From d9653894dffbfd3a58616f31b0967b34faf6f611 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 9 Jan 2024 16:23:05 +0200 Subject: [PATCH 01/42] scripts : script to get Paul Graham essays in txt format (#4838) --- scripts/get-pg.sh | 47 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 47 insertions(+) create mode 100755 scripts/get-pg.sh diff --git a/scripts/get-pg.sh b/scripts/get-pg.sh new file mode 100755 index 000000000..d516db46c --- /dev/null +++ b/scripts/get-pg.sh @@ -0,0 +1,47 @@ +#!/bin/bash + +function usage { + echo "usage: $0" + exit 1 +} + +function has_cmd { + if ! [ -x "$(command -v $1)" ]; then + echo "error: $1 is not available" >&2 + exit 1 + fi +} + +# check for: curl, html2text, tail, sed, fmt +has_cmd curl +has_cmd html2text +has_cmd tail +has_cmd sed + +if [ $# -ne 1 ]; then + usage +fi + +n=$1 + +# get urls +urls="$(curl http://www.aaronsw.com/2002/feeds/pgessays.rss | grep html | sed -e "s/.*http/http/" | sed -e "s/html.*/html/" | head -n $n)" + +printf "urls:\n%s\n" "$urls" + +if [ -f pg.txt ]; then + rm pg.txt +fi + +for url in $urls; do + echo "processing $url" + + curl -L $url | html2text | tail -n +4 | sed -E "s/^[[:space:]]+//g" | fmt -w 80 >> pg.txt + + # don't flood the server + sleep 1 +done + +echo "done. data in pg.txt" + +exit 0 From 18adb4e9bb340b7b4565d8b6715b4449283e7641 Mon Sep 17 00:00:00 2001 From: iohub Date: Wed, 10 Jan 2024 00:45:54 +0800 Subject: [PATCH 02/42] readme : add 3rd party collama reference to UI list (#4840) Add a VSCode extension for llama.cpp reference to UI list --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index a0d86a6ef..866aa87b4 100644 --- a/README.md +++ b/README.md @@ -137,6 +137,7 @@ as the main playground for developing new features for the [ggml](https://github - [semperai/amica](https://github.com/semperai/amica) - [psugihara/FreeChat](https://github.com/psugihara/FreeChat) - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal) +- [iohub/collama](https://github.com/iohub/coLLaMA) --- From 9a818f7c42761984ac99e08e613cc20634f8410e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 9 Jan 2024 19:20:45 +0200 Subject: [PATCH 03/42] scripts : improve get-pg.sh (#4838) --- scripts/get-pg.sh | 25 ++++++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-) diff --git a/scripts/get-pg.sh b/scripts/get-pg.sh index d516db46c..b027793e1 100755 --- a/scripts/get-pg.sh +++ b/scripts/get-pg.sh @@ -2,6 +2,22 @@ function usage { echo "usage: $0" + echo "note: n is the number of essays to download" + echo "for specific n, the resulting pg.txt file will have the following number of tokens:" + echo "n | tokens" + echo "--- | ---" + echo "1 | 6230" + echo "2 | 23619" + echo "5 | 25859" + echo "10 | 36888" + echo "15 | 50188" + echo "20 | 59094" + echo "25 | 88764" + echo "30 | 103121" + echo "32 | 108338" + echo "35 | 113403" + echo "40 | 127699" + echo "45 | 135896" exit 1 } @@ -33,10 +49,17 @@ if [ -f pg.txt ]; then rm pg.txt fi +c=1 for url in $urls; do echo "processing $url" - curl -L $url | html2text | tail -n +4 | sed -E "s/^[[:space:]]+//g" | fmt -w 80 >> pg.txt + cc=$(printf "%03d" $c) + + curl -L $url | html2text | tail -n +4 | sed -E "s/^[[:space:]]+//g" | fmt -w 80 >> pg-$cc-one.txt + cat pg-$cc-one.txt >> pg.txt + + cp -v pg.txt pg-$cc-all.txt + c=$((c+1)) # don't flood the server sleep 1 From 4dccb38d9abab7f9f2d1f9a6977df4185d490132 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 9 Jan 2024 19:37:08 +0200 Subject: [PATCH 04/42] metal : improve dequantize precision to match CPU (#4836) ggml-ci --- ggml-metal.metal | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/ggml-metal.metal b/ggml-metal.metal index 0cc535ac7..229efb8b6 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -3841,8 +3841,8 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4]; int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2) : (scale_2&kmask2) | ((scale_1&kmask1) << 4); - half dl = il<8 ? d_all * (dl_int - 32.h) : d_all * (dl_int / 16.h - 32.h); - const half ml = 4.h * dl; + float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f); + const float ml = 4.f * dl; il = (il/2) & 3; const half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h); @@ -3909,7 +3909,7 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg uint8_t ul = 1 << (il/2); il = il & 3; const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales); - const float d = il < 2 ? xb->d : xb->d / 16.h; + const float d = il < 2 ? xb->d : xb->d / 16.f; const float min = xb->dmin; const float dl = d * sc[0]; const float ml = min * sc[1]; @@ -3942,17 +3942,17 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg #if QK_K == 256 ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1); qh = qh + 32*(il/8) + 16*(il&1); - half sc = scales[(il%2) + 2 * ((il/2))]; + float sc = scales[(il%2) + 2 * ((il/2))]; il = (il/2) & 3; #else ql = ql + 16 * (il&1); - half sc = scales[il]; + float sc = scales[il]; #endif const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F; - const half coef = il>1 ? 1.f/16.h : 1.h; - const half ml = d_all * sc * 32.h; - const half dl = d_all * sc * coef; + const float coef = il>1 ? 1.f/16.f : 1.f; + const float ml = d_all * sc * 32.f; + const float dl = d_all * sc * coef; for (int i = 0; i < 16; ++i) { const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2)) : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4)); From 36e5a08b203542dca53cca4eaf172c5dc4bbc991 Mon Sep 17 00:00:00 2001 From: Justine Tunney Date: Tue, 9 Jan 2024 09:59:14 -0800 Subject: [PATCH 05/42] llava-cli : don't crash if --image flag is invalid (#4835) This change fixes an issue where supplying `--image missing-file` would result in a segfault due to a null pointer being dereferenced. This can result in distracting info being printed if robust crash analysis tools are being used. --- examples/llava/llava-cli.cpp | 3 +++ 1 file changed, 3 insertions(+) diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index 502b788b1..d94795fe3 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -243,6 +243,9 @@ int main(int argc, char ** argv) { } auto image_embed = load_image(ctx_llava, ¶ms); + if (!image_embed) { + return 1; + } // process the prompt process_prompt(ctx_llava, image_embed, ¶ms, params.prompt); From 6efb8eb30e7025b168f3fda3ff83b9b386428ad6 Mon Sep 17 00:00:00 2001 From: Austin <77757836+teleprint-me@users.noreply.github.com> Date: Tue, 9 Jan 2024 13:46:46 -0500 Subject: [PATCH 06/42] convert.py : fix vanilla LLaMA model conversion (#4818) * Update Imports and Add Notes for Future Reference - Updated import statements in `convert.py`. - Added import for `AutoTokenizer` from `transformers` module. - Added conditional import for `gguf` from the local directory. - Added comments and notes for future reference. Additional Notes: - Noted removal of a redundant `TypeAlias` import. - Noted the removal of a `gguf` debug statement. - Commented on the presence of `ARCH` and `NDArray` definitions. - Commented on cleaning up and refactoring data type definitions. * Refine Model Hyperparameters and Params Class - Updated type annotations to use `Optional` for clarity. - Improved method names and attribute consistency. - Removed unnecessary variables for better code readability. Additional Notes: - Highlighted the use of `Optional` for clearer intent. - Ensured backward and forward compatibility. * Restore BpeVocab and SentencePieceVocab classes - Restored the BpeVocab class for handling BPE tokenization. - Restored the SentencePieceVocab class for SentencePiece tokenization. These classes are essential for maintaining the original behavior of the codebase. * refactor: Standardize vocabulary handling with HfVocab - Replaced VocabLoader with HfVocab, aligning vocabulary handling across classes. - Updated initialization of HfVocab with local_files_only=True for AutoTokenizer. - Introduced optional parameter fname_added_tokens for flexible added token management. - Streamlined added token handling for clarity and conciseness. - Maintained special tokens and IDs, enhancing token management. - Simplified token processing methods for improved readability. - Added a placeholder for score computation with a default value of -1000.0. - Optimized newline token check for efficiency. - Updated __repr__ function for clarity in representation. - Adjusted type alias Vocab to include BpeVocab, SentencePieceVocab, and HfVocab. - Removed redundant code related to special token handling, reverse vocabulary mapping, and vocabulary file detection. This refactoring promotes a standardized and modular approach to vocabulary management, facilitating future integration with a VocabFactory and improving code maintainability and scalability. * refactor: Enhance readability, functionality, and code quality - Improved code formatting and readability for better maintainability. - Refactored LazyUnpickler's CLASSES dictionary for clarity. - Added print statements and warnings in check_vocab_size for user feedback. - Removed find_vocab_file_path, as it's superseded by VocabFactory. - Preparatory changes for upcoming classes: OutputFile and VocabFactory. - Overall focus on code quality, error handling, and consistency. These changes reflect a continuous effort to refine the codebase, ensuring it meets best practices and prepares for future enhancements, such as the VocabFactory. * refactor: Update OutputFile class for enhanced model vocabulary management - Restructured the constructor for improved readability. - Updated `add_meta_arch` method for flexible model name determination. - Introduced `handle_tokenizer_model` for mapping vocab types to supported tokenizer models. - Streamlined vocabulary extraction with `extract_vocabulary_from_model`. - Simplified vocabulary metadata addition using `add_meta_vocab`. - Refactored `add_tensor_info` for clarity and consistency. - Improved error handling for better user feedback. These changes signify the development of a versatile and comprehensive `OutputFile` class, enabling efficient management of model conversion output, metadata, vocabulary, and tensor information. * feat: Introduce VocabFactory for flexible vocabulary management in model conversion - The VocabFactory class is added to facilitate modular vocabulary handling. - The constructor initializes a directory path and detects vocabulary-related files. - The _select_file method provides file paths based on vocabulary type (e.g., BPE, SentencePiece). - _create_special_vocab generates special vocabularies, accommodating different types. - The load_vocab method loads vocabularies, handling BPE, SentencePiece, and Hugging Face Fast