float eps) { + float eps) { return ggml_rms_norm_impl(ctx, a, eps, false); } struct ggml_tensor * ggml_rms_norm_inplace( struct ggml_context * ctx, struct ggml_tensor * a, - float eps) { + float eps) { return ggml_rms_norm_impl(ctx, a, eps, true); } @@ -5713,20 +5587,12 @@ struct ggml_tensor * ggml_rms_norm_back( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - float eps) { - bool is_node = false; - - if (a->grad) { - // TODO: implement backward - is_node = true; - } - + float eps) { struct ggml_tensor * result = ggml_dup_tensor(ctx, a); ggml_set_op_params(result, &eps, sizeof(eps)); - result->op = GGML_OP_RMS_NORM_BACK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_RMS_NORM_BACK; result->src[0] = a; result->src[1] = b; @@ -5736,43 +5602,35 @@ struct ggml_tensor * ggml_rms_norm_back( // ggml_group_norm static struct ggml_tensor * ggml_group_norm_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_groups, - float eps, - bool inplace) { - - bool is_node = false; - if (!inplace && (a->grad)) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_groups, + float eps, + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params_i32(result, 0, n_groups); ggml_set_op_params_f32(result, 1, eps); - result->op = GGML_OP_GROUP_NORM; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_GROUP_NORM; result->src[0] = a; return result; } struct ggml_tensor * ggml_group_norm( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_groups, - float eps) { + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_groups, + float eps) { return ggml_group_norm_impl(ctx, a, n_groups, eps, false); } struct ggml_tensor * ggml_group_norm_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_groups, - float eps) { + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_groups, + float eps) { return ggml_group_norm_impl(ctx, a, n_groups, eps, true); } @@ -5785,17 +5643,10 @@ struct ggml_tensor * ggml_mul_mat( GGML_ASSERT(ggml_can_mul_mat(a, b)); GGML_ASSERT(!ggml_is_transposed(a)); - bool is_node = false; - - if (a->grad || b->grad) { - is_node = true; - } - const int64_t ne[4] = { a->ne[1], b->ne[1], b->ne[2], b->ne[3] }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - result->op = GGML_OP_MUL_MAT; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MUL_MAT; result->src[0] = a; result->src[1] = b; @@ -5841,17 +5692,10 @@ struct ggml_tensor * ggml_mul_mat_id( GGML_ASSERT(as->ne[0] == b->ne[0]); // can_mul_mat GGML_ASSERT(ids->ne[0] % b->ne[1] == 0); // can broadcast - bool is_node = false; - - if (as->grad || b->grad) { - is_node = true; - } - const int64_t ne[4] = { as->ne[1], ids->ne[0], b->ne[2], 1 }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - result->op = GGML_OP_MUL_MAT_ID; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MUL_MAT_ID; result->src[0] = as; result->src[1] = b; result->src[2] = ids; @@ -5868,18 +5712,11 @@ struct ggml_tensor * ggml_out_prod( GGML_ASSERT(ggml_can_out_prod(a, b)); GGML_ASSERT(!ggml_is_transposed(a)); - bool is_node = false; - - if (a->grad || b->grad) { - is_node = true; - } - // a is broadcastable to b for ne[2] and ne[3] -> use b->ne[2] and b->ne[3] const int64_t ne[4] = { a->ne[0], b->ne[0], b->ne[2], b->ne[3] }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - result->op = GGML_OP_OUT_PROD; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_OUT_PROD; result->src[0] = a; result->src[1] = b; @@ -5892,21 +5729,14 @@ static struct ggml_tensor * ggml_scale_impl( struct ggml_context * ctx, struct ggml_tensor * a, float s, - bool inplace) { + bool inplace) { GGML_ASSERT(ggml_is_padded_1d(a)); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, &s, sizeof(s)); - result->op = GGML_OP_SCALE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SCALE; result->src[0] = a; return result; @@ -5914,15 +5744,15 @@ static struct ggml_tensor * ggml_scale_impl( struct ggml_tensor * ggml_scale( struct ggml_context * ctx, - struct ggml_tensor * a, - float s) { + struct ggml_tensor * a, + float s) { return ggml_scale_impl(ctx, a, s, false); } struct ggml_tensor * ggml_scale_inplace( struct ggml_context * ctx, - struct ggml_tensor * a, - float s) { + struct ggml_tensor * a, + float s) { return ggml_scale_impl(ctx, a, s, true); } @@ -5936,15 +5766,9 @@ static struct ggml_tensor * ggml_set_impl( size_t nb2, size_t nb3, size_t offset, - bool inplace) { + bool inplace) { GGML_ASSERT(ggml_nelements(a) >= ggml_nelements(b)); - bool is_node = false; - - if (a->grad || b->grad) { - is_node = true; - } - // make a view of the destination struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); @@ -5952,8 +5776,7 @@ static struct ggml_tensor * ggml_set_impl( int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_SET; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SET; result->src[0] = a; result->src[1] = b; @@ -5962,8 +5785,8 @@ static struct ggml_tensor * ggml_set_impl( struct ggml_tensor * ggml_set( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, + struct ggml_tensor * a, + struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, @@ -5973,8 +5796,8 @@ struct ggml_tensor * ggml_set( struct ggml_tensor * ggml_set_inplace( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, + struct ggml_tensor * a, + struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, @@ -5984,24 +5807,24 @@ struct ggml_tensor * ggml_set_inplace( struct ggml_tensor * ggml_set_1d( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, + struct ggml_tensor * a, + struct ggml_tensor * b, size_t offset) { return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, false); } struct ggml_tensor * ggml_set_1d_inplace( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, + struct ggml_tensor * a, + struct ggml_tensor * b, size_t offset) { return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, true); } struct ggml_tensor * ggml_set_2d( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, + struct ggml_tensor * a, + struct ggml_tensor * b, size_t nb1, size_t offset) { return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false); @@ -6009,8 +5832,8 @@ struct ggml_tensor * ggml_set_2d( struct ggml_tensor * ggml_set_2d_inplace( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, + struct ggml_tensor * a, + struct ggml_tensor * b, size_t nb1, size_t offset) { return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, true); @@ -6024,13 +5847,6 @@ static struct ggml_tensor * ggml_cpy_impl( struct ggml_tensor * b) { GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b)); - bool is_node = false; - - if (a->grad || b->grad) { - // inplace is false and either one have a grad - is_node = true; - } - // make a view of the destination struct ggml_tensor * result = ggml_view_tensor(ctx, b); if (strlen(b->name) > 0) { @@ -6039,8 +5855,7 @@ static struct ggml_tensor * ggml_cpy_impl( ggml_format_name(result, "%s (copy)", a->name); } - result->op = GGML_OP_CPY; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CPY; result->src[0] = a; result->src[1] = b; @@ -6058,13 +5873,10 @@ struct ggml_tensor * ggml_cast( struct ggml_context * ctx, struct ggml_tensor * a, enum ggml_type type) { - bool is_node = false; - struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne); ggml_format_name(result, "%s (copy)", a->name); - result->op = GGML_OP_CPY; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CPY; result->src[0] = a; result->src[1] = result; @@ -6076,17 +5888,10 @@ struct ggml_tensor * ggml_cast( static struct ggml_tensor * ggml_cont_impl( struct ggml_context * ctx, struct ggml_tensor * a) { - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = ggml_dup_tensor(ctx, a); ggml_format_name(result, "%s (cont)", a->name); - result->op = GGML_OP_CONT; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CONT; result->src[0] = a; return result; @@ -6132,13 +5937,10 @@ struct ggml_tensor * ggml_cont_4d( int64_t ne3) { GGML_ASSERT(ggml_nelements(a) == (ne0*ne1*ne2*ne3)); - bool is_node = false; - struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3); ggml_format_name(result, "%s (cont)", a->name); - result->op = GGML_OP_CONT; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CONT; result->src[0] = a; return result; @@ -6154,22 +5956,10 @@ struct ggml_tensor * ggml_reshape( // as only the shape of b is relevant, and not its memory layout, b is allowed to be non contiguous. GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b)); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - - if (b->grad) { - // gradient propagation is not supported - //GGML_ABORT("fatal error"); - } - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b->ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); - result->op = GGML_OP_RESHAPE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_RESHAPE; result->src[0] = a; return result; @@ -6182,18 +5972,11 @@ struct ggml_tensor * ggml_reshape_1d( GGML_ASSERT(ggml_is_contiguous(a)); GGML_ASSERT(ggml_nelements(a) == ne0); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - const int64_t ne[1] = { ne0 }; struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); - result->op = GGML_OP_RESHAPE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_RESHAPE; result->src[0] = a; return result; @@ -6207,18 +5990,11 @@ struct ggml_tensor * ggml_reshape_2d( GGML_ASSERT(ggml_is_contiguous(a)); GGML_ASSERT(ggml_nelements(a) == ne0*ne1); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - const int64_t ne[2] = { ne0, ne1 }; struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); - result->op = GGML_OP_RESHAPE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_RESHAPE; result->src[0] = a; return result; @@ -6233,18 +6009,11 @@ struct ggml_tensor * ggml_reshape_3d( GGML_ASSERT(ggml_is_contiguous(a)); GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - const int64_t ne[3] = { ne0, ne1, ne2 }; struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); - result->op = GGML_OP_RESHAPE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_RESHAPE; result->src[0] = a; return result; @@ -6260,18 +6029,11 @@ struct ggml_tensor * ggml_reshape_4d( GGML_ASSERT(ggml_is_contiguous(a)); GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - const int64_t ne[4] = { ne0, ne1, ne2, ne3 }; struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); - result->op = GGML_OP_RESHAPE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_RESHAPE; result->src[0] = a; return result; @@ -6283,20 +6045,12 @@ static struct ggml_tensor * ggml_view_impl( int n_dims, const int64_t * ne, size_t offset) { - - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, a, offset); ggml_format_name(result, "%s (view)", a->name); ggml_set_op_params(result, &offset, sizeof(offset)); - result->op = GGML_OP_VIEW; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_VIEW; result->src[0] = a; return result; @@ -6309,7 +6063,6 @@ struct ggml_tensor * ggml_view_1d( struct ggml_tensor * a, int64_t ne0, size_t offset) { - struct ggml_tensor * result = ggml_view_impl(ctx, a, 1, &ne0, offset); return result; @@ -6324,7 +6077,6 @@ struct ggml_tensor * ggml_view_2d( int64_t ne1, size_t nb1, size_t offset) { - const int64_t ne[2] = { ne0, ne1 }; struct ggml_tensor * result = ggml_view_impl(ctx, a, 2, ne, offset); @@ -6347,7 +6099,6 @@ struct ggml_tensor * ggml_view_3d( size_t nb1, size_t nb2, size_t offset) { - const int64_t ne[3] = { ne0, ne1, ne2 }; struct ggml_tensor * result = ggml_view_impl(ctx, a, 3, ne, offset); @@ -6372,7 +6123,6 @@ struct ggml_tensor * ggml_view_4d( size_t nb2, size_t nb3, size_t offset) { - const int64_t ne[4] = { ne0, ne1, ne2, ne3 }; struct ggml_tensor * result = ggml_view_impl(ctx, a, 4, ne, offset); @@ -6405,12 +6155,6 @@ struct ggml_tensor * ggml_permute( GGML_ASSERT(axis1 != axis3); GGML_ASSERT(axis2 != axis3); - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = ggml_view_tensor(ctx, a); ggml_format_name(result, "%s (permuted)", a->name); @@ -6437,8 +6181,7 @@ struct ggml_tensor * ggml_permute( result->nb[2] = nb[2]; result->nb[3] = nb[3]; - result->op = GGML_OP_PERMUTE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_PERMUTE; result->src[0] = a; int32_t params[] = { axis0, axis1, axis2, axis3 }; @@ -6452,12 +6195,6 @@ struct ggml_tensor * ggml_permute( struct ggml_tensor * ggml_transpose( struct ggml_context * ctx, struct ggml_tensor * a) { - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = ggml_view_tensor(ctx, a); ggml_format_name(result, "%s (transposed)", a->name); @@ -6467,8 +6204,7 @@ struct ggml_tensor * ggml_transpose( result->nb[0] = a->nb[1]; result->nb[1] = a->nb[0]; - result->op = GGML_OP_TRANSPOSE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_TRANSPOSE; result->src[0] = a; return result; @@ -6484,12 +6220,6 @@ struct ggml_tensor * ggml_get_rows( GGML_ASSERT(b->ne[3] == 1); GGML_ASSERT(b->type == GGML_TYPE_I32); - bool is_node = false; - - if (a->grad || b->grad) { - is_node = true; - } - // TODO: implement non F32 return enum ggml_type type = GGML_TYPE_F32; if (a->type == GGML_TYPE_I32) { @@ -6497,8 +6227,7 @@ struct ggml_tensor * ggml_get_rows( } struct ggml_tensor * result = ggml_new_tensor_4d(ctx, type, a->ne[0], b->ne[0], b->ne[1], b->ne[2]); - result->op = GGML_OP_GET_ROWS; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_GET_ROWS; result->src[0] = a; result->src[1] = b; @@ -6515,18 +6244,11 @@ struct ggml_tensor * ggml_get_rows_back( GGML_ASSERT(ggml_is_matrix(a) && ggml_is_vector(b) && b->type == GGML_TYPE_I32); GGML_ASSERT(ggml_is_matrix(c) && (a->ne[0] == c->ne[0])); - bool is_node = false; - - if (a->grad || b->grad) { - is_node = true; - } - // TODO: implement non F32 return //struct ggml_tensor * result = ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]); struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, c->ne[0], c->ne[1]); - result->op = GGML_OP_GET_ROWS_BACK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_GET_ROWS_BACK; result->src[0] = a; result->src[1] = b; @@ -6539,17 +6261,11 @@ struct ggml_tensor * ggml_diag( struct ggml_context * ctx, struct ggml_tensor * a) { GGML_ASSERT(a->ne[1] == 1); - bool is_node = false; - - if (a->grad) { - is_node = true; - } const int64_t ne[4] = { a->ne[0], a->ne[0], a->ne[2], a->ne[3] }; struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, 4, ne); - result->op = GGML_OP_DIAG; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_DIAG; result->src[0] = a; return result; @@ -6562,19 +6278,12 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl( struct ggml_tensor * a, int n_past, bool inplace) { - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); int32_t params[] = { n_past }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_DIAG_MASK_INF; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_DIAG_MASK_INF; result->src[0] = a; return result; @@ -6601,19 +6310,12 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl( struct ggml_tensor * a, int