Tokenizer. - Error handling and logging enhance debugging and user feedback. - The modular and flexible design simplifies vocabulary management and supports future extensions. The VocabFactory class enhances code modularity and maintainability, allowing versatile vocabulary handling in the model conversion process. * refactor: Improve code organization, argument parsing, and user interface - Renamed 'default_outfile' to 'default_output_file' for clarity. - Refactored argument parser setup into 'get_argument_parser' function. - Introduced descriptive comments for each argument in the parser. - Added '--vocab-type' argument with choices ["spm", "bpe", "hfft"] for vocabulary processing. - Improved flag naming consistency: '--outfile' to '--out-file' and '--bigendian' to '--big-endian'. - Enhanced error handling to prevent overwriting input data in 'default_output_file'. - Made 'argv' in 'main' an optional parameter for flexibility. - Introduced dynamic import for 'awq.apply_awq' based on 'args.awq_path' for conditional dependency. These changes enhance code clarity, organization, and the user interface of the script, aligning it with Python best practices and improving maintainability. * refactor: Further refine functionality, improve user interaction, and streamline vocabulary handling - Renamed command-line arguments for clarity and consistency. - Improved path resolution and import adjustments for robustness. - Thoughtfully handled 'awq-path' and conditional logic for the weighted model. - Enhanced model and vocabulary loading with the 'VocabFactory' class for structured and adaptable loading. - Strengthened error handling and user feedback for a more user-friendly experience. - Structured output file handling with clear conditions and defaults. - Streamlined and organized the 'main' function for better logic flow. - Passed 'sys.argv[1:]' to 'main' for adaptability and testability. These changes solidify the script's functionality, making it more robust, user-friendly, and adaptable. The use of the 'VocabFactory' class is a notable enhancement in efficient vocabulary handling, reflecting a thoughtful and iterative approach to script development. * chore: Apply ruff formatting to convert.py Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> * Revert to commit 0614c33 * chore: Apply flake8 formatting rules Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> * refactor: Revise `check_vocab_size` for Enhanced Clarity and Correctness - Resolved an unreachable branch issue by reorganizing the conditional structure. - Moved the special case check for `params.n_vocab == -1` to the top for immediate assertion. - Flattened the conditional logic for improved clarity and predictability of the function's behavior. These changes enhance the readability and functional correctness of the `check_vocab_size` function without altering its intended functionality. * py : fix outfile and outtype * py : suggest hint for missing vocab size --------- Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com> Co-authored-by: Georgi Gerganov --- convert.py | 969 ++++++++++++++++++++++++++++++++++++----------------- 1 file changed, 666 insertions(+), 303 deletions(-) diff --git a/convert.py b/convert.py index c3f3fc0a1..3b613eefc 100755 --- a/convert.py +++ b/convert.py @@ -17,29 +17,58 @@ import signal import struct import sys import time +import warnings import zipfile from abc import ABCMeta, abstractmethod -from collections import OrderedDict +from argparse import ArgumentParser from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor from dataclasses import dataclass from pathlib import Path -from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, Optional, TypeVar, cast +from typing import ( + IO, + TYPE_CHECKING, + Any, + Callable, + Iterable, + Literal, + Optional, + Tuple, + TypeVar, +) import numpy as np from sentencepiece import SentencePieceProcessor -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) -import gguf +try: + from transformers import AutoTokenizer +except ModuleNotFoundError as e: + warnings.warn(f"Could not import AutoTokenizer from transformers: {e}") -if TYPE_CHECKING: - from typing import TypeAlias +# If NO_LOCAL_GGUF is not set, try to import gguf from the local gguf-py directory +if "NO_LOCAL_GGUF" not in os.environ: + # Use absolute path to the gguf-py directory + gguf_py_dir = str(Path(__file__).resolve().parent / "gguf-py") + print(gguf_py_dir) # NOTE: Remove this once path is verified after changes are completed + if gguf_py_dir not in sys.path: + sys.path.insert(1, gguf_py_dir) -if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'): +# Import gguf module +try: + import gguf +except ModuleNotFoundError as e: + print(f"Could not import gguf: {e}") + sys.exit(1) + +if TYPE_CHECKING: # NOTE: This isn't necessary. + from typing import TypeAlias # This can technically be omitted. + +if hasattr(faulthandler, "register") and hasattr(signal, "SIGUSR1"): faulthandler.register(signal.SIGUSR1) -NDArray: TypeAlias = 'np.ndarray[Any, Any]' +# NOTE: n-dimensional arrays should be directly referenced +NDArray: TypeAlias = "np.ndarray[Any, Any]" +# Why is this here? LLAMA and GPT are technically the only compatible ARCHs. ARCH = gguf.MODEL_ARCH.LLAMA DEFAULT_CONCURRENCY = 8 @@ -49,6 +78,7 @@ DEFAULT_CONCURRENCY = 8 # +# TODO: Clean up and refactor data types @dataclass(frozen=True) class DataType: name: str @@ -153,65 +183,85 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = { @dataclass class Params: - n_vocab: int - n_embd: int - n_layer: int - n_ctx: int - n_ff: int - n_head: int - n_head_kv: int - n_experts: int | None = None - n_experts_used: int | None = None - f_norm_eps: float | None = None + n_vocab: int + n_embd: int + n_layer: int + n_ctx: int + n_ff: int + n_head: int + n_head_kv: int + f_norm_eps: Optional[float] = None + n_experts: Optional[int] = None + n_experts_used: Optional[int] = None - rope_scaling_type: gguf.RopeScalingType | None = None - f_rope_freq_base: float | None = None - f_rope_scale: float | None = None - n_orig_ctx: int | None = None - rope_finetuned: bool | None = None + rope_scaling_type: Optional[gguf.RopeScalingType] = None + f_rope_freq_base: Optional[float] = None + f_rope_scale: Optional[float] = None + n_orig_ctx: Optional[int] = None + rope_finetuned: Optional[bool] = None - ftype: GGMLFileType | None = None + ftype: Optional[GGMLFileType] = None # path to the directory containing the model files - path_model: Path | None = None + path_model: Optional[Path] = None @staticmethod - def guessed(model: LazyModel) -> Params: + def guessed(model: LazyModel) -> "Params": # try transformer naming first - n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape + n_vocab, n_embd = ( + model["model.embed_tokens.weight"].shape + if "model.embed_tokens.weight" in model + else model["tok_embeddings.weight"].shape + ) # try transformer naming first if "model.layers.0.self_attn.q_proj.weight" in model: - n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model) - elif "model.layers.0.self_attn.W_pack.weight" in model: # next: try baichuan naming - n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model) + n_layer = next( + i + for i in itertools.count() + if f"model.layers.{i}.self_attn.q_proj.weight" not in model + ) + elif ( + "model.layers.0.self_attn.W_pack.weight" in model + ): # next: try baichuan naming + n_layer = next( + i + for i in itertools.count() + if f"model.layers.{i}.self_attn.W_pack.weight" not in model + ) else: - n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model) + n_layer = next( + i + for i in itertools.count() + if f"layers.{i}.attention.wq.weight" not in model + ) if n_layer < 1: - raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n" - "Suggestion: provide 'config.json' of the model in the same directory containing model files.") + raise Exception( + "failed to guess 'n_layer'. This model is unknown or unsupported.\n" + "Suggestion: provide 'config.json' of the model in the same directory containing model files." + ) - n_head = n_embd // 128 # guessed - n_mult = 256 # guessed + n_head = n_embd // 128 # guessed + n_mult = 256 # guessed # TODO: verify this n_ff = int(2 * (4 * n_embd) / 3) n_ff = n_mult * ((n_ff + n_mult - 1) // n_mult) return Params( - n_vocab = n_vocab, - n_embd = n_embd, - n_layer = n_layer, - n_ctx = -1, - n_ff = n_ff, - n_head = n_head, - n_head_kv = n_head, - f_norm_eps = 1e-5, + n_vocab=n_vocab, + n_embd=n_embd, + n_layer=n_layer, + n_ctx=-1, + n_ff=n_ff, + n_head=n_head, + n_head_kv=n_head, + f_norm_eps=1e-5, ) @staticmethod - def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params: + def load_transformers_config(model: LazyModel, config_path: Path) -> "Params": config = json.load(open(config_path)) rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None @@ -224,20 +274,22 @@ class Params: rope_scaling_type = gguf.RopeScalingType.LINEAR elif typ == "yarn": rope_scaling_type = gguf.RopeScalingType.YARN - n_orig_ctx = rope_scaling['original_max_position_embeddings'] - rope_finetuned = rope_scaling['finetuned'] + n_orig_ctx = rope_scaling["original_max_position_embeddings"] + rope_finetuned = rope_scaling["finetuned"] else: - raise NotImplementedError(f'Unknown rope scaling type: {typ}') + raise NotImplementedError(f"Unknown rope scaling type: {typ}") if "max_sequence_length" in config: n_ctx = config["max_sequence_length"] elif "max_position_embeddings" in config: n_ctx = config["max_position_embeddings"] else: - raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n" - "Suggestion: provide 'config.json' of the model in the same directory containing model files.") + raise Exception( + "failed to guess 'n_ctx'. This model is unknown or unsupported.