n_past, bool inplace) { - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); int32_t params[] = { n_past }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_DIAG_MASK_ZERO; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_DIAG_MASK_ZERO; result->src[0] = a; return result; @@ -6656,19 +6358,12 @@ static struct ggml_tensor * ggml_soft_max_impl( GGML_ASSERT(mask); } - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); float params[] = { scale, max_bias }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_SOFT_MAX; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SOFT_MAX; result->src[0] = a; result->src[1] = mask; @@ -6703,16 +6398,9 @@ static struct ggml_tensor * ggml_soft_max_back_impl( struct ggml_tensor * a, struct ggml_tensor * b, bool inplace) { - bool is_node = false; - - if (a->grad || b->grad) { - is_node = true; // TODO : implement backward pass - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - result->op = GGML_OP_SOFT_MAX_BACK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SOFT_MAX_BACK; result->src[0] = a; result->src[1] = b; @@ -6761,12 +6449,6 @@ static struct ggml_tensor * ggml_rope_impl( GGML_ASSERT(c->ne[0] >= n_dims / 2); } - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig }; @@ -6778,8 +6460,7 @@ static struct ggml_tensor * ggml_rope_impl( memcpy(params + 10, &beta_slow, sizeof(float)); ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_ROPE; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_ROPE; result->src[0] = a; result->src[1] = b; result->src[2] = c; @@ -6907,13 +6588,6 @@ struct ggml_tensor * ggml_rope_back( GGML_ASSERT(b->type == GGML_TYPE_I32); GGML_ASSERT(a->ne[2] == b->ne[0]); - bool is_node = false; - - if (a->grad) { - GGML_ASSERT(false && "backwards pass not implemented"); - is_node = false; - } - struct ggml_tensor * result = ggml_dup_tensor(ctx, a); int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig }; @@ -6925,8 +6599,7 @@ struct ggml_tensor * ggml_rope_back( memcpy(params + 10, &beta_slow, sizeof(float)); ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_ROPE_BACK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_ROPE_BACK; result->src[0] = a; result->src[1] = b; result->src[2] = c; @@ -6941,21 +6614,13 @@ struct ggml_tensor * ggml_clamp( struct ggml_tensor * a, float min, float max) { - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - // TODO: when implement backward, fix this: struct ggml_tensor * result = ggml_view_tensor(ctx, a); float params[] = { min, max }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_CLAMP; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CLAMP; result->src[0] = a; return result; @@ -7017,13 +6682,6 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( GGML_ASSERT(p0 == 0); GGML_ASSERT(d0 == 1); - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - const int64_t ne[4] = { ggml_calc_conv_transpose_1d_output_size(b->ne[0], a->ne[0], s0, 0 /*p0*/, 1 /*d0*/), a->ne[1], b->ne[2], 1, @@ -7033,8 +6691,7 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( int32_t params[] = { s0, p0, d0 }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_CONV_TRANSPOSE_1D; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CONV_TRANSPOSE_1D; result->src[0] = a; result->src[1] = b; @@ -7042,17 +6699,17 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( } // ggml_conv_depthwise -struct ggml_tensor * ggml_conv_depthwise_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1) { +struct ggml_tensor * ggml_conv_depthwise_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1) { struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]); struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]), @@ -7072,29 +6729,23 @@ struct ggml_tensor * ggml_conv_depthwise_2d( // b: [N, IC, IH, IW] // result: [N, OH, OW, IC*KH*KW] struct ggml_tensor * ggml_im2col( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1, - bool is_2D, - enum ggml_type dst_type) { - + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1, + bool is_2D, + enum ggml_type dst_type) { if(is_2D) { GGML_ASSERT(a->ne[2] == b->ne[2]); } else { GGML_ASSERT(a->ne[1] == b->ne[1]); GGML_ASSERT(b->ne[3] == 1); } - bool is_node = false; - - if (/*a->grad ||*/ b->grad) { // a is only used for its shape, not its data - is_node = true; - } const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0; const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); @@ -7113,8 +6764,7 @@ struct ggml_tensor * ggml_im2col( int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_IM2COL; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_IM2COL; result->src[0] = a; result->src[1] = b; @@ -7122,30 +6772,22 @@ struct ggml_tensor * ggml_im2col( } struct ggml_tensor * ggml_im2col_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int64_t * ne, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1, - bool is_2D) { - - bool is_node = false; - - if (/*a->grad ||*/ b->grad) { // a is only used for its shape, not its data - is_node = true; - } - + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int64_t * ne, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1, + bool is_2D) { struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_IM2COL_BACK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_IM2COL_BACK; result->src[0] = a; result->src[1] = b; @@ -7159,12 +6801,12 @@ struct ggml_tensor * ggml_conv_2d( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1) { + int s0, + int s1, + int p0, + int p1, + int d0, + int d1) { struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, a->type); // [N, OH, OW, IC * KH * KW] struct ggml_tensor * result = @@ -7180,6 +6822,7 @@ struct ggml_tensor * ggml_conv_2d( } // ggml_conv_2d_sk_p0 + struct ggml_tensor * ggml_conv_2d_sk_p0( struct ggml_context * ctx, struct ggml_tensor * a, @@ -7209,13 +6852,6 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0( int stride) { GGML_ASSERT(a->ne[3] == b->ne[2]); - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - const int64_t ne[4] = { ggml_calc_conv_transpose_output_size(b->ne[0], a->ne[0], stride, 0 /*p0*/), ggml_calc_conv_transpose_output_size(b->ne[1], a->ne[1], stride, 0 /*p1*/), @@ -7226,8 +6862,7 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0( ggml_set_op_params_i32(result, 0, stride); - result->op = GGML_OP_CONV_TRANSPOSE_2D; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CONV_TRANSPOSE_2D; result->src[0] = a; result->src[1] = b; @@ -7249,14 +6884,6 @@ struct ggml_tensor * ggml_pool_1d( int k0, int s0, int p0) { - - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - const int64_t ne[4] = { ggml_calc_pool_output_size(a->ne[0], k0, s0, p0), a->ne[1], @@ -7268,8 +6895,7 @@ struct ggml_tensor * ggml_pool_1d( int32_t params[] = { op, k0, s0, p0 }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_POOL_1D; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_POOL_1D; result->src[0] = a; return result; @@ -7287,13 +6913,6 @@ struct ggml_tensor * ggml_pool_2d( int s1, float p0, float p1) { - - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result; const int64_t ne[4] = { ggml_calc_pool_output_size(a->ne[0], k0, s0, p0), @@ -7306,9 +6925,9 @@ struct ggml_tensor * ggml_pool_2d( int32_t params[] = { op, k0, k1, s0, s1, p0, p1 }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_POOL_2D; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_POOL_2D; result->src[0] = a; + return result; } @@ -7323,100 +6942,74 @@ struct ggml_tensor * ggml_pool_2d_back( int s1, float p0, float p1) { - - bool is_node = false; - - if (a->grad) { - is_node = true; - } - struct ggml_tensor * result; result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, af->ne); int32_t params[] = { op, k0, k1, s0, s1, p0, p1 }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_POOL_2D_BACK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_POOL_2D_BACK; result->src[0] = a; result->src[1] = af; + return result; } // ggml_upscale static struct ggml_tensor * ggml_upscale_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - int ne0, - int ne1, - int ne2, - int ne3) { - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - + struct ggml_context * ctx, + struct ggml_tensor * a, + int ne0, + int ne1, + int ne2, + int ne3) { GGML_ASSERT(a->ne[0] <= ne0); GGML_ASSERT(a->ne[1] <= ne1); GGML_ASSERT(a->ne[2] <= ne2); GGML_ASSERT(a->ne[3] <= ne3); - struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, - ne0, - ne1, - ne2, - ne3 - ); + struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3); - result->op = GGML_OP_UPSCALE; - - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_UPSCALE; result->src[0] = a; return result; } struct ggml_tensor * ggml_upscale( - struct ggml_context * ctx, - struct ggml_tensor * a, - int scale_factor) { + struct ggml_context * ctx, + struct ggml_tensor * a, + int scale_factor) { return ggml_upscale_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3]); } struct ggml_tensor * ggml_upscale_ext( - struct ggml_context * ctx, - struct ggml_tensor * a, - int ne0, - int ne1, - int ne2, - int ne3) { + struct ggml_context * ctx, + struct ggml_tensor * a, + int ne0, + int ne1, + int ne2, + int ne3) { return ggml_upscale_impl(ctx, a, ne0, ne1, ne2, ne3); } // ggml_pad struct ggml_tensor * ggml_pad( - struct ggml_context * ctx, - struct ggml_tensor * a, - int p0, int p1, int p2, int p3) { - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - + struct ggml_context * ctx, + struct ggml_tensor * a, + int p0, + int p1, + int p2, + int p3) { struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, a->ne[0] + p0, a->ne[1] + p1, a->ne[2] + p2, a->ne[3] + p3); - result->op = GGML_OP_PAD; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_PAD; result->src[0] = a; return result; @@ -7425,39 +7018,32 @@ struct ggml_tensor * ggml_pad( // ggml_arange struct ggml_tensor * ggml_arange( - struct ggml_context * ctx, - float start, - float stop, - float step) { - + struct ggml_context * ctx, + float start, + float stop, + float step) { GGML_ASSERT(stop > start); const int64_t steps = (int64_t) ceilf((stop - start) / step); struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps); - result->op = GGML_OP_ARANGE; ggml_set_op_params_f32(result, 0, start); ggml_set_op_params_f32(result, 1, stop); ggml_set_op_params_f32(result, 2, step); + result->op = GGML_OP_ARANGE; + return result; } // ggml_timestep_embedding struct ggml_tensor * ggml_timestep_embedding( - struct ggml_context * ctx, - struct ggml_tensor * timesteps, - int dim, - int max_period) { - bool is_node = false; - - if (timesteps->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - + struct ggml_context * ctx, + struct ggml_tensor * timesteps, + int dim, + int max_period) { int actual_dim = dim; if (dim % 2 != 0) { actual_dim = dim + 1; @@ -7465,11 +7051,10 @@ struct ggml_tensor * ggml_timestep_embedding( struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, actual_dim, timesteps->ne[0]); - result->op = GGML_OP_TIMESTEP_EMBEDDING; ggml_set_op_params_i32(result, 0, dim); ggml_set_op_params_i32(result, 1, max_period); - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_TIMESTEP_EMBEDDING; result->src[0] = timesteps; return result; @@ -7478,22 +7063,14 @@ struct ggml_tensor * ggml_timestep_embedding( // ggml_argsort struct ggml_tensor * ggml_argsort( - struct ggml_context * ctx, - struct ggml_tensor * a, - enum ggml_sort_order order) { - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: not implemented - is_node = true; - } - + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_sort_order order) { struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne); ggml_set_op_params_i32(result, 0, (int32_t) order); - result->op = GGML_OP_ARGSORT; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_ARGSORT; result->src[0] = a; return result; @@ -7546,10 +7123,6 @@ struct ggml_tensor * ggml_flash_attn_ext( bool is_node = false; - if (q->grad || k->grad || v->grad) { - is_node = true; - } - // permute(0, 2, 1, 3) int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); @@ -7676,17 +7249,9 @@ struct ggml_tensor * ggml_ssm_conv( GGML_ASSERT(sx->ne[1] == d_inner); GGML_ASSERT(n_t >= 0); - bool is_node = false; - - if (sx->grad || c->grad) { - GGML_ABORT("fatal error"); // TODO: implement - is_node = true; - } - struct ggml_tensor * result = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_t, n_s); - result->op = GGML_OP_SSM_CONV; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_SSM_CONV; result->src[0] = sx; result->src[1] = c; @@ -7730,18 +7295,10 @@ struct ggml_tensor * ggml_ssm_scan( GGML_ASSERT(B->ne[2] == n_seqs); } - bool is_node = false; - - if (s->grad || x->grad || dt->grad || A->grad || B->grad || C->grad) { - GGML_ABORT("fatal error"); // TODO: implement - is_node = true; - } - // concatenated y + ssm_states struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s)); result->op = GGML_OP_SSM_SCAN; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = s; result->src[1] = x; result->src[2] = dt; @@ -7761,13 +7318,6 @@ struct ggml_tensor * ggml_win_part( GGML_ASSERT(a->ne[3] == 1); GGML_ASSERT(a->type == GGML_TYPE_F32); - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - // padding const int px = (w - a->ne[1]%w)%w; const int py = (w - a->ne[2]%w)%w; @@ -7782,8 +7332,7 @@ struct ggml_tensor * ggml_win_part( int32_t params[] = { npx, npy, w }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_WIN_PART; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_WIN_PART; result->src[0] = a; return result; @@ -7799,21 +7348,13 @@ struct ggml_tensor * ggml_win_unpart( int w) { GGML_ASSERT(a->type == GGML_TYPE_F32); - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - const int64_t ne[4] = { a->ne[0], w0, h0, 1, }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne); int32_t params[] = { w }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_WIN_UNPART; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_WIN_UNPART; result->src[0] = a; return result; @@ -7829,18 +7370,10 @@ struct ggml_tensor * ggml_get_rel_pos( GGML_ASSERT(qh == kh); GGML_ASSERT(2*MAX(qh, kh) - 1 == a->ne[1]); - bool is_node = false; - - if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - const int64_t ne[4] = { a->ne[0], kh, qh, 1, }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 3, ne); - result->op = GGML_OP_GET_REL_POS; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_GET_REL_POS; result->src[0] = a; return result; @@ -7864,17 +7397,10 @@ static struct ggml_tensor * ggml_add_rel_pos_impl( GGML_ASSERT(pw->ne[0]*pw->ne[0] == a->ne[0]); GGML_ASSERT(pw->ne[1]*pw->ne[2] == a->ne[1]); - bool is_node = false; - - if (!inplace && (a->grad || pw->grad || ph->grad)) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params_i32(result, 0, inplace ? 1 : 0); - result->op = GGML_OP_ADD_REL_POS; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_ADD_REL_POS; result->src[0] = a; result->src[1] = pw; result->src[2] = ph; @@ -7902,12 +7428,12 @@ struct ggml_tensor * ggml_add_rel_pos_inplace( struct ggml_tensor * ggml_rwkv_wkv( struct ggml_context * ctx, - struct ggml_tensor * k, - struct ggml_tensor * v, - struct ggml_tensor * r, - struct ggml_tensor * tf, - struct ggml_tensor * td, - struct ggml_tensor * state) { + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * r, + struct ggml_tensor * tf, + struct ggml_tensor * td, + struct ggml_tensor * state) { GGML_ASSERT(ggml_is_contiguous(k)); GGML_ASSERT(ggml_is_contiguous(v)); GGML_ASSERT(ggml_is_contiguous(r)); @@ -7928,19 +7454,11 @@ struct ggml_tensor * ggml_rwkv_wkv( GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs); } - bool is_node = false; - - if (k->grad || v->grad || r->grad || tf->grad || td->grad || state->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward - is_node = true; - } - // concat output and new_state const int64_t ne[4] = { S * H, n_tokens + S * n_seqs, 1, 1 }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - result->op = GGML_OP_RWKV_WKV; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_RWKV_WKV; result->src[0] = k; result->src[1] = v; result->src[2] = r; @@ -7955,23 +7473,16 @@ struct ggml_tensor * ggml_rwkv_wkv( static struct ggml_tensor * ggml_unary_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - enum ggml_unary_op op, - bool inplace) { + struct ggml_tensor * a, + enum ggml_unary_op op, + bool inplace) { GGML_ASSERT(ggml_is_contiguous_1(a)); - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params_i32(result, 0, (int32_t) op); - result->op = GGML_OP_UNARY; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_UNARY; result->src[0] = a; return result; @@ -7980,14 +7491,14 @@ static struct ggml_tensor * ggml_unary_impl( struct ggml_tensor * ggml_unary( struct ggml_context * ctx, struct ggml_tensor * a, - enum ggml_unary_op op) { + enum ggml_unary_op op) { return ggml_unary_impl(ctx, a, op, false); } struct ggml_tensor * ggml_unary_inplace( struct ggml_context * ctx, struct ggml_tensor * a, - enum ggml_unary_op op) { + enum ggml_unary_op op) { return ggml_unary_impl(ctx, a, op, true); } @@ -7996,20 +7507,13 @@ struct ggml_tensor * ggml_unary_inplace( static struct ggml_tensor * ggml_map_unary_impl_f32( struct ggml_context * ctx, struct ggml_tensor * a, - const ggml_unary_op_f32_t fun, - bool inplace) { - bool is_node = false; - - if (!inplace && a->grad) { - is_node = true; - } - + const ggml_unary_op_f32_t fun, + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - result->op = GGML_OP_MAP_UNARY; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MAP_UNARY; result->src[0] = a; return result; @@ -8018,14 +7522,14 @@ static struct ggml_tensor * ggml_map_unary_impl_f32( struct ggml_tensor * ggml_map_unary_f32( struct ggml_context * ctx, struct ggml_tensor * a, - const ggml_unary_op_f32_t fun) { + const ggml_unary_op_f32_t fun) { return ggml_map_unary_impl_f32(ctx, a, fun, false); } struct ggml_tensor * ggml_map_unary_inplace_f32( struct ggml_context * ctx, struct ggml_tensor * a, - const ggml_unary_op_f32_t fun) { + const ggml_unary_op_f32_t fun) { return ggml_map_unary_impl_f32(ctx, a, fun, true); } @@ -8035,22 +7539,15 @@ static struct ggml_tensor * ggml_map_binary_impl_f32( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - const ggml_binary_op_f32_t fun, - bool inplace) { + const ggml_binary_op_f32_t fun, + bool inplace) { GGML_ASSERT(ggml_are_same_shape(a, b)); - bool is_node = false; - - if (!inplace && (a->grad || b->grad)) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - result->op = GGML_OP_MAP_BINARY; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MAP_BINARY; result->src[0] = a; result->src[1] = b; @@ -8061,7 +7558,7 @@ struct ggml_tensor * ggml_map_binary_f32( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - const ggml_binary_op_f32_t fun) { + const ggml_binary_op_f32_t fun) { return ggml_map_binary_impl_f32(ctx, a, b, fun, false); } @@ -8069,7 +7566,7 @@ struct ggml_tensor * ggml_map_binary_inplace_f32( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - const ggml_binary_op_f32_t fun) { + const ggml_binary_op_f32_t fun) { return ggml_map_binary_impl_f32(ctx, a, b, fun, true); } @@ -8079,19 +7576,12 @@ static struct ggml_tensor * ggml_map_custom1_impl_f32( struct ggml_context * ctx, struct ggml_tensor * a, const ggml_custom1_op_f32_t fun, - bool inplace) { - bool is_node = false; - - if (!inplace && a->grad) { - is_node = true; - } - + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - result->op = GGML_OP_MAP_CUSTOM1_F32; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MAP_CUSTOM1_F32; result->src[0] = a; return result; @@ -8118,19 +7608,12 @@ static struct ggml_tensor * ggml_map_custom2_impl_f32( struct ggml_tensor * a, struct ggml_tensor * b, const ggml_custom2_op_f32_t fun, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad || b->grad)) { - is_node = true; - } - + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - result->op = GGML_OP_MAP_CUSTOM2_F32; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MAP_CUSTOM2_F32; result->src[0] = a; result->src[1] = b; @@ -8161,19 +7644,12 @@ static struct ggml_tensor * ggml_map_custom3_impl_f32( struct ggml_tensor * b, struct ggml_tensor * c, const ggml_custom3_op_f32_t fun, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad || b->grad || c->grad)) { - is_node = true; - } - + bool inplace) { struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - result->op = GGML_OP_MAP_CUSTOM3_F32; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MAP_CUSTOM3_F32; result->src[0] = a; result->src[1] = b; result->src[2] = c; @@ -8201,26 +7677,20 @@ struct ggml_tensor * ggml_map_custom3_inplace_f32( // ggml_map_custom1 struct ggml_map_custom1_op_params { - ggml_custom1_op_t fun; - int n_tasks; - void * userdata; + ggml_custom1_op_t fun; + int n_tasks; + void * userdata; }; static struct ggml_tensor * ggml_map_custom1_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_custom1_op_t fun, - int n_tasks, - void * userdata, - bool inplace) { + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_custom1_op_t fun, + int n_tasks, + void * userdata, + bool inplace) { GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0); - bool is_node = false; - - if (!inplace && a->grad) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); struct ggml_map_custom1_op_params params = { @@ -8230,55 +7700,48 @@ static struct ggml_tensor * ggml_map_custom1_impl( }; ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); - result->op = GGML_OP_MAP_CUSTOM1; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MAP_CUSTOM1; result->src[0] = a; return result; } struct ggml_tensor * ggml_map_custom1( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_custom1_op_t fun, - int n_tasks, - void * userdata) { + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_custom1_op_t fun, + int n_tasks, + void * userdata) { return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, false); } struct ggml_tensor * ggml_map_custom1_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_custom1_op_t fun, - int n_tasks, - void * userdata) { + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_custom1_op_t fun, + int n_tasks, + void * userdata) { return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, true); } // ggml_map_custom2 struct ggml_map_custom2_op_params { - ggml_custom2_op_t fun; - int n_tasks; - void * userdata; + ggml_custom2_op_t fun; + int n_tasks; + void * userdata; }; static struct ggml_tensor * ggml_map_custom2_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_custom2_op_t fun, - int n_tasks, - void * userdata, - bool inplace) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_custom2_op_t fun, + int n_tasks, + void * userdata, + bool inplace) { GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0); - bool is_node = false; - - if (!inplace && (a->grad || b->grad)) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); struct ggml_map_custom2_op_params params = { @@ -8288,8 +7751,7 @@ static struct ggml_tensor * ggml_map_custom2_impl( }; ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); - result->op = GGML_OP_MAP_CUSTOM2; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MAP_CUSTOM2; result->src[0] = a; result->src[1] = b; @@ -8297,22 +7759,22 @@ static struct ggml_tensor * ggml_map_custom2_impl( } struct ggml_tensor * ggml_map_custom2( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_custom2_op_t fun, - int n_tasks, - void * userdata) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_custom2_op_t fun, + int n_tasks, + void * userdata) { return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, false); } struct ggml_tensor * ggml_map_custom2_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_custom2_op_t fun, - int n_tasks, - void * userdata) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_custom2_op_t fun, + int n_tasks, + void * userdata) { return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, true); } @@ -8325,22 +7787,16 @@ struct ggml_map_custom3_op_params { }; static struct ggml_tensor * ggml_map_custom3_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - const ggml_custom3_op_t fun, - int n_tasks, - void * userdata, - bool inplace) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + const ggml_custom3_op_t fun, + int n_tasks, + void * userdata, + bool inplace) { GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0); - bool is_node = false; - - if (!inplace && (a->grad || b->grad || c->grad)) { - is_node = true; - } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); struct ggml_map_custom3_op_params params = { @@ -8350,8 +7806,7 @@ static struct ggml_tensor * ggml_map_custom3_impl( }; ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); - result->op = GGML_OP_MAP_CUSTOM3; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_MAP_CUSTOM3; result->src[0] = a; result->src[1] = b; result->src[2] = c; @@ -8360,44 +7815,38 @@ static struct ggml_tensor * ggml_map_custom3_impl( } struct ggml_tensor * ggml_map_custom3( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - const ggml_custom3_op_t fun, - int n_tasks, - void * userdata) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + const ggml_custom3_op_t fun, + int n_tasks, + void * userdata) { return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, false); } struct ggml_tensor * ggml_map_custom3_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - const ggml_custom3_op_t fun, - int n_tasks, - void * userdata) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + const ggml_custom3_op_t fun, + int n_tasks, + void * userdata) { return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true); } // ggml_cross_entropy_loss struct ggml_tensor * ggml_cross_entropy_loss( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { GGML_ASSERT(ggml_are_same_shape(a, b)); - bool is_node = false; - - if (a->grad || b->grad) { - is_node = true; - } struct ggml_tensor * result = ggml_new_tensor_1d(ctx, a->type, 1); - result->op = GGML_OP_CROSS_ENTROPY_LOSS; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->op = GGML_OP_CROSS_ENTROPY_LOSS; result->src[0] = a; result->src[1] = b; @@ -8407,17 +7856,16 @@ struct ggml_tensor * ggml_cross_entropy_loss( // ggml_cross_entropy_loss_back struct ggml_tensor * ggml_cross_entropy_loss_back( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c) { GGML_ASSERT(ggml_are_same_shape(a, b)); GGML_ASSERT(ggml_is_scalar(c)); struct ggml_tensor * result = ggml_dup_tensor(ctx, a); - result->op = GGML_OP_CROSS_ENTROPY_LOSS_BACK; - result->grad = NULL; + result->op = GGML_OP_CROSS_ENTROPY_LOSS_BACK; result->src[0] = a; result->src[1] = b; result->src[2] = c; @@ -8430,12 +7878,14 @@ struct ggml_tensor * ggml_cross_entropy_loss_back( struct ggml_tensor * ggml_opt_step_adamw( struct ggml_context * ctx, struct ggml_tensor * a, + struct ggml_tensor * grad, float alpha, float beta1, float beta2, float eps, float wd) { - GGML_ASSERT(a->grad); + GGML_ASSERT(a->flags & GGML_TENSOR_FLAG_PARAM); + GGML_ASSERT(ggml_are_same_shape(a, grad)); GGML_ASSERT(alpha > 0.0f); GGML_ASSERT(beta1 >= 0.0f && beta1 <= 1.0f); GGML_ASSERT(beta2 >= 0.0f && beta2 <= 1.0f); @@ -8444,13 +7894,6 @@ struct ggml_tensor * ggml_opt_step_adamw( struct ggml_tensor * result = ggml_view_tensor(ctx, a); - result->op = GGML_OP_OPT_STEP_ADAMW; - result->grad = NULL; - result->src[0] = a; - result->src[1] = a->grad; - result->src[2] = ggml_dup_tensor(ctx, a->grad); - result->src[3] = ggml_dup_tensor(ctx, a->grad); - const int64_t iter = 1; memcpy(&result->op_params[0], &iter, sizeof(int64_t)); ggml_set_op_params_f32(result, 2, alpha); @@ -8459,26 +7902,17 @@ struct ggml_tensor * ggml_opt_step_adamw( ggml_set_op_params_f32(result, 5, eps); ggml_set_op_params_f32(result, 6, wd); + result->op = GGML_OP_OPT_STEP_ADAMW; + result->src[0] = a; + result->src[1] = grad; + result->src[2] = ggml_dup_tensor(ctx, grad); + result->src[3] = ggml_dup_tensor(ctx, grad); + return result; } //////////////////////////////////////////////////////////////////////////////// -void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor) { - tensor->flags |= GGML_TENSOR_FLAG_PARAM; - - GGML_ASSERT(tensor->grad == NULL); - tensor->grad = ggml_dup_tensor(ctx, tensor); - ggml_format_name(tensor->grad, "%s (grad)", tensor->name); -} - -void ggml_set_loss(struct ggml_tensor * tensor) { - GGML_ASSERT(ggml_is_scalar(tensor)); - GGML_ASSERT(tensor->type == GGML_TYPE_F32); - GGML_ASSERT(tensor->grad); - tensor->flags |= GGML_TENSOR_FLAG_LOSS; -} - // ggml_compute_forward_dup static void ggml_compute_forward_dup_same_cont( @@ -11394,6 +10828,86 @@ static void ggml_compute_forward_argmax( } } +// ggml_compute_forward_count_equal + +static void ggml_compute_forward_count_equal_i32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS; + + GGML_ASSERT(src0->type == GGML_TYPE_I32); + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_are_same_shape(src0, src1)); + GGML_ASSERT(ggml_is_scalar(dst)); + GGML_ASSERT(dst->type == GGML_TYPE_I64); + + const int64_t nr = ggml_nrows(src0); + + const int ith = params->ith; + const int nth = params->nth; + + int64_t * sums = (int64_t *) params->wdata; + int64_t sum_thread = 0; + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + for (int64_t ir = ir0; ir < ir1; ++ir) { + const int64_t i03 = ir / (ne02*ne01); + const int64_t i02 = (ir - i03*ne03) / ne01; + const int64_t i01 = ir - i03*ne03 - i02*ne02; + + const char * data0 = (const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01; + const char * data1 = (const char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11; + + for (int64_t i00 = 0; i00 < ne00; ++i00) { + const int32_t val0 = *((const int32_t *) (data0 + i00*nb00)); + const int32_t val1 = *((const int32_t *) (data1 + i00*nb10)); + + sum_thread += val0 == val1; + } + } + if (ith != 0) { + sums[ith] = sum_thread; + } + ggml_barrier(params->threadpool); + + if (ith != 0) { + return; + } + + for (int ith_other = 1; ith_other < nth; ++ith_other) { + sum_thread += sums[ith_other]; + } + *((int64_t *) dst->data) = sum_thread; +} + +static void ggml_compute_forward_count_equal( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_I32: + { + ggml_compute_forward_count_equal_i32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + // ggml_compute_forward_repeat static void ggml_compute_forward_repeat_f32( @@ -13357,6 +12871,10 @@ static void ggml_compute_forward_out_prod_f32( GGML_TENSOR_BINARY_OP_LOCALS + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + const int ith = params->ith; const int nth = params->nth; @@ -14686,7 +14204,7 @@ static void ggml_rope_cache_init( } } -GGML_CALL void ggml_rope_yarn_corr_dims( +void ggml_rope_yarn_corr_dims( int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2] ) { // start and end correction dims @@ -17436,41 +16954,40 @@ static void ggml_compute_forward_cross_entropy_loss_f32( const struct ggml_tensor * src0 = dst->src[0]; const struct ggml_tensor * src1 = dst->src[1]; - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(src1)); - GGML_ASSERT(ggml_is_scalar(dst)); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); + GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type)); GGML_ASSERT(ggml_are_same_shape(src0, src1)); + GGML_ASSERT(ggml_is_scalar(dst)); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + // TODO: handle transposed/permuted matrices + const int64_t nc = src0->ne[0]; + const int64_t nr = ggml_nrows(src0); const int ith = params->ith; const int nth = params->nth; - float * sums = (float *) params->wdata; - - // TODO: handle transposed/permuted matrices - const int nc = src0->ne[0]; - const int nr = ggml_nrows(src0); + float * sums = (float *) params->wdata; + float * st = ((float *) params->wdata) + nth + ith*nc; + float sum_thread = 0.0f; GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc)); - if (ith == 0) { - memset(sums, 0, sizeof(float) * (nth + nth * nc)); - } - ggml_barrier(params->threadpool); - // rows per thread - const int dr = (nr + nth - 1)/nth; + const int64_t dr = (nr + nth - 1)/nth; // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); - for (int i1 = ir0; i1 < ir1; i1++) { - float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]); - float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]); - float * st = ((float *) params->wdata) + nth + ith*nc; + for (int64_t i1 = ir0; i1 < ir1; ++i1) { + const float * s0 = (const float *)((const char *) src0->data + i1*src0->nb[1]); + const float * s1 = (const float *)((const char *) src1->data + i1*src1->nb[1]); #ifndef NDEBUG - for (int i = 0; i < nc; ++i) { + for (int64_t i = 0; i < nc; ++i) { //printf("p[%d] = %f\n", i, p[i]); assert(!isnan(s0[i])); assert(!isnan(s1[i])); @@ -17479,23 +16996,24 @@ static void ggml_compute_forward_cross_entropy_loss_f32( float max = -INFINITY; ggml_vec_max_f32(nc, &max, s0); - ggml_float sum = ggml_vec_log_soft_max_f32(nc, st, s0, max); - assert(sum >= 0.0); + const ggml_float sum_softmax = ggml_vec_log_soft_max_f32(nc, st, s0, max); + assert(sum_softmax >= 0.0); - ggml_vec_add1_f32(nc, st, st, -sum); + ggml_vec_add1_f32(nc, st, st, -sum_softmax); ggml_vec_mul_f32(nc, st, st, s1); - float st_sum = 0.0f; - ggml_vec_sum_f32(nc, &st_sum, st); - sums[ith] += st_sum; + float sum_st = 0.0f; + ggml_vec_sum_f32(nc, &sum_st, st); + sum_thread += sum_st; #ifndef NDEBUG - for (int i = 0; i < nc; ++i) { + for (int64_t i = 0; i < nc; ++i) { assert(!isnan(st[i])); assert(!isinf(st[i])); } #endif } + sums[ith] = sum_thread; ggml_barrier(params->threadpool); if (ith == 0) { @@ -17561,7 +17079,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]); #ifndef NDEBUG - for (int i = 0; i < nc; ++i) { + for (int64_t i = 0; i < nc; ++i) { //printf("p[%d] = %f\n", i, p[i]); assert(!isnan(s0[i])); assert(!isnan(s1[i])); @@ -17580,7 +17098,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( ggml_vec_scale_f32(nc, ds0, d_by_nr); #ifndef NDEBUG - for (int i = 0; i < nc; ++i) { + for (int64_t i = 0; i < nc; ++i) { assert(!isnan(ds0[i])); assert(!isinf(ds0[i])); } @@ -17768,6 +17286,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_argmax(params, tensor); } break; + case GGML_OP_COUNT_EQUAL: + { + ggml_compute_forward_count_equal(params, tensor); + } break; case GGML_OP_REPEAT: { ggml_compute_forward_repeat(params, tensor); @@ -18198,7 +17720,7 @@ void ggml_build_backward_gradient_checkpointing( struct ggml_tensor * * checkpoints, int n_checkpoints) { ggml_graph_cpy(gf, gb_tmp); - ggml_build_backward_expand(ctx, gf, gb_tmp, false, true); + ggml_build_backward_expand(ctx, gf, gb_tmp, false); if (n_checkpoints <= 0) { ggml_graph_cpy(gb_tmp, gb); @@ -18518,6 +18040,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor } break; case GGML_OP_MEAN: case GGML_OP_ARGMAX: + case GGML_OP_COUNT_EQUAL: { GGML_ABORT("fatal error"); // TODO: implement } @@ -18850,7 +18373,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor ggml_soft_max_back(ctx, tensor->grad, tensor), zero_table, acc_table); } - + GGML_ASSERT((!src1 || !src1->grad) && "backward pass for softmax mask not implemented"); } break; case GGML_OP_SOFT_MAX_BACK: { @@ -18891,6 +18414,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor beta_slow), zero_table, acc_table); } + GGML_ASSERT((!src2 || !src2->grad) && "gradients for freq factors not implemented"); } break; case GGML_OP_ROPE_BACK: { @@ -19012,6 +18536,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor } case GGML_OP_FLASH_ATTN_EXT: { + GGML_ABORT("FA backward pass not adapted after rework"); struct ggml_tensor * flash_grad = NULL; if (src0->grad || src1->grad || tensor->src[2]->grad) { int32_t t = ggml_get_op_params_i32(tensor, 0); @@ -19186,6 +18711,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor tensor->grad), zero_table, acc_table); } + GGML_ASSERT(!src1->grad && "backward pass for labels not implemented"); } break; case GGML_OP_CROSS_ENTROPY_LOSS_BACK: { @@ -19236,7 +18762,7 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * } } - if (node->op == GGML_OP_NONE && node->grad == NULL) { + if (node->op == GGML_OP_NONE && !(node->flags & GGML_TENSOR_FLAG_PARAM)) { // reached a leaf node, not part of the gradient graph (e.g. a constant) GGML_ASSERT(cgraph->n_leafs < cgraph->size); @@ -19254,9 +18780,6 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * } cgraph->nodes[cgraph->n_nodes] = node; - if (cgraph->grads) { - cgraph->grads[cgraph->n_nodes] = node->grad; - } cgraph->n_nodes++; } } @@ -19284,20 +18807,62 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * ggml_build_forward_impl(cgraph, tensor, true); } -void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate, bool keep) { +void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate) { GGML_ASSERT(gf->n_nodes > 0); GGML_ASSERT(gf->grads); - // if we are keeping the gradient graph, we have to detach the gradient nodes from the original graph - if (keep) { - for (int i = 0; i < gf->n_nodes; i++) { - struct ggml_tensor * node = gf->nodes[i]; + for (int i = 0; i < gf->n_nodes; ++i) { + struct ggml_tensor * node = gf->nodes[i]; - if (node->grad) { - node->grad = ggml_dup_tensor(ctx, node); - gf->grads[i] = node->grad; - } + if (node->type == GGML_TYPE_I32) { + continue; } + + bool needs_grad = node->flags & GGML_TENSOR_FLAG_PARAM; + bool ignore_src[GGML_MAX_SRC] = {false}; + switch (node->op) { + // gradients in node->src[0] for one reason or another have no effect on output gradients + case GGML_OP_IM2COL: // only used for its shape + case GGML_OP_IM2COL_BACK: // same as IM2COL + ignore_src[0] = true; + break; + case GGML_OP_UNARY: { + const enum ggml_unary_op uop = ggml_get_unary_op(node); + // SGN and STEP unary ops are piecewise constant + if (uop == GGML_UNARY_OP_SGN || uop == GGML_UNARY_OP_STEP) { + ignore_src[0] = true; + } + } break; + + // gradients in node->src[1] for one reason or another have no effect on output gradients + case GGML_OP_CPY: // gradients in CPY target are irrelevant + case GGML_OP_GET_ROWS: // row indices not differentiable + case GGML_OP_GET_ROWS_BACK: // same as for GET_ROWS + case GGML_OP_ROPE: // positions not differentiable + ignore_src[1] = true; + break; + + default: + break; + } + for (int j = 0; j < GGML_MAX_SRC; ++j) { + if (!node->src[j] || !node->src[j]->grad || ignore_src[j]) { + continue; + } + GGML_ASSERT(node->src[j]->type == GGML_TYPE_F32 || node->src[j]->type == GGML_TYPE_F16); + needs_grad = true; + break; + } + if (!needs_grad) { + continue; + } + + // inplace operations are currently not supported + GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW || + node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE); + + // create a new tensor with the same type and shape as the node and set it as grad + node->grad = ggml_dup_tensor(ctx, node); } // keep tables of original gradients for replacement/accumulation logic @@ -19359,7 +18924,7 @@ void ggml_build_opt_adamw( if (node->flags & GGML_TENSOR_FLAG_PARAM) { GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node); - struct ggml_tensor * opt_step = ggml_opt_step_adamw(ctx, node, alpha, beta1, beta2, eps, wd); + struct ggml_tensor * opt_step = ggml_opt_step_adamw(ctx, node, node->grad, alpha, beta1, beta2, eps, wd); ggml_build_forward_expand(gb, opt_step); } } @@ -19656,6 +19221,13 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_OP_SUM_ROWS: case GGML_OP_MEAN: case GGML_OP_ARGMAX: + { + n_tasks = 1; + } break; + case GGML_OP_COUNT_EQUAL: + { + n_tasks = n_threads; + } break; case GGML_OP_REPEAT: case GGML_OP_REPEAT_BACK: case GGML_OP_LEAKY_RELU: @@ -20154,6 +19726,10 @@ struct ggml_cplan ggml_graph_plan( cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks; } } break; + case GGML_OP_COUNT_EQUAL: + { + cur = ggml_type_size(node->type)*n_tasks; + } break; case GGML_OP_MUL_MAT: { const enum ggml_type vec_dot_type = type_traits[node->src[0]->type].vec_dot_type; @@ -20597,7 +20173,7 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl } #else if (n_threads > threadpool->n_threads_max) { - GGML_PRINT("WARNING: cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max); + GGML_LOG_WARN("cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max); n_threads = threadpool->n_threads_max; } @@ -21136,30 +20712,30 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context * } void ggml_graph_print(const struct ggml_cgraph * cgraph) { - GGML_PRINT("=== GRAPH ===\n"); + GGML_LOG_INFO("=== GRAPH ===\n"); - GGML_PRINT("n_nodes = %d\n", cgraph->n_nodes); + GGML_LOG_INFO("n_nodes = %d\n", cgraph->n_nodes); for (int i = 0; i < cgraph->n_nodes; i++) { struct ggml_tensor * node = cgraph->nodes[i]; - GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s\n", + GGML_LOG_INFO(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s\n", i, node->ne[0], node->ne[1], node->ne[2], ggml_op_name(node->op), (node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" : node->grad ? "g" : " "); } - GGML_PRINT("n_leafs = %d\n", cgraph->n_leafs); + GGML_LOG_INFO("n_leafs = %d\n", cgraph->n_leafs); for (int i = 0; i < cgraph->n_leafs; i++) { struct ggml_tensor * node = cgraph->leafs[i]; - GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s %16s\n", + GGML_LOG_INFO(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s %16s\n", i, node->ne[0], node->ne[1], ggml_op_name(node->op), ggml_get_name(node)); } - GGML_PRINT("========================================\n"); + GGML_LOG_INFO("========================================\n"); } // check if node is part of the graph @@ -21330,7 +20906,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph fclose(fp); - GGML_PRINT("%s: dot -Tpng %s -o %s.png && open %s.png\n", __func__, filename, filename, filename); + GGML_LOG_INFO("%s: dot -Tpng %s -o %s.png && open %s.png\n", __func__, filename, filename, filename); } //////////////////////////////////////////////////////////////////////////////// @@ -22162,8 +21738,6 @@ enum ggml_opt_result ggml_opt( struct ggml_context * ctx, struct ggml_opt_params params, struct ggml_tensor * f) { - GGML_ASSERT(f->grad && "ggml_set_param called for at least one parent tensor."); - bool free_ctx = false; if (ctx == NULL) { struct ggml_init_params params_ctx = { @@ -22204,7 +21778,7 @@ enum ggml_opt_result ggml_opt_resume( ggml_build_forward_expand(gf, f); struct ggml_cgraph * gb = ggml_graph_dup(ctx, gf); - ggml_build_backward_expand(ctx, gf, gb, false, true); + ggml_build_backward_expand(ctx, gf, gb, false); return ggml_opt_resume_g(ctx, opt, f, gf, gb, NULL, NULL); } @@ -22257,6 +21831,17 @@ void ggml_set_output(struct ggml_tensor * tensor) { tensor->flags |= GGML_TENSOR_FLAG_OUTPUT; } +void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor) { + GGML_UNUSED(ctx); // TODO: remove this parameter + tensor->flags |= GGML_TENSOR_FLAG_PARAM; +} + +void ggml_set_loss(struct ggml_tensor * tensor) { + GGML_ASSERT(ggml_is_scalar(tensor)); + GGML_ASSERT(tensor->type == GGML_TYPE_F32); + tensor->flags |= GGML_TENSOR_FLAG_LOSS; +} + //////////////////////////////////////////////////////////////////////////////// void ggml_quantize_init(enum ggml_type type) { @@ -23816,4 +23401,9 @@ int ggml_cpu_get_sve_cnt(void) { return 0; #endif } + +void ggml_log_set(ggml_log_callback log_callback, void * user_data) { + g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default; + g_logger_state.log_callback_user_data = user_data; +} //////////////////////////////////////////////////////////////////////////////// diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index ebe66a4a3..e08617ba2 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -814,6 +814,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FACTORS_LONG, + MODEL_TENSOR.ROPE_FACTORS_SHORT, MODEL_TENSOR.ATTN_NORM, MODEL_TENSOR.ATTN_QKV, MODEL_TENSOR.ATTN_Q, @@ -892,6 +894,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FACTORS_LONG, + MODEL_TENSOR.ROPE_FACTORS_SHORT, MODEL_TENSOR.ATTN_NORM, MODEL_TENSOR.ATTN_Q_A, MODEL_TENSOR.ATTN_Q_B, diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index e7e9b6fd5..f4a787c56 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -87,6 +87,9 @@ class TensorNameMap: "rope.freqs", # llama-pth "rotary_pos_emb.inv_freq", # chatglm ), + + MODEL_TENSOR.ROPE_FACTORS_LONG: (), + MODEL_TENSOR.ROPE_FACTORS_SHORT: (), } block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index dc5749913..f2645f921 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -122,8 +122,30 @@ class SpecialVocab: tokenizer = json.load(f) if self.load_merges: merges = tokenizer.get('model', {}).get('merges') - if isinstance(merges, list) and merges and isinstance(merges[0], str): - self.merges = merges + if isinstance(merges, list) and merges: + if