\n" + "Suggestion: provide 'config.json' of the model in the same directory containing model files." + ) - n_experts = None + n_experts = None n_experts_used = None if "num_local_experts" in config: @@ -245,30 +297,30 @@ class Params: n_experts_used = config["num_experts_per_tok"] return Params( - n_vocab = config["vocab_size"], - n_embd = config["hidden_size"], - n_layer = config["num_hidden_layers"], - n_ctx = n_ctx, - n_ff = config["intermediate_size"], - n_head = (n_head := config["num_attention_heads"]), - n_head_kv = config.get("num_key_value_heads", n_head), - n_experts = n_experts, - n_experts_used = n_experts_used, - f_norm_eps = config["rms_norm_eps"], - f_rope_freq_base = config.get("rope_theta"), - rope_scaling_type = rope_scaling_type, - f_rope_scale = f_rope_scale, - n_orig_ctx = n_orig_ctx, - rope_finetuned = rope_finetuned, + n_vocab=config["vocab_size"], + n_embd=config["hidden_size"], + n_layer=config["num_hidden_layers"], + n_ctx=n_ctx, + n_ff=config["intermediate_size"], + n_head=(n_head := config["num_attention_heads"]), + n_head_kv=config.get("num_key_value_heads", n_head), + n_experts=n_experts, + n_experts_used=n_experts_used, + f_norm_eps=config["rms_norm_eps"], + f_rope_freq_base=config.get("rope_theta"), + rope_scaling_type=rope_scaling_type, + f_rope_scale=f_rope_scale, + n_orig_ctx=n_orig_ctx, + rope_finetuned=rope_finetuned, ) # LLaMA v2 70B params.json # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1} @staticmethod - def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params: + def load_torch_params(model: LazyModel, config_path: Path) -> "Params": config = json.load(open(config_path)) - n_experts = None + n_experts = None n_experts_used = None f_rope_freq_base = None @@ -291,129 +343,249 @@ class Params: if config.get("moe"): n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0] - n_experts = config["moe"]["num_experts"] + n_experts = config["moe"]["num_experts"] n_experts_used = config["moe"]["num_experts_per_tok"] f_rope_freq_base = 1e6 return Params( - n_vocab = model["tok_embeddings.weight"].shape[0], - n_embd = config["dim"], - n_layer = config["n_layers"], - n_ctx = n_ctx, - n_ff = n_ff, - n_head = (n_head := config["n_heads"]), - n_head_kv = config.get("n_kv_heads", n_head), - n_experts = n_experts, - n_experts_used = n_experts_used, - f_norm_eps = config["norm_eps"], - f_rope_freq_base = config.get("rope_theta", f_rope_freq_base), + n_vocab=config.get("vocab_size", model["tok_embeddings.weight"].shape[0]), + n_embd=config["dim"], + n_layer=config["n_layers"], + n_ctx=n_ctx, + n_ff=n_ff, + n_head=(n_head := config["n_heads"]), + n_head_kv=config.get("n_kv_heads", n_head), + n_experts=n_experts, + n_experts_used=n_experts_used, + f_norm_eps=config["norm_eps"], + f_rope_freq_base=config.get("rope_theta", f_rope_freq_base), ) @staticmethod - def load(model_plus: ModelPlus) -> Params: - hf_config_path = model_plus.paths[0].parent / "config.json" + def load(model_plus: ModelPlus) -> "Params": + hf_config_path = model_plus.paths[0].parent / "config.json" orig_config_path = model_plus.paths[0].parent / "params.json" if hf_config_path.exists(): - params = Params.loadHFTransformerJson(model_plus.model, hf_config_path) + params = Params.load_transformers_config(model_plus.model, hf_config_path) elif orig_config_path.exists(): - params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path) - elif model_plus.format != 'none': + params = Params.load_torch_params(model_plus.model, orig_config_path) + elif model_plus.format != "none": params = Params.guessed(model_plus.model) else: - raise ValueError('Cannot guess params when model format is none') + raise ValueError("Cannot guess params when model format is none") params.path_model = model_plus.paths[0].parent return params -class VocabLoader: - def __init__(self, params: Params, fname_tokenizer: Path) -> None: - try: - from transformers import AutoTokenizer - except ImportError as e: - raise ImportError( - "To use VocabLoader, please install the `transformers` package. " - "You can install it with `pip install transformers`." - ) from e +class BpeVocab: # GPT + def __init__( + self, fname_tokenizer: Path, fname_added_tokens: Optional[Path] + ) -> None: + self.bpe_tokenizer = json.loads( + open(str(fname_tokenizer), encoding="utf-8").read() + ) + added_tokens: dict[str, int] + if fname_added_tokens is not None: + # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab. + added_tokens = json.load(open(fname_added_tokens, encoding="utf-8")) + else: + # Fall back to trying to find the added tokens in tokenizer.json + tokenizer_json_file = fname_tokenizer.parent / "tokenizer.json" + if not tokenizer_json_file.is_file(): + added_tokens = {} + else: + tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8")) + added_tokens = dict( + (item["content"], item["id"]) + for item in tokenizer_json.get("added_tokens", []) + # Added tokens here can be duplicates of the main vocabulary. + if item["content"] not in self.bpe_tokenizer + ) - try: - self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer), trust_remote_code=True) - except ValueError: - self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer), use_fast=False, trust_remote_code=True) + vocab_size: int = len(self.bpe_tokenizer) + expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) + actual_ids = sorted(added_tokens.values()) + if expected_ids != actual_ids: + expected_end_id = vocab_size + len(actual_ids) - 1 + raise Exception( + f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}" + ) - self.added_tokens_dict: OrderedDict[str, int] = OrderedDict() + items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) + self.added_tokens_list = [text for (text, idx) in items] + self.vocab_size_base: int = vocab_size + self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + self.fname_added_tokens = fname_added_tokens - for tok, tokidx in sorted(self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]): - if tokidx >= params.n_vocab or tokidx < self.tokenizer.vocab_size: - continue + def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + tokenizer = self.bpe_tokenizer + reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()} - self.added_tokens_dict[tok] = tokidx + for i, _ in enumerate(tokenizer): + yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL - self.unk_token_id: int = self.tokenizer.unk_token_id - self.specials: dict[str, int] = { + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score, gguf.TokenType.CONTROL + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.bpe_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class SentencePieceVocab: # LlaMa + def __init__( + self, fname_tokenizer: Path, fname_added_tokens: Optional[Path] + ) -> None: + self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer)) + added_tokens: dict[str, int] + if fname_added_tokens is not None: + added_tokens = json.load(open(fname_added_tokens, encoding="utf-8")) + else: + added_tokens = {} + + vocab_size: int = self.sentencepiece_tokenizer.vocab_size() + + new_tokens = { + id: piece for piece, id in added_tokens.items() if id >= vocab_size + } + expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) + actual_new_ids = sorted(new_tokens.keys()) + + if expected_new_ids != actual_new_ids: + raise ValueError( + f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}" + ) + + # Token pieces that were added to the base vocabulary. + self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + self.fname_added_tokens = fname_added_tokens + + def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + tokenizer = self.sentencepiece_tokenizer + for i in range(tokenizer.vocab_size()): + piece = tokenizer.id_to_piece(i) + text: bytes = piece.encode("utf-8") + score: float = tokenizer.get_score(i) + + toktype = gguf.TokenType.NORMAL + if tokenizer.is_unknown(i): + toktype = gguf.TokenType.UNKNOWN + if tokenizer.is_control(i): + toktype = gguf.TokenType.CONTROL + + # NOTE: I think added_tokens are user defined. + # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto + # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED + + if tokenizer.is_unused(i): + toktype = gguf.TokenType.UNUSED + if tokenizer.is_byte(i): + toktype = gguf.TokenType.BYTE + + yield text, score, toktype + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.sentencepiece_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class HfVocab: + def __init__( + self, + fname_tokenizer: Path, + fname_added_tokens: Optional[Path] = None, + ) -> None: + print("fname_tokenizer:", fname_tokenizer) + # Allow the tokenizer to default to slow or fast versions. + # Explicitly set tokenizer to use local paths. + self.tokenizer = AutoTokenizer.from_pretrained( + fname_tokenizer, + cache_dir=fname_tokenizer, + local_files_only=True, + ) + + # Initialize lists and dictionaries for added tokens + self.added_tokens_list = [] + self.added_tokens_dict = dict() + self.added_tokens_ids = set() + + # Process added tokens + for tok, tokidx in sorted( + self.tokenizer.get_added_vocab().items(), key=lambda x: x[1] + ): + # Only consider added tokens that are not in the base vocabulary + if tokidx >= self.tokenizer.vocab_size: + self.added_tokens_list.append(tok) + self.added_tokens_dict[tok] = tokidx + self.added_tokens_ids.add(tokidx) + + # Store special tokens and their IDs + self.specials = { tok: self.tokenizer.get_vocab()[tok] for tok in self.tokenizer.all_special_tokens } - self.special_ids: set[int] = set(self.tokenizer.all_special_ids) - self.reverse_vocab = {id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()} - self.vocab_size_base: int = self.tokenizer.vocab_size - self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_dict) - self.fname_tokenizer: Path = fname_tokenizer + self.special_ids = set(self.tokenizer.all_special_ids) - vocab_file = "tokenizer.model" - path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file) - if path_candidate is not None: - self.spm = SentencePieceProcessor(str(path_candidate)) - print(self.spm.vocab_size(), self.vocab_size_base) - else: - self.spm = None + # Set vocabulary sizes + self.vocab_size_base = self.tokenizer.vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - added_tokens_ids = set(self.added_tokens_dict.values()) + self.fname_tokenizer = fname_tokenizer + self.fname_added_tokens = fname_added_tokens - for i in range(self.vocab_size_base): - if i in added_tokens_ids: + def hf_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + reverse_vocab = { + id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items() + } + + for token_id in range(self.vocab_size_base): + # Skip processing added tokens here + if token_id in self.added_tokens_ids: continue - text = self.reverse_vocab[i].encode("utf-8") - yield text, self.get_token_score(i), self.get_token_type(i) + # Convert token text to bytes + token_text = reverse_vocab[token_id].encode("utf-8") - def get_token_type(self, token_id: int) -> gguf.TokenType: - toktype = gguf.TokenType.NORMAL + # Yield token text, score, and type + yield token_text, self.get_token_score(token_id), self.get_token_type( + token_id, self.special_ids # Reuse already stored special IDs + ) - if self.spm is not None and token_id < self.spm.vocab_size(): - if self.spm.is_unknown(token_id): - toktype = gguf.TokenType.UNKNOWN - if self.spm.is_control(token_id): - toktype = gguf.TokenType.CONTROL - if self.spm.is_unused(token_id): - toktype = gguf.TokenType.UNUSED - if self.spm.is_byte(token_id): - toktype = gguf.TokenType.BYTE - else: - token = self.reverse_vocab[token_id] - if token_id == self.unk_token_id: - toktype = gguf.TokenType.UNKNOWN - elif token_id in self.special_ids: - toktype = gguf.TokenType.CONTROL - elif len(token) == 6 and token.startswith("<0x") and token.endswith(">"): - toktype = gguf.TokenType.BYTE - - return toktype + def get_token_type(self, token_id: int, special_ids: set) -> gguf.TokenType: + # Determine token type based on whether it's a special token + return ( + gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL + ) def get_token_score(self, token_id: int) -> float: - if self.spm is not None and token_id < self.spm.vocab_size(): - return cast(float, self.spm.get_score(token_id)) - return 0.0 + # Placeholder for actual logic to determine the token's score + # This needs to be implemented based on specific requirements + return -1000.0 # Default score def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - - for text in self.added_tokens_dict: + for text in self.added_tokens_list: if text in self.specials: - - toktype = self.get_token_type(self.specials[text]) + toktype = self.get_token_type(self.specials[text], self.special_ids) score = self.get_token_score(self.specials[text]) else: @@ -422,45 +594,18 @@ class VocabLoader: yield text.encode("utf-8"), score, toktype - def has_newline_token(self) -> bool: - return '<0x0A>' in self.tokenizer.vocab or '\n' in self.tokenizer.vocab + def has_newline_token(self): + return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: yield from self.hf_tokens() yield from self.added_tokens() - def get_vocab_type(self) -> str: - path_candidates = [] - vocab_file = "tokenizer.model" - path_candidates.append(vocab_file) - path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file) - if path_candidate is not None: - return "llama" - - vocab_file = "vocab.json" - path_candidates.append(vocab_file) - path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file) - if path_candidate is not None: - return "gpt2" - - vocab_file = "tokenizer.json" - path_candidates.append(vocab_file) - path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file) - if path_candidate: - if not self.has_newline_token(): - return "gpt2" - return "llama" - - raise FileNotFoundError( - f"Could not find {path_candidates} in {self.fname_tokenizer} or its parent; " - "if it's in another directory, pass the directory as --vocab-dir" - ) - def __repr__(self) -> str: - return f"" + return f"" -Vocab: TypeAlias = 'VocabLoader' +Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab" # @@ -724,13 +869,17 @@ class LazyUnpickler(pickle.Unpickler): CLASSES: dict[tuple[str, str], Any] = { # getattr used here as a workaround for mypy not being smart enough to determine # the staticmethods have a __func__ attribute. - ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'), - ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'), - ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16), - ('torch', 'HalfStorage'): LazyStorageKind(DT_F16), - ('torch', 'FloatStorage'): LazyStorageKind(DT_F32), - ('torch', 'IntStorage'): LazyStorageKind(DT_I32), - ('torch', 'Tensor'): LazyTensor, + ("torch._tensor", "_rebuild_from_type_v2"): getattr( + rebuild_from_type_v2, "__func__" + ), + ("torch._utils", "_rebuild_tensor_v2"): getattr( + lazy_rebuild_tensor_v2, "__func__" + ), + ("torch", "BFloat16Storage"): LazyStorageKind(DT_BF16), + ("torch", "HalfStorage"): LazyStorageKind(DT_F16), + ("torch", "FloatStorage"): LazyStorageKind(DT_F32), + ("torch", "IntStorage"): LazyStorageKind(DT_I32), + ("torch", "Tensor"): LazyTensor, } def find_class(self, module: str, name: str) -> Any: @@ -839,32 +988,43 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None: - if params.n_vocab != vocab.vocab_size: - if params.n_vocab == vocab.vocab_size: - print("Ignoring added_tokens.json since model matches vocab size without it.") - vocab.added_tokens_dict = OrderedDict() - vocab.vocab_size = vocab.vocab_size - return + # Handle special case where the model's vocab size is not set + if params.n_vocab == -1: + raise ValueError( + f"The model's vocab size is set to -1 in params.json. Please update it manually. Maybe {vocab.vocab_size}?" + ) - if pad_vocab and params.n_vocab > vocab.vocab_size: - pad_count = params.n_vocab - vocab.vocab_size - print(f'Padding vocab with {pad_count} token(s) - through ') - for i in range(1, (params.n_vocab - vocab.vocab_size) + 1): - vocab.added_tokens_dict[f''] = -1 - vocab.vocab_size = params.n_vocab - return - msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer}" - msg += f" has {vocab.vocab_size})." - if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20: - msg += f" Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})." - if vocab.vocab_size < params.n_vocab: - msg += " Possibly try using the --padvocab option." - raise Exception(msg) + # Check for a vocab size mismatch + if params.n_vocab == vocab.vocab_size: + print("Ignoring added_tokens.json since model matches vocab size without it.") + return + + if pad_vocab and params.n_vocab > vocab.vocab_size: + pad_count = params.n_vocab - vocab.vocab_size + print( + f"Padding vocab with {pad_count} token(s) - through " + ) + for i in range(1, pad_count + 1): + vocab.added_tokens_dict[f""] = -1 + vocab.vocab_size = params.n_vocab + return + + msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer} has {vocab.vocab_size})." + if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20: + msg += f" Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})." + if vocab.vocab_size < params.n_vocab: + msg += " Add the --pad-vocab option and try again." + + raise Exception(msg) class OutputFile: - def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None: - self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess) + def __init__( + self, fname_out: Path, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE + ) -> None: + self.gguf = gguf.GGUFWriter( + fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess + ) def add_meta_arch(self, params: Params) -> None: name = "LLaMA" @@ -873,16 +1033,21 @@ class OutputFile: if params.n_ctx == 4096: name = "LLaMA v2" elif params.path_model is not None: - name = str(params.path_model.parent).split('/')[-1] + name = str(params.path_model.parent).split("/")[-1] - self.gguf.add_name (name) - self.gguf.add_context_length (params.n_ctx) - self.gguf.add_embedding_length (params.n_embd) - self.gguf.add_block_count (params.n_layer) - self.gguf.add_feed_forward_length (params.n_ff) + self.gguf.add_name(name) + self.gguf.add_context_length(params.n_ctx) + self.gguf.add_embedding_length(params.n_embd) + self.gguf.add_block_count(params.n_layer) + self.gguf.add_feed_forward_length(params.n_ff) self.gguf.add_rope_dimension_count(params.n_embd // params.n_head) - self.gguf.add_head_count (params.n_head) - self.gguf.add_head_count_kv (params.n_head_kv) + self.gguf.add_head_count(params.n_head) + self.gguf.add_head_count_kv(params.n_head_kv) + + if params.f_norm_eps is None: + raise ValueError("f_norm_eps is None") + + self.gguf.add_layer_norm_rms_eps(params.f_norm_eps) if params.n_experts: self.gguf.add_expert_count(params.n_experts) @@ -890,11 +1055,6 @@ class OutputFile: if params.n_experts_used: self.gguf.add_expert_used_count(params.n_experts_used) - if params.f_norm_eps: - self.gguf.add_layer_norm_rms_eps(params.f_norm_eps) - else: - raise ValueError('f_norm_eps is None') - if params.f_rope_freq_base is not None: self.gguf.add_rope_freq_base(params.f_rope_freq_base) @@ -912,18 +1072,44 @@ class OutputFile: if params.ftype is not None: self.gguf.add_file_type(params.ftype) - def add_meta_vocab(self, vocab: Vocab) -> None: + def handle_tokenizer_model(self, vocab: Vocab) -> str: + # Map the vocab types to the supported tokenizer models + tokenizer_model = { + SentencePieceVocab: "llama", + HfVocab: "llama", + BpeVocab: "gpt2", + }.get(type(vocab)) + + # Block if vocab type is not predefined + if tokenizer_model is None: + raise ValueError("Unknown vocab type: Not supported") + + return tokenizer_model + + def extract_vocabulary_from_model(self, vocab: Vocab) -> Tuple[list, list, list]: tokens = [] scores = [] toktypes = [] + # NOTE: `all_tokens` returns the base vocabulary and added tokens for text, score, toktype in vocab.all_tokens(): tokens.append(text) scores.append(score) toktypes.append(toktype) - vocab_type = vocab.get_vocab_type() - self.gguf.add_tokenizer_model(vocab_type) + return tokens, scores, toktypes + + def add_meta_vocab(self, vocab: Vocab) -> None: + # Handle the tokenizer model + tokenizer_model = self.handle_tokenizer_model(vocab) + + # Ensure that tokenizer_model is added to the GGUF model + self.gguf.add_tokenizer_model(tokenizer_model) + + # Extract model vocabulary for model conversion + tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab) + + # Add extracted token information for model conversion self.gguf.add_token_list(tokens) self.gguf.add_token_scores(scores) self.gguf.add_token_types(toktypes) @@ -933,10 +1119,14 @@ class OutputFile: def add_tensor_info(self, name: str, tensor: LazyTensor) -> None: n_elements = int(np.prod(tensor.shape)) - raw_dtype = getattr(tensor.data_type, 'ggml_type', None) - data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype + raw_dtype = getattr(tensor.data_type, "ggml_type", None) + data_type = ( + getattr(tensor.data_type, "quantized_type", None) or tensor.data_type.dtype + ) data_nbytes = tensor.data_type.elements_to_bytes(n_elements) - self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype = raw_dtype) + self.gguf.add_tensor_info( + name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype + ) def write_meta(self) -> None: self.gguf.write_header_to_file() @@ -950,11 +1140,14 @@ class OutputFile: @staticmethod def write_vocab_only( - fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, + fname_out: Path, + params: Params, + vocab: Vocab, + svocab: gguf.SpecialVocab, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, ) -> None: - check_vocab_size(params, vocab, pad_vocab = pad_vocab) + check_vocab_size(params, vocab, pad_vocab=pad_vocab) of = OutputFile(fname_out, endianess=endianess) @@ -982,12 +1175,17 @@ class OutputFile: @staticmethod def write_all( - fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, + fname_out: Path, + ftype: GGMLFileType, + params: Params, + model: LazyModel, + vocab: Vocab, + svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, ) -> None: - check_vocab_size(params, vocab, pad_vocab = pad_vocab) + check_vocab_size(params, vocab, pad_vocab=pad_vocab) of = OutputFile(fname_out, endianess=endianess) @@ -1004,18 +1202,30 @@ class OutputFile: of.write_tensor_info() # tensor data - ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency) + ndarrays_inner = bounded_parallel_map( + OutputFile.do_item, model.items(), concurrency=concurrency + ) if ftype == GGMLFileType.MostlyQ8_0: - ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, use_processpool_executor = True) + ndarrays = bounded_parallel_map( + OutputFile.maybe_do_quantize, + ndarrays_inner, + concurrency=concurrency, + max_workers=concurrency, + use_processpool_executor=True, + ) else: ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner) start = time.time() - for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)): + for i, ((name, lazy_tensor), ndarray) in enumerate( + zip(model.items(), ndarrays) + ): elapsed = time.time() - start - size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) + size = " x ".join(f"{dim:6d}" for dim in lazy_tensor.shape) padi = len(str(len(model))) - print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}") + print( + f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}" + ) of.gguf.write_tensor_data(ndarray) of.close() @@ -1145,30 +1355,95 @@ def load_some_model(path: Path) -> ModelPlus: return model_plus -def find_vocab_file_path(path: Path, vocab_file: str) -> Optional[Path]: - path2 = path / vocab_file - # Use `.parent` instead of /.. to handle the symlink case better. - path3 = path.parent / vocab_file +class VocabFactory: + def __init__(self, path: Path): + self.path = path + self.files = { + "tokenizer.model": None, + "vocab.json": None, + "tokenizer.json": None, + } + self._detect_files() - if path2.exists(): - return path2 - if path3.exists(): - return path3 + def _detect_files(self): + for file in self.files.keys(): + file_path = self.path / file + parent_file_path = self.path.parent / file + if file_path.exists(): + self.files[file] = file_path + elif parent_file_path.exists(): + self.files[file] = parent_file_path - return None + def _select_file(self, vocabtype: Optional[str]) -> Path: + if vocabtype in ["spm", "bpe"]: + # For SentencePiece and BPE, return specific files as before + file_key = "tokenizer.model" if vocabtype == "spm" else "vocab.json" + if self.files[file_key]: + return self.files[file_key] + else: + raise FileNotFoundError(f"{vocabtype} {file_key} not found.") + elif vocabtype == "hfft": + # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file + return self.path + else: + raise ValueError(f"Unsupported vocabulary type {vocabtype}") + + def _create_special_vocab( + self, + vocab: Vocab, + vocabtype: str, + model_parent_path: Path, + ) -> gguf.SpecialVocab: + load_merges = vocabtype == "bpe" + n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None + return gguf.SpecialVocab( + model_parent_path, + load_merges=load_merges, + special_token_types=None, # Predetermined or passed as a parameter + n_vocab=n_vocab, + ) + + def load_vocab( + self, vocabtype: str, model_parent_path: Path + ) -> Tuple[Vocab, gguf.SpecialVocab]: + path = self._select_file(vocabtype) + print(f"Loading vocab file '{path}', type '{vocabtype}'") + + added_tokens_path = path.parent / "added_tokens.json" + if vocabtype == "bpe": + vocab = BpeVocab( + path, added_tokens_path if added_tokens_path.exists() else None + ) + elif vocabtype == "spm": + vocab = SentencePieceVocab( + path, added_tokens_path if added_tokens_path.exists() else None + ) + elif vocabtype == "hfft": + vocab = HfVocab( + path, added_tokens_path if added_tokens_path.exists() else None + ) + else: + raise ValueError(f"Unsupported vocabulary type {vocabtype}") + special_vocab = self._create_special_vocab( + vocab, + vocabtype, + model_parent_path, + ) + return vocab, special_vocab -def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path: +def default_output_file(model_paths: list[Path], file_type: GGMLFileType) -> Path: namestr = { - GGMLFileType.AllF32: "f32", + GGMLFileType.AllF32: "f32", GGMLFileType.MostlyF16: "f16", - GGMLFileType.MostlyQ8_0:"q8_0", + GGMLFileType.MostlyQ8_0: "q8_0", }[file_type] ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf" if ret in model_paths: sys.stderr.write( f"Error: Default output path ({ret}) would overwrite the input. " - "Please explicitly specify a path using --outfile.\n") + "Please explicitly specify a path using --outfile.\n" + ) sys.exit(1) return ret @@ -1178,32 +1453,111 @@ def do_dump_model(model_plus: ModelPlus) -> None: print(f"model_plus.format = {model_plus.format!r}") print(f"model_plus.vocab = {model_plus.vocab!r}") for name, lazy_tensor in model_plus.model.items(): - print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") + print( + f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}" + ) -def main(args_in: list[str] | None = None) -> None: +def get_argument_parser() -> ArgumentParser: output_choices = ["f32", "f16"] if np.uint32(1) == np.uint32(1).newbyteorder("<"): # We currently only support Q8_0 output on little endian systems. output_choices.append("q8_0") - parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file") - parser.add_argument("--awq-path", type=Path, help="Path to scale awq cache file", default=None) - parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") - parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") - parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") - parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)") - parser.add_argument("--ctx", type=int, help="model training context (default: based on input)") - parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY) - parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine") - parser.add_argument("--padvocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides") - args = parser.parse_args(args_in) + parser = argparse.ArgumentParser( + description="Convert a LLaMa model to a GGML compatible file" + ) + + parser.add_argument( + "model", + type=Path, + help="Directory containing the model file or the model file itself (*.pth, *.pt, *.bin)", + ) + + parser.add_argument( + "--awq-path", + type=Path, + help="Path to the Activation-aware Weight Quantization cache file", + default=None, + ) + + parser.add_argument( + "--dump", + action="store_true", + help="Display the model content without converting it", + ) + + parser.add_argument( + "--dump-single", + action="store_true", + help="Display the content of a single model file without conversion", + ) + + parser.add_argument( + "--vocab-only", + action="store_true", + help="Extract and output only the vocabulary", + ) + + parser.add_argument( + "--outtype", + choices=output_choices, + help="Output format - note: q8_0 may be very slow (default: f16 or f32 based on input)", + ) + + parser.add_argument( + "--vocab-dir", + type=Path, + help="Directory containing the tokenizer.model, if separate from the model file", + ) + + parser.add_argument( + "--vocab-type", + choices=["spm", "bpe", "hfft"], # hfft: Hugging Face Fast Tokenizer + default="spm", + help="The vocabulary format used to define the tokenizer model (default: spm)", + ) + + parser.add_argument( + "--pad-vocab", + action="store_true", + help="Add padding tokens when the model's vocabulary size exceeds the tokenizer metadata", + ) + + parser.add_argument( + "--outfile", + type=Path, + help="Specify the path for the output file (default is based on input)", + ) + + parser.add_argument( + "--ctx", type=int, help="Model training context (default is based on input)" + ) + + parser.add_argument( + "--concurrency", + type=int, + help=f"Concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", + default=DEFAULT_CONCURRENCY, + ) + + parser.add_argument( + "--big-endian", + action="store_true", + help="Indicate that the model is executed on a big-endian machine", + ) + + return parser + + +def main(argv: Optional[list[str]] = None) -> None: + parser = get_argument_parser() + args = parser.parse_args(argv) + if args.awq_path: - sys.path.insert(1, str(Path(__file__).parent / 'awq-py')) + sys.path.insert(1, str(Path(__file__).resolve().parent / "awq-py")) from awq.apply_awq import add_scale_weights + tmp_model_path = args.model / "weighted_model" if tmp_model_path.is_dir(): print(f"{tmp_model_path} exists as a weighted model.") @@ -1222,22 +1576,27 @@ def main(args_in: list[str] | None = None) -> None: if not args.vocab_only: model_plus = load_some_model(args.model) else: - model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None) + model_plus = ModelPlus( + model={}, paths=[args.model / "dummy"], format="none", vocab=None + ) if args.dump: do_dump_model(model_plus) return + endianess = gguf.GGUFEndian.LITTLE - if args.bigendian: + if args.big_endian: endianess = gguf.GGUFEndian.BIG params = Params.load(model_plus) if params.n_ctx == -1: if args.ctx is None: - raise Exception("The model doesn't have a context size, and you didn't specify one with --ctx\n" - "Please specify one with --ctx:\n" - " - LLaMA v1: --ctx 2048\n" - " - LLaMA v2: --ctx 4096\n") + raise Exception( + "The model doesn't have a context size, and you didn't specify one with --ctx\n" + "Please specify one with --ctx:\n" + " - LLaMA v1: --ctx 2048\n" + " - LLaMA v2: --ctx 4096\n" + ) params.n_ctx = args.ctx if args.outtype: @@ -1249,47 +1608,51 @@ def main(args_in: list[str] | None = None) -> None: print(f"params = {params}") - vocab: Vocab + model_parent_path = model_plus.paths[0].parent + vocab_path = Path(args.vocab_dir or args.model or model_parent_path) + vocab_factory = VocabFactory(vocab_path) + vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type, model_parent_path) + if args.vocab_only: if not args.outfile: raise ValueError("need --outfile if using --vocab-only") - # FIXME: Try to respect vocab_dir somehow? - vocab = VocabLoader(params, args.vocab_dir or args.model) - special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, - load_merges = True, - n_vocab = vocab.vocab_size) outfile = args.outfile - OutputFile.write_vocab_only(outfile, params, vocab, special_vocab, - endianess = endianess, pad_vocab = args.padvocab) + OutputFile.write_vocab_only( + outfile, + params, + vocab, + special_vocab, + endianess=endianess, + pad_vocab=args.pad_vocab, + ) print(f"Wrote {outfile}") return if model_plus.vocab is not None and args.vocab_dir is None: vocab = model_plus.vocab - else: - vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent - vocab = VocabLoader(params, vocab_dir) - # FIXME: Try to respect vocab_dir somehow? - print(f"Vocab info: {vocab}") - special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, - load_merges = True, - n_vocab = vocab.vocab_size) - - print(f"Special vocab info: {special_vocab}") - model = model_plus.model - model = convert_model_names(model, params) - ftype = pick_output_type(model, args.outtype) - model = convert_to_output_type(model, ftype) - outfile = args.outfile or default_outfile(model_plus.paths, ftype) + model = model_plus.model + model = convert_model_names(model, params) + ftype = pick_output_type(model, args.outtype) + model = convert_to_output_type(model, ftype) + outfile = args.outfile or default_output_file(model_plus.paths, ftype) params.ftype = ftype print(f"Writing {outfile}, format {ftype}") - OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, - concurrency = args.concurrency, endianess = endianess, pad_vocab = args.padvocab) + OutputFile.write_all( + outfile, + ftype, + params, + model, + vocab, + special_vocab, + concurrency=args.concurrency, + endianess=endianess, + pad_vocab=args.pad_vocab, + ) print(f"Wrote {outfile}") -if __name__ == '__main__': - main() +if __name__ == "__main__": + main(sys.argv[1:]) # Exclude the first element (script name) from sys.argv From 4f56458d34cb13dcbf69aca650e9bf77d5497e6f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Wed, 10 Jan 2024 01:04:33 +0100 Subject: [PATCH 07/42] Python script to compare commits with llama-bench (#4844) --- scripts/compare-llama-bench.py | 356 +++++++++++++++++++++++++++++++++ 1 file changed, 356 insertions(+) create mode 100755 scripts/compare-llama-bench.py diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py new file mode 100755 index 000000000..bc1714487 --- /dev/null +++ b/scripts/compare-llama-bench.py @@ -0,0 +1,356 @@ +#!