isinstance(merges[0], str): + self.merges = merges + elif isinstance(merges[0], list) and len(merges[0]) == 2 and isinstance(merges[0][0], str): + # New format since transformers 4.45 to support spaces in merges + # ref: https://github.com/ggerganov/llama.cpp/issues/9692 + # TODO: internally store as the new format instead of converting to old + if any(' ' in s for pair in merges for s in pair): + logger.warning(f'Spaces in merges detected, encoding as {chr(ord(" ") + 256)!r}') + self.merges = [ + ' '.join( + [ + # ensure the spaces are properly encoded + ''.join( + chr(ord(c) + 256) if c == ' ' else c + for c in part + ) + for part in pair + ] + ) + for pair in merges + ] + else: + raise ValueError("Unknown tokenizer merges format") added_tokens = tokenizer.get('added_tokens', {}) else: added_tokens = {} diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh index f16336594..ffce2aab0 100755 --- a/scripts/sync-ggml-am.sh +++ b/scripts/sync-ggml-am.sh @@ -122,7 +122,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then # src/ggml-aarch64.h -> ggml/src/ggml-aarch64.h # src/ggml-alloc.c -> ggml/src/ggml-alloc.c # src/ggml-backend-impl.h -> ggml/src/ggml-backend-impl.h - # src/ggml-backend.c -> ggml/src/ggml-backend.c + # src/ggml-backend.cpp -> ggml/src/ggml-backend.cpp # src/ggml-cann/* -> ggml/src/ggml-cann/ # src/ggml-cann.cpp -> ggml/src/ggml-cann.cpp # src/ggml-common.h -> ggml/src/ggml-common.h @@ -169,7 +169,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.h/\1ggml\/src\/ggml-aarch64.h/g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-alloc\.c/\1ggml\/src\/ggml-alloc.c/g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-backend-impl\.h/\1ggml\/src\/ggml-backend-impl.h/g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-backend\.c/\1ggml\/src\/ggml-backend.c/g' \ + -e 's/([[:space:]]|[ab]\/)src\/ggml-backend\.cpp/\1ggml\/src\/ggml-backend.cpp/g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\.cpp/\1ggml\/src\/ggml-cann.cpp/g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-common\.h/\1ggml\/src\/ggml-common.h/g' \ diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index aa301462a..3cca9cc2f 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -9a24b8c8c40eab7262d067e91d08df160678df8d +564f42082f858f9674b2a2e06e9e779d9ed2c754 diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh index 30a62e088..f6ff5e683 100755 --- a/scripts/sync-ggml.sh +++ b/scripts/sync-ggml.sh @@ -9,7 +9,7 @@ cp -rpv ../ggml/src/ggml-aarch64.c ./ggml/src/ggml-aarch64.c cp -rpv ../ggml/src/ggml-aarch64.h ./ggml/src/ggml-aarch64.h cp -rpv ../ggml/src/ggml-alloc.c ./ggml/src/ggml-alloc.c cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml/src/ggml-backend-impl.h -cp -rpv ../ggml/src/ggml-backend.c ./ggml/src/ggml-backend.c +cp -rpv ../ggml/src/ggml-backend.cpp ./ggml/src/ggml-backend.cpp cp -rpv ../ggml/src/ggml-cann/* ./ggml/src/ggml-cann/ cp -rpv ../ggml/src/ggml-cann.cpp ./ggml/src/ggml-cann.cpp cp -rpv ../ggml/src/ggml-common.h ./ggml/src/ggml-common.h diff --git a/src/llama-vocab.h b/src/llama-vocab.h index 069bdc423..28bad9135 100644 --- a/src/llama-vocab.h +++ b/src/llama-vocab.h @@ -40,17 +40,17 @@ struct llama_vocab { id special_bos_id = 1; id special_eos_id = 2; id special_unk_id = 0; - id special_sep_id = -1; - id special_pad_id = -1; - id special_cls_id = -1; - id special_mask_id = -1; + id special_sep_id = LLAMA_TOKEN_NULL; + id special_pad_id = LLAMA_TOKEN_NULL; + id special_cls_id = LLAMA_TOKEN_NULL; + id special_mask_id = LLAMA_TOKEN_NULL; id linefeed_id = 13; - id special_prefix_id = -1; - id special_suffix_id = -1; - id special_middle_id = -1; - id special_eot_id = -1; // TODO: move above after "eos_id", and here add "file separator" token - id special_eom_id = -1; + id special_prefix_id = LLAMA_TOKEN_NULL; + id special_suffix_id = LLAMA_TOKEN_NULL; + id special_middle_id = LLAMA_TOKEN_NULL; + id special_eot_id = LLAMA_TOKEN_NULL; // TODO: move above after "eos_id", and here add "file separator" token + id special_eom_id = LLAMA_TOKEN_NULL; // set of all tokens that cause "end of generation" std::set special_eog_ids; diff --git a/src/llama.cpp b/src/llama.cpp index 8a34e205f..77df74723 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -12,9 +12,7 @@ # include "ggml-rpc.h" #endif -#ifdef GGML_USE_CUDA -# include "ggml-cuda.h" -#elif defined(GGML_USE_VULKAN) +#if defined(GGML_USE_VULKAN) # include "ggml-vulkan.h" #elif defined(GGML_USE_SYCL) # include "ggml-sycl.h" @@ -28,10 +26,6 @@ # include "ggml-blas.h" #endif -#ifdef GGML_USE_METAL -# include "ggml-metal.h" -#endif - // TODO: replace with ggml API call #define QK_K 256 @@ -610,7 +604,7 @@ enum llm_tensor { LLM_TENSOR_CLS_OUT, }; -static const std::map> LLM_TENSOR_NAMES = { +static const std::map> LLM_TENSOR_NAMES = { { LLM_ARCH_LLAMA, { @@ -1566,32 +1560,32 @@ struct LLM_TN { return LLM_TENSOR_NAMES.at(arch).at(tensor); } - std::string operator()(llm_tensor tensor, const std::string & suffix) const { + std::string operator()(llm_tensor tensor, const char * suffix) const { if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return LLM_TENSOR_NAMES.at(arch).at(tensor) + "." + suffix; + return std::string(LLM_TENSOR_NAMES.at(arch).at(tensor)) + "." + suffix; } std::string operator()(llm_tensor tensor, int bid) const { if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid); + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor), bid); } - std::string operator()(llm_tensor tensor, const std::string & suffix, int bid) const { + std::string operator()(llm_tensor tensor, const char * suffix, int bid) const { if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid) + "." + suffix; + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor), bid) + "." + suffix; } - std::string operator()(llm_tensor tensor, const std::string & suffix, int bid, int xid) const { + std::string operator()(llm_tensor tensor, const char * suffix, int bid, int xid) const { if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid, xid) + "." + suffix; + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor), bid, xid) + "." + suffix; } }; @@ -2264,59 +2258,16 @@ static std::string llama_token_to_piece(const struct llama_model * model, llama_ return piece; } -static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer) { - ggml_backend_buffer_type_t buft = nullptr; - -#if defined(GGML_USE_CUDA) - // host buffers should only be used when data is expected to be copied to/from the GPU - if (host_buffer) { - buft = ggml_backend_cuda_host_buffer_type(); - } -#elif defined(GGML_USE_SYCL) - if (host_buffer) { - buft = ggml_backend_sycl_host_buffer_type(); - } -#elif defined(GGML_USE_CANN) - if (host_buffer) { - buft = ggml_backend_cann_host_buffer_type(); - } -#elif defined(GGML_USE_CPU_HBM) - buft = ggml_backend_cpu_hbm_buffer_type(); -#elif defined(GGML_USE_VULKAN) - if (host_buffer) { - buft = ggml_backend_vk_host_buffer_type(); - } -#endif - - if (buft == nullptr) { - buft = ggml_backend_cpu_buffer_type(); - } - return buft; - - GGML_UNUSED(host_buffer); -} - // // globals // -struct llama_state { - llama_state() { -#ifdef GGML_USE_METAL - ggml_backend_metal_log_set_callback(log_callback, log_callback_user_data); -#elif defined(GGML_USE_CUDA) - ggml_backend_cuda_log_set_callback(log_callback, log_callback_user_data); -#elif defined(GGML_USE_CANN) - ggml_backend_cann_log_set_callback(log_callback, log_callback_user_data); -#endif - } - - // We save the log callback globally +struct llama_logger_state { ggml_log_callback log_callback = llama_log_callback_default; void * log_callback_user_data = nullptr; }; -static llama_state g_state; +static llama_logger_state g_logger_state; // available llama models enum e_model { @@ -2457,7 +2408,7 @@ struct llama_hparams { // needed by encoder-decoder models (e.g. T5, FLAN-T5) // ref: https://github.com/ggerganov/llama.cpp/pull/8141 - llama_token dec_start_token_id = -1; + llama_token dec_start_token_id = LLAMA_TOKEN_NULL; enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; @@ -2920,14 +2871,17 @@ struct llama_model { std::vector layers; + // gguf metadata + std::unordered_map gguf_kv; + llama_split_mode split_mode; int main_gpu; int n_gpu_layers; - std::vector rpc_servers; + // list of devices used in this model + std::vector devices; - // gguf metadata - std::unordered_map gguf_kv; + std::vector rpc_servers; // layer -> buffer type mapping struct layer_buft { @@ -2970,11 +2924,6 @@ struct llama_model { ggml_free(ctx); } for (ggml_backend_buffer_t buf : bufs) { -#ifdef GGML_USE_CUDA - if (ggml_backend_buffer_get_type(buf) == ggml_backend_cpu_buffer_type()) { - ggml_backend_cuda_unregister_host_buffer(ggml_backend_buffer_get_base(buf)); - } -#endif ggml_backend_buffer_free(buf); } while (!lora_adapters.empty()) { @@ -3339,9 +3288,6 @@ struct llama_context { std::unordered_map lora_adapters; std::vector backends; -#ifdef GGML_USE_METAL - ggml_backend_t backend_metal = nullptr; -#endif #ifdef GGML_USE_BLAS ggml_backend_t backend_blas = nullptr; #endif @@ -3460,72 +3406,112 @@ struct llama_lora_adapter { } }; -static size_t llama_get_device_count(const llama_model & model) { - size_t count = 1; -#if defined(GGML_USE_CUDA) - count = ggml_backend_cuda_get_device_count(); -#elif defined(GGML_USE_SYCL) - count = ggml_backend_sycl_get_device_count(); -#elif defined(GGML_USE_VULKAN) - count = ggml_backend_vk_get_device_count(); -#elif defined(GGML_USE_CANN) - return ggml_backend_cann_get_device_count(); -#endif +static int llama_get_device_count(const llama_model & model) { + int count = (int) model.devices.size(); + #if defined(GGML_USE_RPC) - count += model.rpc_servers.size(); + count += (int) model.rpc_servers.size(); #endif + +#if defined(GGML_USE_SYCL) + count += ggml_backend_sycl_get_device_count(); +#elif defined(GGML_USE_VULKAN) + count += ggml_backend_vk_get_device_count(); +#elif defined(GGML_USE_CANN) + count += ggml_backend_cann_get_device_count(); +#endif + return count; + GGML_UNUSED(model); } -static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int gpu) { +static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(const llama_model & model, bool host_buffer) { ggml_backend_buffer_type_t buft = nullptr; -#ifdef GGML_USE_RPC - int rpc_count = (int)model.rpc_servers.size(); -#else - int rpc_count = 0; -#endif - int local_gpu = gpu - rpc_count; -#if defined(GGML_USE_RPC) - if (gpu < rpc_count) { - const char * endpoint = model.rpc_servers[gpu].c_str(); - return ggml_backend_rpc_buffer_type(endpoint); + if (host_buffer) { + for (auto * dev : model.devices) { + buft = ggml_backend_dev_host_buffer_type(dev); + if (buft != nullptr) { + break; + } + } } -#endif -#if defined(GGML_USE_METAL) - buft = ggml_backend_metal_buffer_type(); -#elif defined(GGML_USE_CUDA) - buft = ggml_backend_cuda_buffer_type(local_gpu); -#elif defined(GGML_USE_VULKAN) - buft = ggml_backend_vk_buffer_type(local_gpu); -#elif defined(GGML_USE_SYCL) - buft = ggml_backend_sycl_buffer_type(local_gpu); -#elif defined(GGML_USE_KOMPUTE) - buft = ggml_backend_kompute_buffer_type(local_gpu); - if (buft == nullptr) { - LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, local_gpu); + +#if defined(GGML_USE_SYCL) + if (host_buffer) { + buft = ggml_backend_sycl_host_buffer_type(); } #elif defined(GGML_USE_CANN) - buft = ggml_backend_cann_buffer_type(local_gpu); + if (host_buffer) { + buft = ggml_backend_cann_host_buffer_type(); + } +#elif defined(GGML_USE_CPU_HBM) + buft = ggml_backend_cpu_hbm_buffer_type(); +#elif defined(GGML_USE_VULKAN) + if (host_buffer) { + buft = ggml_backend_vk_host_buffer_type(); + } #endif if (buft == nullptr) { - buft = llama_default_buffer_type_cpu(true); + buft = ggml_backend_cpu_buffer_type(); } return buft; + + GGML_UNUSED(host_buffer); +} + +static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int device) { + ggml_backend_buffer_type_t buft = nullptr; + +#if defined(GGML_USE_RPC) + int rpc_count = (int)model.rpc_servers.size(); + if (device < rpc_count) { + const char * endpoint = model.rpc_servers[device].c_str(); + return ggml_backend_rpc_buffer_type(endpoint); + } + device -= rpc_count; +#endif + + if (device < (int)model.devices.size()) { + return ggml_backend_dev_buffer_type(model.devices[device]); + } + device -= (int)model.devices.size(); + +#if defined(GGML_USE_VULKAN) + buft = ggml_backend_vk_buffer_type(device); +#elif defined(GGML_USE_SYCL) + buft = ggml_backend_sycl_buffer_type(device); +#elif defined(GGML_USE_KOMPUTE) + buft = ggml_backend_kompute_buffer_type(device); +#elif defined(GGML_USE_CANN) + buft = ggml_backend_cann_buffer_type(device); +#endif + + if (buft == nullptr) { + buft = llama_default_buffer_type_cpu(model, true); + } + return buft; + GGML_UNUSED(model); - GGML_UNUSED(local_gpu); } static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_model & model, int fallback_gpu, const float * tensor_split) { ggml_backend_buffer_type_t buft = nullptr; -#ifdef GGML_USE_CUDA - if (ggml_backend_cuda_get_device_count() > 1) { - buft = ggml_backend_cuda_split_buffer_type(tensor_split); + // find a backend that supports split buffers + for (size_t i = 0; i < ggml_backend_reg_count(); ++i) { + ggml_backend_reg_t reg = ggml_backend_reg_get(i); + + auto ggml_backend_split_buffer_type_fn = (ggml_backend_split_buffer_type_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_split_buffer_type"); + if (ggml_backend_split_buffer_type_fn) { + buft = ggml_backend_split_buffer_type_fn(tensor_split); + if (buft != nullptr) { + break; + } + } } -#endif #ifdef GGML_USE_SYCL if (ggml_backend_sycl_get_device_count() > 1) { @@ -3542,13 +3528,8 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_mo } static size_t llama_get_device_memory(const llama_model & model, int device) { -#ifdef GGML_USE_RPC - int rpc_count = (int)model.rpc_servers.size(); -#else - int rpc_count = 0; -#endif - int local_device = device - rpc_count; #if defined(GGML_USE_RPC) + int rpc_count = (int)model.rpc_servers.size(); if (device < rpc_count) { size_t total; size_t free; @@ -3556,32 +3537,37 @@ static size_t llama_get_device_memory(const llama_model & model, int device) { ggml_backend_rpc_get_device_memory(endpoint, &free, &total); return free; } + device = device - rpc_count; #endif -#if defined(GGML_USE_CUDA) + + if (device < (int)model.devices.size()) { + ggml_backend_dev_t dev = model.devices[device]; + size_t total; + size_t free; + ggml_backend_dev_memory(dev, &free, &total); + return free; + } + +#if defined(GGML_USE_SYCL) size_t total; size_t free; - ggml_backend_cuda_get_device_memory(local_device, &free, &total); - return free; -#elif defined(GGML_USE_SYCL) - size_t total; - size_t free; - ggml_backend_sycl_get_device_memory(local_device, &free, &total); + ggml_backend_sycl_get_device_memory(device, &free, &total); return free; #elif defined(GGML_USE_VULKAN) size_t total; size_t free; - ggml_backend_vk_get_device_memory(local_device, &free, &total); + ggml_backend_vk_get_device_memory(device, &free, &total); return free; #elif defined(GGML_USE_CANN) size_t total; size_t free; - ggml_backend_cann_get_device_memory(local_device, &free, &total); + ggml_backend_cann_get_device_memory(device, &free, &total); return free; #else return 1; #endif GGML_UNUSED(model); - GGML_UNUSED(local_device); + GGML_UNUSED(device); } // @@ -3624,7 +3610,7 @@ static bool llama_kv_cache_init( buft_layer_count[model.buft_layer[i].buft]++; } } else { - buft_layer_count[llama_default_buffer_type_cpu(true)] = n_layer; + buft_layer_count[llama_default_buffer_type_cpu(model, true)] = n_layer; } // create a context for each buffer type @@ -4916,7 +4902,7 @@ struct llama_model_loader { static const int TENSOR_NOT_REQUIRED = 1; static const int TENSOR_DUPLICATED = 2; - struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector & ne, int flags = 0) { + struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::initializer_list & ne, int flags = 0) { const struct ggml_tensor * cur = check_tensor_dims(name, ne, !(flags & TENSOR_NOT_REQUIRED)); if (cur == NULL) { @@ -4926,7 +4912,7 @@ struct llama_model_loader { return create_tensor_for(ctx, cur, flags & TENSOR_DUPLICATED); } - struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::vector & ne, size_t offset, bool required = true) { + struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::initializer_list & ne, size_t offset, bool required = true) { const struct ggml_tensor * cur = check_tensor_dims(name, ne, required); if (cur == NULL) { @@ -4939,7 +4925,7 @@ struct llama_model_loader { std::array dims; for (size_t i = 0; i < GGML_MAX_DIMS; ++i) { - dims[i] = i < ne.size() ? ne[i] : 1; + dims[i] = i < ne.size() ? ne.begin()[i] : 1; } struct ggml_tensor * tensor = ggml_view_4d(ctx, base, @@ -5037,7 +5023,7 @@ struct llama_model_loader { // Returns false if cancelled by progress_callback bool load_all_data( struct ggml_context * ctx, - llama_buf_map & bufs_mmap, + llama_buf_map & bufs, llama_mlocks * lmlocks, llama_progress_callback progress_callback, void * progress_callback_user_data) { @@ -5046,79 +5032,94 @@ struct llama_model_loader { std::vector> read_buf; std::vector>> validation_result; -#if defined(GGML_USE_CUDA) // 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives. // NVMe raid configurations might require more / larger buffers. constexpr size_t n_buffers = 4; constexpr size_t buffer_size = 1 * 1024 * 1024; // 1MB std::vector host_buffers; - std::vector host_ptrs; std::vector events; + std::vector host_ptrs; size_t buffer_idx = 0; // buffer to use for async loads - - ggml_backend_t cuda_backend = nullptr; - if (!use_mmap && !check_tensors) { + ggml_backend_t upload_backend = [&](const char * fn) -> ggml_backend_t { + if (use_mmap || check_tensors) { + return nullptr; + } // When not using mmaped io use async uploads from pinned memory to GPU memory. - // First determine if the CUDA backend is active, and if so, determine the device ID. - ggml_backend_buffer_t buf = bufs_mmap.count(0) ? bufs_mmap.at(0) : nullptr; - if (buf) { - ggml_backend_buffer_type_t buffer_type = ggml_backend_buffer_get_type(buf); - for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) { - auto * cuda_buffer_type = ggml_backend_cuda_buffer_type(i); - if (buffer_type == cuda_buffer_type) { - cuda_backend = ggml_backend_cuda_init(i); - break; - } - } + // First determine if the backend supports the necessary features for async uploads. + auto * buf = bufs.count(0) ? bufs.at(0) : nullptr; + if (!buf) { + LLAMA_LOG_DEBUG("%s: no buffer found for async uploads\n", fn); + return nullptr; } - // If the cuda backend is active create pinned memory buffers and events for synchronisation. - if (cuda_backend) { - for (size_t idx = 0; idx < n_buffers; ++idx) { - host_buffers.emplace_back(ggml_backend_buft_alloc_buffer(llama_default_buffer_type_cpu(true), buffer_size)); - host_ptrs.emplace_back(ggml_backend_buffer_get_base(host_buffers[idx])); - events.emplace_back(ggml_backend_event_new(cuda_backend)); - } + auto * buft = ggml_backend_buffer_get_type(buf); + auto * dev = ggml_backend_buft_get_device(buft); + if (!dev) { + LLAMA_LOG_DEBUG("%s: no device found for buffer type %s for async uploads\n", fn, + ggml_backend_buft_name(buft)); + return nullptr; } + + if (buft != ggml_backend_dev_buffer_type(dev)) { + LLAMA_LOG_DEBUG("%s: buffer type %s is not the default buffer type for device %s for async uploads\n", fn, + ggml_backend_buft_name(buft), ggml_backend_dev_name(dev)); + return nullptr; + } + + ggml_backend_dev_props props; + ggml_backend_dev_get_props(dev, &props); + if (!props.caps.async || !props.caps.host_buffer || !props.caps.events) { + LLAMA_LOG_DEBUG("%s: device %s does not support async, host buffers or events\n", fn, + ggml_backend_dev_name(dev)); + return nullptr; + } + + auto * host_buft = ggml_backend_dev_host_buffer_type(dev); + if (!host_buft) { + LLAMA_LOG_DEBUG("%s: no host buffer type found for device %s\n", fn, + ggml_backend_dev_name(dev)); + return nullptr; + } + + // If the backend is supported, create pinned memory buffers and events for synchronisation. + for (size_t idx = 0; idx < n_buffers; ++idx) { + auto * buf = ggml_backend_buft_alloc_buffer(host_buft, buffer_size); + if (!buf) { + LLAMA_LOG_DEBUG("%s: failed to allocate host buffer for async uploads for device %s\n", fn, + ggml_backend_dev_name(dev)); + return nullptr; + } + + host_buffers.emplace_back(buf); + host_ptrs.emplace_back(ggml_backend_buffer_get_base(buf)); + + auto * event = ggml_backend_event_new(dev); + if (!event) { + LLAMA_LOG_DEBUG("%s: failed to create event for async uploads for device %s\n", fn, + ggml_backend_dev_name(dev)); + return nullptr; + } + + events.emplace_back(event); + } + + ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr); + if (!backend) { + LLAMA_LOG_DEBUG("%s: failed to initialize backend for device %s for async uploads\n", fn, + ggml_backend_dev_name(dev)); + return nullptr; + } + + return backend; + }(__func__); + + if (upload_backend) { + LLAMA_LOG_DEBUG("%s: using async uploads for device %s, buffer type %s, backend %s\n", __func__, + ggml_backend_dev_name(ggml_backend_get_device(upload_backend)), + ggml_backend_buft_name(ggml_backend_buffer_get_type(bufs.at(0))), + ggml_backend_name(upload_backend)); } -#elif defined(GGML_USE_SYCL) - // 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives. - // NVMe raid configurations might require more / larger buffers. - constexpr size_t n_buffers = 4; - constexpr size_t buffer_size = 1 * 1024 * 1024; // 1MB - - std::vector host_buffers; - std::vector host_ptrs; - std::vector events; - size_t buffer_idx = 0; // buffer to use for async loads - - ggml_backend_t sycl_backend = nullptr; - if (!use_mmap && !check_tensors) { - // When not using mmaped io use async uploads from pinned memory to GPU memory. - // First determine if the SYCL backend is active, and if so, determine the device ID. - ggml_backend_buffer_t buf = bufs_mmap.count(0) ? bufs_mmap.at(0) : nullptr; - if (buf) { - ggml_backend_buffer_type_t buffer_type = ggml_backend_buffer_get_type(buf); - for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) { - auto * sycl_buffer_type = ggml_backend_sycl_buffer_type(i); - if (buffer_type == sycl_buffer_type) { - sycl_backend = ggml_backend_sycl_init(i); - break; - } - } - } - - // If the sycl backend is active create pinned memory buffers and events for synchronisation. - if (sycl_backend) { - for (size_t idx = 0; idx < n_buffers; ++idx) { - host_buffers.emplace_back(ggml_backend_buft_alloc_buffer(llama_default_buffer_type_cpu(true), buffer_size)); - host_ptrs.emplace_back(ggml_backend_buffer_get_base(host_buffers[idx])); - events.emplace_back(ggml_backend_event_new(sycl_backend)); - } - } - } -#endif for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) { const auto * weight = get_weight(ggml_get_name(cur)); @@ -5138,8 +5139,8 @@ struct llama_model_loader { if (use_mmap) { const auto & mapping = mappings.at(weight->idx); ggml_backend_buffer_t buf_mmap = nullptr; - if (bufs_mmap.count(weight->idx)) { - buf_mmap = bufs_mmap.at(weight->idx); + if (bufs.count(weight->idx)) { + buf_mmap = bufs.at(weight->idx); } uint8_t * data = (uint8_t *) mapping->addr + weight->offs; @@ -5175,9 +5176,8 @@ struct llama_model_loader { })); } } else { -#if defined(GGML_USE_CUDA) - // If cuda_backend is valid load the tensor in chunks to pinned memory and upload the buffers asynchronously to the GPU. - if (cuda_backend) { + // If upload_backend is valid load the tensor in chunks to pinned memory and upload the buffers asynchronously to the GPU. + if (upload_backend) { file->seek(weight->offs, SEEK_SET); size_t bytes_read = 0; @@ -5187,40 +5187,14 @@ struct llama_model_loader { ggml_backend_event_synchronize(events[buffer_idx]); file->read_raw(host_ptrs[buffer_idx], read_iteration); - ggml_backend_tensor_set_async(cuda_backend, cur, host_ptrs[buffer_idx], bytes_read, read_iteration); - ggml_backend_event_record(events[buffer_idx]); + ggml_backend_tensor_set_async(upload_backend, cur, host_ptrs[buffer_idx], bytes_read, read_iteration); + ggml_backend_event_record(events[buffer_idx], upload_backend); + bytes_read += read_iteration; ++buffer_idx; buffer_idx %= n_buffers; } - - } - else -#elif defined(GGML_USE_SYCL) - // If sycl_backend is valid load the tensor in chunks to - // pinned memory and upload the buffers asynchronously to the - // GPU. - if (sycl_backend) { - file->seek(weight->offs, SEEK_SET); - - size_t bytes_read = 0; - - while (bytes_read < n_size) { - size_t read_iteration = - std::min(buffer_size, n_size - bytes_read); - ggml_backend_event_synchronize(events[buffer_idx]); - file->read_raw(host_ptrs[buffer_idx], read_iteration); - ggml_backend_tensor_set_async(sycl_backend, cur, - host_ptrs[buffer_idx], - bytes_read, read_iteration); - ggml_backend_event_record(events[buffer_idx]); - bytes_read += read_iteration; - ++buffer_idx; - buffer_idx %= n_buffers; - } - } else -#endif - { + } else { read_buf.resize(n_size); file->seek(weight->offs, SEEK_SET); file->read_raw(read_buf.data(), n_size); @@ -5235,27 +5209,15 @@ struct llama_model_loader { size_done += n_size; } -#if defined(GGML_USE_CUDA) - // free temporary resources used for async cuda uploads - if (cuda_backend) { - for (size_t idx = 0; idx < n_buffers;++idx) { - ggml_backend_event_synchronize(events[idx]); - ggml_backend_event_free(events[idx]); - ggml_backend_buffer_free(host_buffers[idx]); - } - ggml_backend_free(cuda_backend); + // free temporary resources used for async uploads + for (auto * event : events) { + ggml_backend_event_synchronize(event); + ggml_backend_event_free(event); } -#elif defined(GGML_USE_SYCL) - // free temporary resources used for async cuda uploads - if (sycl_backend) { - for (size_t idx = 0; idx < n_buffers;++idx) { - ggml_backend_event_synchronize(events[idx]); - ggml_backend_event_free(events[idx]); - ggml_backend_buffer_free(host_buffers[idx]); - } - ggml_backend_free(sycl_backend); + for (auto * buf : host_buffers) { + ggml_backend_buffer_free(buf); } -#endif + ggml_backend_free(upload_backend); // check validation results bool validation_failed = false; @@ -5571,8 +5533,10 @@ static void llm_load_hparams( } } else { switch (hparams.n_layer) { + case 16: model.type = e_model::MODEL_1B; break; // Llama 3.2 1B case 22: model.type = e_model::MODEL_1B; break; case 26: model.type = e_model::MODEL_3B; break; + case 28: model.type = e_model::MODEL_3B; break; // Llama 3.2 3B // granite uses a vocab with len 49152 case 32: model.type = hparams.n_vocab == 49152 ? e_model::MODEL_3B : (hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B); break; case 36: model.type = e_model::MODEL_8B; break; // granite @@ -6989,6 +6953,13 @@ static bool llm_load_tensors( void * progress_callback_user_data) { auto & hparams = model.hparams; + // check if the value of main_gpu is valid + if (llama_get_device_count(model) > 0 && + split_mode != LLAMA_SPLIT_MODE_LAYER && + (main_gpu < 0 || main_gpu >= llama_get_device_count(model))) { + throw std::runtime_error(format("invalid value for main_gpu: %d (available devices: %d)", main_gpu, llama_get_device_count(model))); + } + model.split_mode = split_mode; model.main_gpu = main_gpu; model.n_gpu_layers = n_gpu_layers; @@ -6998,14 +6969,14 @@ static bool llm_load_tensors( bool use_mmap_buffer = true; // there is very little benefit to offloading the input layer, so always keep it on the CPU - model.buft_input = llama_default_buffer_type_cpu(true); + model.buft_input = llama_default_buffer_type_cpu(model, true); //model.buft_input = llama_default_buffer_type_offload(main_gpu); model.buft_layer.resize(n_layer); // assign cpu layers for (int i = 0; i < i_gpu_start; ++i) { - model.buft_layer[i] = llama_default_buffer_type_cpu(true); + model.buft_layer[i] = llama_default_buffer_type_cpu(model, true); } if (split_mode == LLAMA_SPLIT_MODE_LAYER) { @@ -7043,7 +7014,7 @@ static bool llm_load_tensors( int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(act_gpu_layers - 1)/act_gpu_layers) - splits.begin(); model.buft_output = llama_default_buffer_type_offload(model, layer_gpu); } else { - model.buft_output = llama_default_buffer_type_cpu(true); + model.buft_output = llama_default_buffer_type_cpu(model, true); } } else { ggml_backend_buffer_type_t split_buft; @@ -7067,7 +7038,7 @@ static bool llm_load_tensors( llama_default_buffer_type_offload(model, main_gpu) }; } else { - model.buft_output = llama_default_buffer_type_cpu(true); + model.buft_output = llama_default_buffer_type_cpu(model, true); } } @@ -8936,55 +8907,39 @@ static bool llm_load_tensors( llama_buf_map bufs; bufs.reserve(n_max_backend_buffer); - // only the mmap region containing the tensors in the model is mapped to the backend buffer - // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers - // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size - if (ml.use_mmap && use_mmap_buffer && buft == llama_default_buffer_type_cpu(true)) { + // check if this backend device supports buffer_from_host_ptr + ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft); + bool buffer_from_host_ptr_supported = false; + if (dev) { + ggml_backend_dev_props props; + ggml_backend_dev_get_props(dev, &props); + buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr; + } + + if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported) { for (uint32_t idx = 0; idx < ml.files.size(); idx++) { + // only the mmap region containing the tensors in the model is mapped to the backend buffer + // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers + // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size void * addr = nullptr; - size_t first, last; + size_t first, last; // NOLINT