/usr/bin/env python3 + +import argparse +import heapq +import sys +import os +from glob import glob +import sqlite3 + +try: + import git + from tabulate import tabulate +except ImportError: + print("ERROR: the following Python libraries are required: GitPython, tabulate.") + sys.exit(1) + +# Properties by which to differentiate results per commit: +KEY_PROPERTIES = [ + "cuda", "opencl", "metal", "gpu_blas", "blas", "cpu_info", "gpu_info", "model_filename", + "model_type", "model_size", "model_n_params", "n_batch", "n_threads", "type_k", "type_v", + "n_gpu_layers", "main_gpu", "no_kv_offload", "mul_mat_q", "tensor_split", "n_prompt", "n_gen" +] + +# Properties that are boolean and are converted to Yes/No for the table: +BOOL_PROPERTIES = ["cuda", "opencl", "metal", "gpu_blas", "blas"] + +# Header names for the table: +PRETTY_NAMES = { + "cuda": "CUDA", "opencl": "OpenCL", "metal": "Metal", "gpu_blas": "GPU BLAS", "blas": "BLAS", + "cpu_info": "CPU", "gpu_info": "GPU", "model_filename": "File", "model_type": "Model", + "model_size": "Model Size [GiB]", "model_n_params": "Num. of Parameters", + "n_batch": "Batch size", "n_threads": "Threads", "type_k": "K type", "type_v": "V type", + "n_gpu_layers": "GPU layers", "main_gpu": "Main GPU", "no_kv_offload": "NKVO", + "mul_mat_q": "MMQ", "tensor_split": "Tensor split" +} + +DEFAULT_SHOW = ["model_type"] # Always show these properties by default. +DEFAULT_HIDE = ["model_filename"] # Always hide these properties by default. +GPU_NAME_STRIP = ["NVIDIA GeForce ", "Tesla ", "AMD Radeon "] # Strip prefixes for smaller tables. + +DESCRIPTION = """Creates tables from llama-bench data written to an SQLite database. Example usage (Linux): + +$ git checkout master +$ make clean && make llama-bench +$ ./llama-bench -o sql | sqlite3 llama-bench.sqlite +$ git checkout some_branch +$ make clean && make llama-bench +$ ./llama-bench -o sql | sqlite3 llama-bench.sqlite +$ ./scripts/compare-llama-bench.py + +Performance numbers from multiple runs per commit are averaged WITHOUT being weighted by the --repetitions parameter of llama-bench. +""" + +parser = argparse.ArgumentParser( + description=DESCRIPTION, formatter_class=argparse.RawDescriptionHelpFormatter) +help_b = ( + "The baseline commit to compare performance to. " + "Accepts either a branch name, tag name, or commit hash. " + "Defaults to latest master commit with data." +) +parser.add_argument("-b", "--baseline", help=help_b) +help_c = ( + "The commit whose performance is to be compared to the baseline. " + "Accepts either a branch name, tag name, or commit hash. " + "Defaults to the non-master commit for which llama-bench was run most recently." +) +parser.add_argument("-c", "--compare", help=help_c) +help_i = ( + "Input SQLite file for comparing commits. " + "Defaults to 'llama-bench.sqlite' in the current working directory. " + "If no such file is found and there is exactly one .sqlite file in the current directory, " + "that file is instead used as input." +) +parser.add_argument("-i", "--input", help=help_i) +help_o = ( + "Output format for the table. " + "Defaults to 'pipe' (GitHub compatible). " + "Also supports e.g. 'latex' or 'mediawiki'. " + "See tabulate documentation for full list." +) +parser.add_argument("-o", "--output", help=help_o, default="pipe") +help_s = ( + "Columns to add to the table. " + "Accepts a comma-separated list of values. " + f"Legal values: {', '.join(KEY_PROPERTIES[:-2])}. " + "Defaults to model name (model_type) and CPU and/or GPU name (cpu_info, gpu_info) " + "plus any column where not all data points are the same. " + "If the columns are manually specified, then the results for each unique combination of the " + "specified values are averaged WITHOUT weighing by the --repetitions parameter of llama-bench." +) +parser.add_argument("-s", "--show", help=help_s) + +known_args, unknown_args = parser.parse_known_args() + +if unknown_args: + print(f"ERROR: Received unknown args: {unknown_args}.") + print() + parser.print_help() + sys.exit(1) + +input_file = known_args.input +if input_file is None and os.path.exists("./llama-bench.sqlite"): + input_file = "llama-bench.sqlite" +if input_file is None: + sqlite_files = glob("*.sqlite") + if len(sqlite_files) == 1: + input_file = sqlite_files[0] + +if input_file is None: + print("ERROR: Cannot find a suitable input file, please provide one.") + print() + parser.print_help() + sys.exit(1) + +connection = sqlite3.connect(input_file) +cursor = connection.cursor() +builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall() + +try: + repo = git.Repo(".", search_parent_directories=True) +except git.exc.InvalidGitRepositoryError: + repo = None + + +def find_parent_in_data(commit): + """Helper function to find the most recent parent measured in number of commits for which there is data.""" + heap = [(0, commit)] + seen_hexsha8 = set() + while heap: + depth, current_commit = heapq.heappop(heap) + current_hexsha8 = commit.hexsha[:8] + if (current_hexsha8,) in builds: + return current_hexsha8 + for parent in commit.parents: + parent_hexsha8 = parent.hexsha[:8] + if parent_hexsha8 not in seen_hexsha8: + seen_hexsha8.add(parent_hexsha8) + heapq.heappush(heap, (depth + 1, parent)) + return None + + +def get_all_parent_hexsha8s(commit): + """Helper function to recursively get hexsha8 values for all parents of a commit.""" + unvisited = [commit] + visited = [] + + while unvisited: + current_commit = unvisited.pop(0) + visited.append(current_commit.hexsha[:8]) + for parent in current_commit.parents: + if parent.hexsha[:8] not in visited: + unvisited.append(parent) + + return visited + + +def get_commit_name(hexsha8): + """Helper function to find a human-readable name for a commit if possible.""" + if repo is None: + return hexsha8 + for h in repo.heads: + if h.commit.hexsha[:8] == hexsha8: + return h.name + for t in repo.tags: + if t.commit.hexsha[:8] == hexsha8: + return t.name + return hexsha8 + + +def get_commit_hexsha8(name): + """Helper function to search for a commit given a human-readable name.""" + if repo is None: + return None + for h in repo.heads: + if h.name == name: + return h.commit.hexsha[:8] + for t in repo.tags: + if t.name == name: + return t.commit.hexsha[:8] + return None + + +hexsha8_baseline = name_baseline = None + +# If the user specified a baseline, try to find a commit for it: +if known_args.baseline is not None: + if (known_args.baseline,) in builds: + hexsha8_baseline = known_args.baseline + if hexsha8_baseline is None: + hexsha8_baseline = get_commit_hexsha8(known_args.baseline) + name_baseline = known_args.baseline + if hexsha8_baseline is None: + print(f"ERROR: cannot find data for baseline={known_args.baseline}.") + sys.exit(1) +# Otherwise, search for the most recent parent of master for which there is data: +elif repo is not None: + hexsha8_baseline = find_parent_in_data(repo.heads.master.commit) + + if hexsha8_baseline is None: + print("ERROR: No baseline was provided and did not find data for any master branch commits.") + print() + parser.print_help() + sys.exit(1) +else: + print( + "ERROR: No baseline was provided and the current working directory " + "is not part of a git repository from which a baseline could be inferred." + ) + print() + parser.print_help() + sys.exit(1) + + +name_baseline = get_commit_name(hexsha8_baseline) + +hexsha8_compare = name_compare = None + +# If the user has specified a compare value, try to find a corresponding commit: +if known_args.compare is not None: + if (known_args.compare,) in builds: + hexsha8_compare = known_args.compare + if hexsha8_compare is None: + hexsha8_compare = get_commit_hexsha8(known_args.compare) + name_compare = known_args.compare + if hexsha8_compare is None: + print(f"ERROR: cannot find data for baseline={known_args.compare}.") + sys.exit(1) +# Otherwise, search for the commit for llama-bench was most recently run +# and that is not a parent of master: +elif repo is not None: + hexsha8s_master = get_all_parent_hexsha8s(repo.heads.master.commit) + builds_timestamp = cursor.execute( + "SELECT build_commit, test_time FROM test ORDER BY test_time;").fetchall() + for (hexsha8, _) in reversed(builds_timestamp): + if hexsha8 not in hexsha8s_master: + hexsha8_compare = hexsha8 + break + + if hexsha8_compare is None: + print("ERROR: No compare target was provided and did not find data for any non-master commits.") + print() + parser.print_help() + sys.exit(1) +else: + print( + "ERROR: No compare target was provided and the current working directory " + "is not part of a git repository from which a compare target could be inferred." + ) + print() + parser.print_help() + sys.exit(1) + +name_compare = get_commit_name(hexsha8_compare) + + +def get_rows(properties): + """ + Helper function that gets table rows for some list of properties. + Rows are created by combining those where all provided properties are equal. + The resulting rows are then grouped by the provided properties and the t/s values are averaged. + The returned rows are unique in terms of property combinations. + """ + select_string = ", ".join( + [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"]) + equal_string = " AND ".join( + [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [ + f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"] + ) + group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt"]) + query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} " + f"GROUP BY {group_order_string} ORDER BY {group_order_string};") + return cursor.execute(query).fetchall() + + +# If the user provided columns to group the results by, use them: +if known_args.show is not None: + show = known_args.show.split(",") + unknown_cols = [] + for prop in show: + if prop not in KEY_PROPERTIES[:-2]: # Last two values are n_prompt, n_gen. + unknown_cols.append(prop) + if unknown_cols: + print(f"ERROR: Unknown values for --show: {', '.join(unknown_cols)}") + print() + parser.print_usage() + sys.exit(1) + rows_show = get_rows(show) +# Otherwise, select those columns where the values are not all the same: +else: + rows_full = get_rows(KEY_PROPERTIES) + properties_different = [] + for i, kp_i in enumerate(KEY_PROPERTIES): + if kp_i in DEFAULT_SHOW or kp_i == "n_prompt" or kp_i == "n_gen": + continue + for row_full in rows_full: + if row_full[i] != rows_full[0][i]: + properties_different.append(kp_i) + break + + show = [] + # Show CPU and/or GPU by default even if the hardware for all results is the same: + if "gpu_blas" not in properties_different and "n_gpu_layers" not in properties_different: + gpu_blas = bool(rows_full[0][KEY_PROPERTIES.index("gpu_blas")]) + ngl = int(rows_full[0][KEY_PROPERTIES.index("n_gpu_layers")]) + + if not gpu_blas or ngl != 99 and "cpu_info" not in properties_different: + show.append("cpu_info") + if gpu_blas and "gpu_info" not in properties_different: + show.append("gpu_info") + + show += DEFAULT_SHOW + show += properties_different + for prop in DEFAULT_HIDE: + try: + show.remove(prop) + except ValueError: + pass + rows_show = get_rows(show) + +table = [] +for row in rows_show: + n_prompt = int(row[-4]) + n_gen = int(row[-3]) + assert n_prompt == 0 or n_gen == 0 + test_name = f"tg{n_gen}" if n_prompt == 0 else f"pp{n_prompt}" + # Regular columns test name avg t/s values Speedup + # VVVVVVVVVVVVV VVVVVVVVV VVVVVVVVVVVVVV VVVVVVV + table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])]) + +# Some a-posteriori fixes to make the table contents prettier: +for bool_property in BOOL_PROPERTIES: + if bool_property in show: + ip = show.index(bool_property) + for row_table in table: + row_table[ip] = "Yes" if int(row_table[ip]) == 1 else "No" + +if "model_size" in show: + ip = show.index("model_size") + for row_table in table: + row_table[ip] = float(row_table[ip]) / 1024 ** 3 + +if "gpu_info" in show: + ip = show.index("gpu_info") + for gns in GPU_NAME_STRIP: + for row_table in table: + row_table[ip] = row_table[ip].replace(gns, "") + +headers = [PRETTY_NAMES[p] for p in show] +headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"] + +print(tabulate( + table, + headers=headers, + floatfmt=".2f", + tablefmt=known_args.output +)) From d34633d8db6c2e400355de4862cd699154ecc73f Mon Sep 17 00:00:00 2001 From: John <78893154+cmp-nct@users.noreply.github.com> Date: Wed, 10 Jan 2024 14:37:09 +0100 Subject: [PATCH 08/42] clip : support more quantization types (#4846) Uses ggml functions instead of hardcoded names and adds support to quantize into the modern Q-K variants. This is just the bare minimum to get k-types working - a more refined choice of types would be needed to get best quality on low quantizations. I ran a few tests, it doesn't break anything I could notice and a Q6_K ViT works almost as well as Q8_0 but 3 times the inference speed. --- examples/llava/clip.cpp | 62 ++++++++++++++++------------------------- 1 file changed, 24 insertions(+), 38 deletions(-) diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index cfb79e789..2ae8853d3 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -126,24 +126,7 @@ static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::str } static std::string get_ftype(int ftype) { - switch (ftype) { - case 0: - return "f32"; - case 1: - return "f16"; - case 2: - return "q4_0"; - case 3: - return "q4_1"; - case 6: - return "q5_0"; - case 7: - return "q5_1"; - case 8: - return "q8_0"; - default: - throw std::runtime_error(format("%s: Unrecognized file type: %d\n", __func__, ftype)); - } + return ggml_type_name(static_cast(ftype)); } // @@ -533,6 +516,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { buffer_size += n_tensors * 128 /* CLIP PADDING */; clip_ctx * new_clip = new clip_ctx; + #ifdef GGML_USE_CUBLAS new_clip->backend = ggml_backend_cuda_init(0); printf("%s: CLIP using CUDA backend\n", __func__); @@ -543,6 +527,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { printf("%s: CLIP using Metal backend\n", __func__); #endif + if (!new_clip->backend) { new_clip->backend = ggml_backend_cpu_init(); printf("%s: CLIP using CPU backend\n", __func__); @@ -931,26 +916,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i ggml_type type = GGML_TYPE_Q4_1; - switch (itype) { - case 2: - type = GGML_TYPE_Q4_0; - break; - case 3: - type = GGML_TYPE_Q4_1; - break; - case 6: - type = GGML_TYPE_Q5_0; - break; - case 7: - type = GGML_TYPE_Q5_1; - break; - case 8: - type = GGML_TYPE_Q8_0; - break; - default: - fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); - return false; - }; + assert(itype < GGML_TYPE_COUNT); + type = static_cast(itype); auto * ctx_clip = clip_model_load(fname_inp, 2); @@ -1010,6 +977,10 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i if (quantize) { new_type = type; + if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) { + new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type + // fprintf(stderr, "%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type)); + } const size_t n_elms = ggml_nelements(cur); float * f32_data; @@ -1054,6 +1025,21 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i case GGML_TYPE_Q8_0: { new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); } break; + case GGML_TYPE_Q2_K: { + new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); + } break; + case GGML_TYPE_Q3_K: { + new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); + } break; + case GGML_TYPE_Q4_K: { + new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); + } break; + case GGML_TYPE_Q5_K: { + new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); + } break; + case GGML_TYPE_Q6_K: { + new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); + } break; default: { fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type); return false; From 329ff615699d32f596d4ebf8baba654c30064e0d Mon Sep 17 00:00:00 2001 From: Austin <77757836+teleprint-me@users.noreply.github.com> Date: Wed, 10 Jan 2024 08:39:09 -0500 Subject: [PATCH 09/42] llama : recognize 1B phi models (#4847) This update categorizes models with 24 layers as MODEL_1B, ensuring compatibility with different Phi model variants without impacting existing Phi-2 model functionality. --- llama.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/llama.cpp b/llama.cpp index 8e0717db9..0f09d0c2b 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2829,6 +2829,7 @@ static void llm_load_hparams( ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_3B; break; default: model.type = e_model::MODEL_UNKNOWN; } From 57d016ba2d46a6e22517a31a75cebb48f9e234b6 Mon Sep 17 00:00:00 2001 From: Brian Date: Thu, 11 Jan 2024 01:09:53 +1100 Subject: [PATCH 10/42] llama : add additional suffixes for model params (#4834) * llm_load_print_meta: Add additional suffixs for model params * Update llama.cpp model param log remove unneeded comments and convert from > to >= --- llama.cpp | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 0f09d0c2b..e1f1932ba 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3146,7 +3146,15 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); - LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, ml.n_elements*1e-9); + if (ml.n_elements >= 1e12) { + LLAMA_LOG_INFO("%s: model params = %.2f T\n", __func__, ml.n_elements*1e-12); + } else if (ml.n_elements >= 1e9) { + LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, ml.n_elements*1e-9); + } else if (ml.n_elements >= 1e6) { + LLAMA_LOG_INFO("%s: model params = %.2f M\n", __func__, ml.n_elements*1e-6); + } else { + LLAMA_LOG_INFO("%s: model params = %.2f K\n", __func__, ml.n_elements*1e-3); + } if (ml.n_bytes < GiB) { LLAMA_LOG_INFO("%s: model size = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); } else { From cd108e641dbdedd8c5641c4cec1762f751f38136 Mon Sep 17 00:00:00 2001 From: Behnam M <58621210+ibehnam@users.noreply.github.com> Date: Wed, 10 Jan 2024 14:56:05 -0500 Subject: [PATCH 11/42] server : add a `/health` endpoint (#4860) * added /health endpoint to the server * added comments on the additional /health endpoint * Better handling of server state When the model is being loaded, the server state is `LOADING_MODEL`. If model-loading fails, the server state becomes `ERROR`, otherwise it becomes `READY`. The `/health` endpoint provides more granular messages now according to the server_state value. * initialized server_state * fixed a typo * starting http server before initializing the model * Update server.cpp * Update server.cpp * fixes * fixes * fixes * made ServerState atomic and turned two-line spaces into one-line --- examples/server/server.cpp | 199 +++++++++++++++++++++---------------- 1 file changed, 113 insertions(+), 86 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 6c7fcd176..1cca634d5 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -26,6 +26,7 @@ #include #include #include +#include #ifndef SERVER_VERBOSE #define SERVER_VERBOSE 1 @@ -146,6 +147,12 @@ static std::vector base64_decode(const std::string & encoded_string) // parallel // +enum ServerState { + LOADING_MODEL, // Server is starting up, model not fully loaded yet + READY, // Server is ready and model is loaded + ERROR // An error occurred, load_model failed +}; + enum task_type { COMPLETION_TASK, CANCEL_TASK @@ -2453,7 +2460,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } } - static std::string random_string() { static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"); @@ -2790,15 +2796,117 @@ int main(int argc, char **argv) {"system_info", llama_print_system_info()}, }); - // load the model - if (!llama.load_model(params)) + httplib::Server svr; + + std::atomic server_state{LOADING_MODEL}; + + svr.set_default_headers({{"Server", "llama.cpp"}, + {"Access-Control-Allow-Origin", "*"}, + {"Access-Control-Allow-Headers", "content-type"}}); + + svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) { + ServerState