ml.get_mapping_range(&first, &last, &addr, idx, ctx); if (first >= last) { continue; } - ggml_backend_buffer_t buf = ggml_backend_cpu_buffer_from_ptr((char *) addr + first, last - first); - if (buf == nullptr) { - throw std::runtime_error("unable to allocate backend CPU buffer"); - } - model.bufs.push_back(buf); - bufs.emplace(idx, buf); -#ifdef GGML_USE_CUDA - if (n_layer >= n_gpu_layers) { - ggml_backend_cuda_register_host_buffer( - ggml_backend_buffer_get_base(buf), - ggml_backend_buffer_get_size(buf)); - } -#endif - } - } -#ifdef GGML_USE_METAL - else if (ml.use_mmap && use_mmap_buffer && buft == ggml_backend_metal_buffer_type()) { - for (uint32_t idx = 0; idx < ml.files.size(); idx++) { const size_t max_size = ggml_get_max_tensor_size(ctx); - void * addr = nullptr; - size_t first, last; - ml.get_mapping_range(&first, &last, &addr, idx, ctx); - if (first >= last) { - continue; - } - ggml_backend_buffer_t buf = ggml_backend_metal_buffer_from_ptr((char *) addr + first, last - first, max_size); + ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size); if (buf == nullptr) { - throw std::runtime_error("unable to allocate backend metal buffer"); + throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); } model.bufs.push_back(buf); bufs.emplace(idx, buf); } } -#endif else { ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); if (buf == nullptr) { - throw std::runtime_error("unable to allocate backend buffer"); + throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); } model.bufs.push_back(buf); if (use_mlock && ggml_backend_buffer_is_host(buf)) { @@ -17023,7 +16978,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { lctx.embd = nullptr; } - lctx.buf_output = ggml_backend_buft_alloc_buffer(llama_default_buffer_type_cpu(true), new_size); + lctx.buf_output = ggml_backend_buft_alloc_buffer(llama_default_buffer_type_cpu(lctx.model, true), new_size); if (lctx.buf_output == nullptr) { LLAMA_LOG_ERROR("%s: failed to allocate output buffer of size %.2f MiB\n", __func__, new_size / (1024.0 * 1024.0)); return 0; @@ -17092,12 +17047,6 @@ static void llama_graph_compute( ggml_cgraph * gf, int n_threads, ggml_threadpool * threadpool) { -#ifdef GGML_USE_METAL - if (ggml_backend_is_metal(lctx.backend_metal)) { - ggml_backend_metal_set_n_cb(lctx.backend_metal, n_threads); - } -#endif - if (lctx.backend_cpu != nullptr) { ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads); ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool); @@ -19060,21 +19009,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() { } size_t llama_max_devices(void) { -#if defined(GGML_USE_RPC) - return GGML_RPC_MAX_SERVERS; -#elif defined(GGML_USE_METAL) - return 1; -#elif defined(GGML_USE_CUDA) - return GGML_CUDA_MAX_DEVICES; -#elif defined(GGML_USE_SYCL) - return GGML_SYCL_MAX_DEVICES; -#elif defined(GGML_USE_VULKAN) - return GGML_VK_MAX_DEVICES; -#elif defined(GGML_USE_CANN) - return GGML_CANN_MAX_DEVICES; -#else - return 1; -#endif + return 16; } bool llama_supports_mmap(void) { @@ -19086,12 +19021,13 @@ bool llama_supports_mlock(void) { } bool llama_supports_gpu_offload(void) { -#if defined(GGML_USE_CUDA) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \ +#if defined(GGML_USE_VULKAN) || \ defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_RPC) // Defined when llama.cpp is compiled with support for offloading model layers to GPU. return true; #else - return false; + return ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU) != nullptr || + ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL) != nullptr; #endif } @@ -19156,17 +19092,30 @@ struct llama_model * llama_load_model_from_file( return true; }; } + if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') { // split the servers set them into model->rpc_servers std::string servers(params.rpc_servers); size_t pos = 0; - while ((pos = servers.find(",")) != std::string::npos) { + while ((pos = servers.find(',')) != std::string::npos) { std::string server = servers.substr(0, pos); model->rpc_servers.push_back(server); servers.erase(0, pos + 1); } model->rpc_servers.push_back(servers); } + + // create list of devices to use with this model + // currently, we use all available devices + // TODO: rework API to give user more control over device selection + for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { + ggml_backend_dev_t dev = ggml_backend_dev_get(i); + // skip the CPU backend since it is handled separately + if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU_FULL) { + model->devices.push_back(dev); + } + } + int status = llama_model_load(path_model, *model, params); GGML_ASSERT(status <= 0); if (status < 0) { @@ -19328,6 +19277,36 @@ struct llama_context * llama_new_context_with_model( if (!hparams.vocab_only) { // initialize backends + int main_gpu = model->main_gpu; + + // with registry + if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { + if (main_gpu >= 0 && main_gpu < (int)model->devices.size()) { + ggml_backend_dev_t main_dev = model->devices[main_gpu]; + ggml_backend_t backend = ggml_backend_dev_init(main_dev, nullptr); + if (backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(main_dev)); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } + } else { + // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU + for (auto * dev : model->devices) { + ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr); + if (backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev)); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } + } + if (main_gpu >= (int)model->devices.size()) { + main_gpu -= (int)model->devices.size(); + } + #if defined(GGML_USE_RPC) if (model->n_gpu_layers > 0) { for (const auto & endpoint : model->rpc_servers) { @@ -19340,48 +19319,19 @@ struct llama_context * llama_new_context_with_model( ctx->backends.push_back(backend); } } + if (main_gpu >= (int)model->rpc_servers.size()) { + main_gpu -= (int)model->rpc_servers.size(); + } #endif -#if defined(GGML_USE_METAL) - if (model->n_gpu_layers > 0) { - ctx->backend_metal = ggml_backend_metal_init(); - if (ctx->backend_metal == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize Metal backend\n", __func__); - llama_free(ctx); - return nullptr; - } - ctx->backends.push_back(ctx->backend_metal); - } -#elif defined(GGML_USE_CUDA) - if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { - // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used - ggml_backend_t backend = ggml_backend_cuda_init(model->main_gpu); - if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize CUDA%d backend\n", __func__, model->main_gpu); - llama_free(ctx); - return nullptr; - } - ctx->backends.push_back(backend); - } else { - // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU - for (int device = 0; device < ggml_backend_cuda_get_device_count(); ++device) { - ggml_backend_t backend = ggml_backend_cuda_init(device); - if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize CUDA%d backend\n", __func__, device); - llama_free(ctx); - return nullptr; - } - ctx->backends.push_back(backend); - } - } -#elif defined(GGML_USE_VULKAN) +#if defined(GGML_USE_VULKAN) if (model->split_mode == LLAMA_SPLIT_MODE_ROW) { LLAMA_LOG_ERROR("%s: Row split not supported. Failed to initialize Vulkan backend\n", __func__); llama_free(ctx); return nullptr; } if (model->split_mode == LLAMA_SPLIT_MODE_NONE) { - ggml_backend_t backend = ggml_backend_vk_init(model->main_gpu); + ggml_backend_t backend = ggml_backend_vk_init(main_gpu); if (backend == nullptr) { LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__); llama_free(ctx); @@ -19402,9 +19352,9 @@ struct llama_context * llama_new_context_with_model( #elif defined(GGML_USE_SYCL) // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { - ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu); + ggml_backend_t backend = ggml_backend_sycl_init(main_gpu); if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, model->main_gpu); + LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, main_gpu); llama_free(ctx); return nullptr; } @@ -19423,7 +19373,7 @@ struct llama_context * llama_new_context_with_model( } #elif defined(GGML_USE_KOMPUTE) if (model->n_gpu_layers > 0) { - auto * backend = ggml_backend_kompute_init(model->main_gpu); + auto * backend = ggml_backend_kompute_init(main_gpu); if (backend == nullptr) { LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__); llama_free(ctx); @@ -19432,29 +19382,29 @@ struct llama_context * llama_new_context_with_model( ctx->backends.push_back(backend); } #elif defined(GGML_USE_CANN) - // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used - // TODO: ggml_backend_cann is not support split tensor now, just leave code here. - if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { - ggml_backend_t backend = ggml_backend_cann_init(model->main_gpu); - if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, model->main_gpu); - llama_free(ctx); - return nullptr; - } - ctx->backends.push_back(backend); - } else { - // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU - // TODO: currently, CANN can't use multi-gpus, just leave code here for further cann version. - for (int32_t device = 0; device < ggml_backend_cann_get_device_count(); ++device) { - ggml_backend_t backend = ggml_backend_cann_init(device); + // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used + // TODO: ggml_backend_cann is not support split tensor now, just leave code here. + if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { + ggml_backend_t backend = ggml_backend_cann_init(main_gpu); if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, device); + LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, main_gpu); llama_free(ctx); return nullptr; } ctx->backends.push_back(backend); + } else { + // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU + // TODO: currently, CANN can't use multi-gpus, just leave code here for further cann version. + for (int32_t device = 0; device < ggml_backend_cann_get_device_count(); ++device) { + ggml_backend_t backend = ggml_backend_cann_init(device); + if (backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, device); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } } - } #endif #ifdef GGML_USE_BLAS @@ -19519,7 +19469,7 @@ struct llama_context * llama_new_context_with_model( for (auto * backend : ctx->backends) { if (ggml_backend_is_cpu(backend)) { // use host buffers for the CPU backend compute buffer - backend_buft.push_back(llama_default_buffer_type_cpu(true)); + backend_buft.push_back(llama_default_buffer_type_cpu(*model, true)); } else { backend_buft.push_back(ggml_backend_get_default_buffer_type(backend)); } @@ -19530,17 +19480,37 @@ struct llama_context * llama_new_context_with_model( // buffer used to store the computation graph and the tensor meta data ctx->buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false)); + // TODO: move these checks to ggml_backend_sched // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary bool pipeline_parallel = llama_get_device_count(*model) > 1 && model->n_gpu_layers > (int)model->hparams.n_layer && model->split_mode == LLAMA_SPLIT_MODE_LAYER && params.offload_kqv; -#ifndef GGML_USE_CUDA - // pipeline parallelism requires support for async compute and events - // currently this is only implemented in the CUDA backend - pipeline_parallel = false; -#endif + + // pipeline parallelism requires support for async compute and events in all devices + if (pipeline_parallel) { + for (auto * backend : ctx->backends) { + if (ggml_backend_is_cpu(backend)) { + // ignore CPU backend + continue; + } + auto * dev = ggml_backend_get_device(backend); + if (!dev) { + // backend is using old interface, not supported + pipeline_parallel = false; + break; + } + ggml_backend_dev_props props; + ggml_backend_dev_get_props(dev, &props); + if (!props.caps.async || !props.caps.events) { + // device does not support async compute or events + pipeline_parallel = false; + break; + } + } + } + ctx->sched = ggml_backend_sched_new(ctx->backends.data(), backend_buft.data(), ctx->backends.size(), max_nodes, pipeline_parallel); if (pipeline_parallel) { @@ -21845,15 +21815,9 @@ const std::vector> & llama_internal } void llama_log_set(ggml_log_callback log_callback, void * user_data) { - g_state.log_callback = log_callback ? log_callback : llama_log_callback_default; - g_state.log_callback_user_data = user_data; -#ifdef GGML_USE_METAL - ggml_backend_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data); -#elif defined(GGML_USE_CUDA) - ggml_backend_cuda_log_set_callback(g_state.log_callback, g_state.log_callback_user_data); -#elif defined(GGML_USE_CANN) - ggml_backend_cann_log_set_callback(g_state.log_callback, g_state.log_callback_user_data); -#endif + ggml_log_set(log_callback, user_data); + g_logger_state.log_callback = log_callback ? log_callback : llama_log_callback_default; + g_logger_state.log_callback_user_data = user_data; } static void llama_log_internal_v(ggml_log_level level, const char * format, va_list args) { @@ -21862,12 +21826,12 @@ static void llama_log_internal_v(ggml_log_level level, const char * format, va_l char buffer[128]; int len = vsnprintf(buffer, 128, format, args); if (len < 128) { - g_state.log_callback(level, buffer, g_state.log_callback_user_data); + g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data); } else { char * buffer2 = new char[len + 1]; vsnprintf(buffer2, len + 1, format, args_copy); buffer2[len] = 0; - g_state.log_callback(level, buffer2, g_state.log_callback_user_data); + g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data); delete[] buffer2; } va_end(args_copy); diff --git a/src/unicode-data.cpp b/src/unicode-data.cpp index 02bdf7823..07424bbab 100644 --- a/src/unicode-data.cpp +++ b/src/unicode-data.cpp @@ -7,7 +7,7 @@ #include #include -const std::vector> unicode_ranges_flags = { // start, flags // last=next_start-1 +const std::initializer_list> unicode_ranges_flags = { // start, flags // last=next_start-1 {0x000000, 0x0080}, {0x000020, 0x0008}, {0x000021, 0x0020}, @@ -2311,7 +2311,8 @@ const std::unordered_set unicode_set_whitespace = { 0x003000, }; -const std::unordered_map unicode_map_lowercase = { +// list is always in ascending order, to enable binary searh +const std::initializer_list> unicode_map_lowercase = { {0x000041, 0x000061}, {0x000042, 0x000062}, {0x000043, 0x000063}, @@ -3747,7 +3748,8 @@ const std::unordered_map unicode_map_lowercase = { {0x01E921, 0x01E943}, }; -const std::unordered_map unicode_map_uppercase = { +// list is always in ascending order, to enable binary searh +const std::initializer_list> unicode_map_uppercase = { {0x000061, 0x000041}, {0x000062, 0x000042}, {0x000063, 0x000043}, @@ -5200,7 +5202,7 @@ const std::unordered_map unicode_map_uppercase = { {0x01E943, 0x01E921}, }; -const std::vector unicode_ranges_nfd = { // start, last, nfd +const std::initializer_list unicode_ranges_nfd = { // start, last, nfd {0x000000, 0x000000, 0x000000}, {0x0000C0, 0x0000C5, 0x000041}, {0x0000C7, 0x0000C7, 0x000043}, diff --git a/src/unicode-data.h b/src/unicode-data.h index e27fe1770..f6973ebd2 100644 --- a/src/unicode-data.h +++ b/src/unicode-data.h @@ -13,8 +13,8 @@ struct range_nfd { static const uint32_t MAX_CODEPOINTS = 0x110000; -extern const std::vector> unicode_ranges_flags; +extern const std::initializer_list> unicode_ranges_flags; extern const std::unordered_set unicode_set_whitespace; -extern const std::unordered_map unicode_map_lowercase; -extern const std::unordered_map unicode_map_uppercase; -extern const std::vector unicode_ranges_nfd; +extern const std::initializer_list> unicode_map_lowercase; +extern const std::initializer_list> unicode_map_uppercase; +extern const std::initializer_list unicode_ranges_nfd; diff --git a/src/unicode.cpp b/src/unicode.cpp index f4e941cd1..50b35bbbc 100644 --- a/src/unicode.cpp +++ b/src/unicode.cpp @@ -123,11 +123,11 @@ uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) { static std::vector unicode_cpt_flags_array() { std::vector cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED); - assert (unicode_ranges_flags.front().first == 0); - assert (unicode_ranges_flags.back().first == MAX_CODEPOINTS); + assert (unicode_ranges_flags.begin()[0].first == 0); + assert (unicode_ranges_flags.begin()[unicode_ranges_flags.size()-1].first == MAX_CODEPOINTS); for (size_t i = 1; i < unicode_ranges_flags.size(); ++i) { - const auto range_ini = unicode_ranges_flags[i-1]; // codepoint_ini, flags - const auto range_end = unicode_ranges_flags[i]; // codepoint_end, flags + const auto range_ini = unicode_ranges_flags.begin()[i-1]; // codepoint_ini, flags + const auto range_end = unicode_ranges_flags.begin()[i]; // codepoint_end, flags for (uint32_t cpt = range_ini.first; cpt < range_end.first; ++cpt) { cpt_flags[cpt] = range_ini.second; } @@ -597,7 +597,7 @@ std::vector unicode_cpts_normalize_nfd(const std::vector & c std::vector result(cpts.size()); for (size_t i = 0; i < cpts.size(); ++i) { const uint32_t cpt = cpts[i]; - auto it = std::upper_bound(unicode_ranges_nfd.cbegin(), unicode_ranges_nfd.cend(), cpt, comp) - 1; + auto it = std::upper_bound(unicode_ranges_nfd.begin(), unicode_ranges_nfd.end(), cpt, comp) - 1; result[i] = (it->first <= cpt && cpt <= it->last) ? it->nfd : cpt; } return result; @@ -639,8 +639,15 @@ uint8_t unicode_utf8_to_byte(const std::string & utf8) { } uint32_t unicode_tolower(uint32_t cp) { - auto it = unicode_map_lowercase.find(cp); - return it == unicode_map_lowercase.end() ? cp : it->second; + // binary search + auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cp, + [](const std::pair & pair, uint32_t value) { + return pair.first < value; + }); + if (it != unicode_map_lowercase.end() && it->first == cp) { + return it->second; + } + return cp; // Return the original code point if no lowercase mapping is found } std::vector unicode_regex_split(const std::string & text, const std::vector & regex_exprs) { diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index d2cfe06b5..a10d98e35 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -1,6 +1,6 @@ // This file defines tests for various GGML ops and backends. // For the forward pass it asserts that the results of multiple backends computing the same GGML ops are consistent. -// For the backwards pass it asserts that the gradients from backpropagation are consistent +// For the backward pass it asserts that the gradients from backpropagation are consistent // with the gradients obtained via the method of finite differences ("grad" mode, this is optional). // It is also possible to check the performance ("perf" mode). // @@ -116,6 +116,11 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m } else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) { // This is going to create some weird integers though. ggml_backend_tensor_set(tensor, data.data(), 0, ggml_nbytes(tensor)); + } else if (tensor->type == GGML_TYPE_I64) { + // Integers with a size of 8 bytes can be set by mirroring the float data, the specific values are again not really meaningful. + const size_t nbytes_half = ggml_nbytes(tensor)/2; + ggml_backend_tensor_set(tensor, data.data(), 0*nbytes_half, nbytes_half); + ggml_backend_tensor_set(tensor, data.data(), 1*nbytes_half, nbytes_half); } else { GGML_ABORT("fatal error"); } @@ -145,6 +150,8 @@ static std::vector tensor_to_float(const ggml_tensor * t) { tv.push_back(ggml_bf16_to_fp32(*(ggml_bf16_t*)&buf[i])); } else if (t->type == GGML_TYPE_F32) { tv.push_back(*(float *) &buf[i]); + } else if (t->type == GGML_TYPE_I64) { + tv.push_back((float)*(int64_t *) &buf[i]); } else if (t->type == GGML_TYPE_I32) { tv.push_back((float)*(int32_t *) &buf[i]); } else if (t->type == GGML_TYPE_I16) { @@ -672,14 +679,11 @@ struct test_case { } // run - ggml_backend_synchronize(backend); - int64_t total_time_us = 0; int total_runs = 0; do { int64_t start_time = ggml_time_us(); ggml_backend_graph_compute(backend, gf); - ggml_backend_synchronize(backend); int64_t end_time = ggml_time_us(); total_time_us += end_time - start_time; @@ -740,7 +744,7 @@ struct test_case { ggml_tensor * out = build_graph(ctx); - if (op_name != nullptr && op_desc(out) != op_name) { + if ((op_name != nullptr && op_desc(out) != op_name) || out->op == GGML_OP_OPT_STEP_ADAMW) { //printf(" %s: skipping\n", op_desc(out).c_str()); ggml_free(ctx); return true; @@ -749,11 +753,6 @@ struct test_case { printf(" %s(%s): ", op_desc(out).c_str(), vars().c_str()); fflush(stdout); - if (out->grad == nullptr) { - printf("backwards pass not supported \n"); - ggml_free(ctx); - return true; - } if (out->type != GGML_TYPE_F32) { ggml_free(ctx); printf("not supported [%s->type != FP32]\n", out->name); @@ -762,18 +761,26 @@ struct test_case { // check if the backend supports the ops bool supported = true; + bool any_params = false; for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { if (!ggml_backend_supports_op(backend, t)) { printf("not supported [%s] ", ggml_backend_name(backend)); supported = false; break; } - if ((t->flags & GGML_TENSOR_FLAG_PARAM) && t->type != GGML_TYPE_F32) { - printf("not supported [%s->type != FP32] ", t->name); - supported = false; - break; + if ((t->flags & GGML_TENSOR_FLAG_PARAM)) { + any_params = true; + if (t->type != GGML_TYPE_F32) { + printf("not supported [%s->type != FP32] ", t->name); + supported = false; + break; + } } } + if (!any_params) { + printf("not supported [%s] \n", op_name); + supported = false; + } if (!supported) { printf("\n"); ggml_free(ctx); @@ -801,7 +808,7 @@ struct test_case { ggml_build_forward_expand(gf, out); ggml_graph_cpy(gf, gb); - ggml_build_backward_expand(ctx, gf, gb, false, false); + ggml_build_backward_expand(ctx, gf, gb, false); if (expect.size() != 1 || expect[0] != 0.0f) { GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf)); for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { @@ -984,7 +991,7 @@ struct test_example : public test_case { } // In order to also check the gradients for your op, add calls like ggml_set_param(ctx, a) // immediately after you create the tensors. - // This is optional and only makes sense if a backwards pass has actually been implemented for the new op. + // This is optional and only makes sense if a backward pass has actually been implemented for the new op. }; @@ -1116,6 +1123,71 @@ struct test_get_rows : public test_case { } }; +// GGML_OP_ARGMAX +struct test_argmax : public test_case { + const ggml_type type; + const std::array ne; + + std::string vars() override { + return VARS_TO_STR2(type, ne); + } + + test_argmax(ggml_type type = GGML_TYPE_F32, + std::array ne = {10, 100, 1, 1}) + : type(type), ne(ne) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); + + ggml_tensor * out = ggml_argmax(ctx, a); + ggml_set_name(out, "out"); + + return out; + } + + double max_nmse_err() override { + return 0.0; + } +}; + +// GGML_OP_COUNT_EQUAL +struct test_count_equal : public test_case { + const ggml_type type; + const std::array ne; + + std::string vars() override { + return VARS_TO_STR2(type, ne); + } + + test_count_equal(ggml_type type = GGML_TYPE_F32, + std::array ne = {4, 500, 1, 1}) + : type(type), ne(ne) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); + + ggml_tensor * a_argmax = ggml_argmax(ctx, a); + ggml_set_name(a_argmax, "a_argmax"); + + ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(b, "b"); + + ggml_tensor * b_argmax = ggml_argmax(ctx, a); + ggml_set_name(b_argmax, "b_argmax"); + + ggml_tensor * out = ggml_count_equal(ctx, a_argmax, b_argmax); + ggml_set_name(out, "out"); + + return out; + } + + double max_nmse_err() override { + return 0.0; + } +}; + // GGML_OP_REPEAT struct test_repeat : public test_case { const ggml_type type; @@ -1223,7 +1295,7 @@ struct test_set : public test_case { offset += ((ne_dst[i] - ne[i])/2)*dst->nb[i]; } ggml_tensor * out = ggml_set(ctx, dst, src, - // The backwards pass requires setting a contiguous region: + // The backward pass requires setting a contiguous region: src->nb[1], src->nb[2], src->nb[3], offset); ggml_set_name(out, "out"); @@ -1335,7 +1407,7 @@ struct test_bin_bcast : public test_case { ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data()); ggml_set_name(b, "b"); - // The backwards pass supports broadcasting only for GGML_ADD: + // The backward pass supports broadcasting only for GGML_ADD: const bool grad_supported = op == ggml_add || ggml_are_same_shape(a, b); if (grad_supported) { ggml_set_param(ctx, a); @@ -1830,7 +1902,7 @@ struct test_log : public test_case { void initialize_tensors(ggml_context * ctx) override { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { - // log(1) == 0, cluster values there to keep the sum low for better precision in the backwards pass: + // log(1) == 0, cluster values there to keep the sum low for better precision in the backward pass: init_tensor_uniform(t, 0.9f, 1.1f); } } @@ -2748,7 +2820,10 @@ struct test_opt_step_adamw : public test_case { ggml_set_param(ctx, a); // Despite tensor a having gradients the output tensor will not. ggml_set_name(a, "a"); - ggml_tensor * out = ggml_opt_step_adamw(ctx, a, alpha, beta1, beta2, eps, wd); + ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]); + ggml_set_name(grad, "grad"); + + ggml_tensor * out = ggml_opt_step_adamw(ctx, a, grad, alpha, beta1, beta2, eps, wd); ggml_set_name(out, "out"); return out; @@ -3257,7 +3332,10 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1)); test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1)); - for (int ne3 : {1, 3}) { // CUDA backwards pass only supports ne3 == 1 + test_cases.emplace_back(new test_argmax()); + test_cases.emplace_back(new test_count_equal()); + + for (int ne3 : {1, 3}) { // CUDA backward pass only supports ne3 == 1 test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, ne3}, {1, 1, 1, 1})); test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, ne3}, {2, 1, 1, 1})); test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, ne3}, {1, 2, 1, 1})); @@ -3275,8 +3353,8 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_dup(GGML_TYPE_F16, {10, 10, 5, 1}, {0, 2, 1, 3})); // dup by rows test_cases.emplace_back(new test_dup(GGML_TYPE_F32, {10, 10, 5, 1}, {1, 0, 2, 3})); test_cases.emplace_back(new test_dup(GGML_TYPE_F16, {10, 10, 5, 1}, {1, 0, 2, 3})); // dup dst not-contiguous - test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {0, 2, 1, 3})); - test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {1, 2, 0, 3})); + test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {0, 2, 1, 3})); + test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {1, 2, 0, 3})); for (int dim = 1; dim < GGML_MAX_DIMS; ++dim) { test_cases.emplace_back(new test_set(GGML_TYPE_F32, GGML_TYPE_F32, {6, 5, 4, 3}, dim)); @@ -3717,20 +3795,22 @@ int main(int argc, char ** argv) { } // enumerate backends - printf("Testing %zu backends\n\n", ggml_backend_reg_get_count()); + printf("Testing %zu devices\n\n", ggml_backend_dev_count()); size_t n_ok = 0; - for (size_t i = 0; i < ggml_backend_reg_get_count(); i++) { - printf("Backend %zu/%zu (%s)\n", i + 1, ggml_backend_reg_get_count(), ggml_backend_reg_get_name(i)); + for (size_t i = 0; i < ggml_backend_dev_count(); i++) { + ggml_backend_dev_t dev = ggml_backend_dev_get(i); - if (backend_filter != NULL && strcmp(backend_filter, ggml_backend_reg_get_name(i)) != 0) { + printf("Backend %zu/%zu: %s\n", i + 1, ggml_backend_dev_count(), ggml_backend_dev_name(dev)); + + if (backend_filter != NULL && strcmp(backend_filter, ggml_backend_dev_name(dev)) != 0) { printf(" Skipping\n"); n_ok++; continue; } - ggml_backend_t backend = ggml_backend_reg_init_backend(i, NULL); + ggml_backend_t backend = ggml_backend_dev_init(dev, NULL); GGML_ASSERT(backend != NULL); if (backend_filter == NULL && ggml_backend_is_cpu(backend) && mode != MODE_GRAD) { @@ -3745,7 +3825,11 @@ int main(int argc, char ** argv) { ggml_backend_cpu_set_n_threads(backend, std::thread::hardware_concurrency() / 2); } - printf(" Backend name: %s\n", ggml_backend_name(backend)); + printf(" Device description: %s\n", ggml_backend_dev_description(dev)); + size_t free, total; // NOLINT + ggml_backend_dev_memory(dev, &free, &total); + printf(" Device memory: %zu MB (%zu MB free)\n", total / 1024 / 1024, free / 1024 / 1024); + printf("\n"); bool ok = test_backend(backend, mode, op_name_filter); @@ -3762,9 +3846,9 @@ int main(int argc, char ** argv) { ggml_backend_free(backend); } - printf("%zu/%zu backends passed\n", n_ok, ggml_backend_reg_get_count()); + printf("%zu/%zu backends passed\n", n_ok, ggml_backend_dev_count()); - if (n_ok != ggml_backend_reg_get_count()) { + if (n_ok != ggml_backend_dev_count()) { printf("\033[1;31mFAIL\033[0m\n"); return 1; } diff --git a/tests/test-grad0.cpp b/tests/test-grad0.cpp index 2ef606d2c..2200ad93d 100644 --- a/tests/test-grad0.cpp +++ b/tests/test-grad0.cpp @@ -240,12 +240,14 @@ static bool check_gradient( struct ggml_cgraph * gb = ggml_new_graph_custom(ctx0, GGML_DEFAULT_GRAPH_SIZE, true); ggml_build_forward_expand(gf, f); ggml_graph_cpy(gf, gb); - ggml_build_backward_expand(ctx0, gf, gb, false, false); + ggml_build_backward_expand(ctx0, gf, gb, false); ggml_graph_compute_with_ctx(ctx0, gf, n_threads); - ggml_graph_reset (gf); - ggml_set_f32 (f->grad, 1.0f); + ggml_graph_reset(gb); + if (f->grad) { + ggml_set_f32(f->grad, 1.0f); + } ggml_graph_compute_with_ctx(ctx0, gb, n_threads); @@ -298,8 +300,10 @@ static bool check_gradient( ggml_set_f32_1d(x[i], k, x0); // compute gradient using backward graph - ggml_graph_reset (gf); - ggml_set_f32 (f->grad, 1.0f); + ggml_graph_reset(gb); + if (f->grad) { + ggml_set_f32(f->grad, 1.0f); + } ggml_graph_compute_with_ctx(ctx0, gb, n_threads);