current_state = server_state.load(); + switch(current_state) { + case READY: + res.set_content(R"({"status": "ok"})", "application/json"); + res.status = 200; // HTTP OK + break; + case LOADING_MODEL: + res.set_content(R"({"status": "loading model"})", "application/json"); + res.status = 503; // HTTP Service Unavailable + break; + case ERROR: + res.set_content(R"({"status": "error", "error": "Model failed to load"})", "application/json"); + res.status = 500; // HTTP Internal Server Error + break; + } + }); + + svr.set_logger(log_server_request); + + svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep) + { + const char fmt[] = "500 Internal Server Error\n%s"; + char buf[BUFSIZ]; + try + { + std::rethrow_exception(std::move(ep)); + } + catch (std::exception &e) + { + snprintf(buf, sizeof(buf), fmt, e.what()); + } + catch (...) + { + snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); + } + res.set_content(buf, "text/plain; charset=utf-8"); + res.status = 500; + }); + + svr.set_error_handler([](const httplib::Request &, httplib::Response &res) + { + if (res.status == 401) + { + res.set_content("Unauthorized", "text/plain; charset=utf-8"); + } + if (res.status == 400) + { + res.set_content("Invalid request", "text/plain; charset=utf-8"); + } + else if (res.status == 404) + { + res.set_content("File Not Found", "text/plain; charset=utf-8"); + res.status = 404; + } + }); + + // set timeouts and change hostname and port + svr.set_read_timeout (sparams.read_timeout); + svr.set_write_timeout(sparams.write_timeout); + + if (!svr.bind_to_port(sparams.hostname, sparams.port)) { + fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port); return 1; } - llama.initialize(); + // Set the base directory for serving static files + svr.set_base_dir(sparams.public_path); - httplib::Server svr; + // to make it ctrl+clickable: + LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port); + + std::unordered_map log_data; + log_data["hostname"] = sparams.hostname; + log_data["port"] = std::to_string(sparams.port); + + if (!sparams.api_key.empty()) { + log_data["api_key"] = "api_key: ****" + sparams.api_key.substr(sparams.api_key.length() - 4); + } + + LOG_INFO("HTTP server listening", log_data); + // run the HTTP server in a thread - see comment below + std::thread t([&]() + { + if (!svr.listen_after_bind()) + { + server_state.store(ERROR); + return 1; + } + + return 0; + }); + + // load the model + if (!llama.load_model(params)) + { + server_state.store(ERROR); + return 1; + } else { + llama.initialize(); + server_state.store(READY); + } // Middleware for API key validation auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool { @@ -2826,10 +2934,6 @@ int main(int argc, char **argv) return false; }; - svr.set_default_headers({{"Server", "llama.cpp"}, - {"Access-Control-Allow-Origin", "*"}, - {"Access-Control-Allow-Headers", "content-type"}}); - // this is only called if no index.html is found in the public --path svr.Get("/", [](const httplib::Request &, httplib::Response &res) { @@ -2937,8 +3041,6 @@ int main(int argc, char **argv) } }); - - svr.Get("/v1/models", [¶ms](const httplib::Request&, httplib::Response& res) { std::time_t t = std::time(0); @@ -3157,81 +3259,6 @@ int main(int argc, char **argv) return res.set_content(result.result_json.dump(), "application/json; charset=utf-8"); }); - svr.set_logger(log_server_request); - - svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep) - { - const char fmt[] = "500 Internal Server Error\n%s"; - char buf[BUFSIZ]; - try - { - std::rethrow_exception(std::move(ep)); - } - catch (std::exception &e) - { - snprintf(buf, sizeof(buf), fmt, e.what()); - } - catch (...) - { - snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); - } - res.set_content(buf, "text/plain; charset=utf-8"); - res.status = 500; - }); - - svr.set_error_handler([](const httplib::Request &, httplib::Response &res) - { - if (res.status == 401) - { - res.set_content("Unauthorized", "text/plain; charset=utf-8"); - } - if (res.status == 400) - { - res.set_content("Invalid request", "text/plain; charset=utf-8"); - } - else if (res.status == 404) - { - res.set_content("File Not Found", "text/plain; charset=utf-8"); - res.status = 404; - } - }); - - // set timeouts and change hostname and port - svr.set_read_timeout (sparams.read_timeout); - svr.set_write_timeout(sparams.write_timeout); - - if (!svr.bind_to_port(sparams.hostname, sparams.port)) - { - fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port); - return 1; - } - - // Set the base directory for serving static files - svr.set_base_dir(sparams.public_path); - - // to make it ctrl+clickable: - LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port); - - std::unordered_map log_data; - log_data["hostname"] = sparams.hostname; - log_data["port"] = std::to_string(sparams.port); - - if (!sparams.api_key.empty()) { - log_data["api_key"] = "api_key: ****" + sparams.api_key.substr(sparams.api_key.length() - 4); - } - - LOG_INFO("HTTP server listening", log_data); - // run the HTTP server in a thread - see comment below - std::thread t([&]() - { - if (!svr.listen_after_bind()) - { - return 1; - } - - return 0; - }); - // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!? // "Bus error: 10" - this is on macOS, it does not crash on Linux //std::thread t2([&]() From 5c1980d8d4c4e0c0af77359f81cc44d90b3f250b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 11 Jan 2024 09:10:34 +0200 Subject: [PATCH 12/42] server : fix build + rename enums (#4870) --- examples/server/server.cpp | 36 ++++++++++++++++++------------------ 1 file changed, 18 insertions(+), 18 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 1cca634d5..4a0714997 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -147,15 +147,15 @@ static std::vector base64_decode(const std::string & encoded_string) // parallel // -enum ServerState { - LOADING_MODEL, // Server is starting up, model not fully loaded yet - READY, // Server is ready and model is loaded - ERROR // An error occurred, load_model failed +enum server_state { + SERVER_STATE_LOADING_MODEL, // Server is starting up, model not fully loaded yet + SERVER_STATE_READY, // Server is ready and model is loaded + SERVER_STATE_ERROR // An error occurred, load_model failed }; enum task_type { - COMPLETION_TASK, - CANCEL_TASK + TASK_TYPE_COMPLETION, + TASK_TYPE_CANCEL, }; struct task_server { @@ -1402,7 +1402,7 @@ struct llama_server_context task.data = std::move(data); task.infill_mode = infill; task.embedding_mode = embedding; - task.type = COMPLETION_TASK; + task.type = TASK_TYPE_COMPLETION; task.multitask_id = multitask_id; // when a completion task's prompt array is not a singleton, we split it into multiple requests @@ -1524,7 +1524,7 @@ struct llama_server_context std::unique_lock lock(mutex_tasks); task_server task; task.id = id_gen++; - task.type = CANCEL_TASK; + task.type = TASK_TYPE_CANCEL; task.target_id = task_id; queue_tasks.push_back(task); condition_tasks.notify_one(); @@ -1560,7 +1560,7 @@ struct llama_server_context queue_tasks.erase(queue_tasks.begin()); switch (task.type) { - case COMPLETION_TASK: { + case TASK_TYPE_COMPLETION: { llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); if (slot == nullptr) { @@ -1589,7 +1589,7 @@ struct llama_server_context break; } } break; - case CANCEL_TASK: { // release slot linked with the task id + case TASK_TYPE_CANCEL: { // release slot linked with the task id for (auto & slot : slots) { if (slot.task_id == task.target_id) @@ -2798,24 +2798,24 @@ int main(int argc, char **argv) httplib::Server svr; - std::atomic server_state{LOADING_MODEL}; + std::atomic state{SERVER_STATE_LOADING_MODEL}; svr.set_default_headers({{"Server", "llama.cpp"}, {"Access-Control-Allow-Origin", "*"}, {"Access-Control-Allow-Headers", "content-type"}}); svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) { - ServerState current_state = server_state.load(); + server_state current_state = state.load(); switch(current_state) { - case READY: + case SERVER_STATE_READY: res.set_content(R"({"status": "ok"})", "application/json"); res.status = 200; // HTTP OK break; - case LOADING_MODEL: + case SERVER_STATE_LOADING_MODEL: res.set_content(R"({"status": "loading model"})", "application/json"); res.status = 503; // HTTP Service Unavailable break; - case ERROR: + case SERVER_STATE_ERROR: res.set_content(R"({"status": "error", "error": "Model failed to load"})", "application/json"); res.status = 500; // HTTP Internal Server Error break; @@ -2891,7 +2891,7 @@ int main(int argc, char **argv) { if (!svr.listen_after_bind()) { - server_state.store(ERROR); + state.store(SERVER_STATE_ERROR); return 1; } @@ -2901,11 +2901,11 @@ int main(int argc, char **argv) // load the model if (!llama.load_model(params)) { - server_state.store(ERROR); + state.store(SERVER_STATE_ERROR); return 1; } else { llama.initialize(); - server_state.store(READY); + state.store(SERVER_STATE_READY); } // Middleware for API key validation From 7a9f75c38b5e62fe27b8a5a3ed823b4a3714024b Mon Sep 17 00:00:00 2001 From: Behnam M <58621210+ibehnam@users.noreply.github.com> Date: Thu, 11 Jan 2024 02:12:05 -0500 Subject: [PATCH 13/42] server : update readme to document the new `/health` endpoint (#4866) * added /health endpoint to the server * added comments on the additional /health endpoint * Better handling of server state When the model is being loaded, the server state is `LOADING_MODEL`. If model-loading fails, the server state becomes `ERROR`, otherwise it becomes `READY`. The `/health` endpoint provides more granular messages now according to the server_state value. * initialized server_state * fixed a typo * starting http server before initializing the model * Update server.cpp * Update server.cpp * fixes * fixes * fixes * made ServerState atomic and turned two-line spaces into one-line * updated `server` readme to document the `/health` endpoint too --- examples/server/README.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/examples/server/README.md b/examples/server/README.md index d85a14f89..dc27e72b9 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -110,6 +110,10 @@ node index.js ``` ## API Endpoints +- **GET** `/health`: Returns the current state of the server: + - `{"status": "loading model"}` if the model is still being loaded. + - `{"status": "error"}` if the model failed to load. + - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below. - **POST** `/completion`: Given a `prompt`, it returns the predicted completion. From f34432ca1e0b288129390c1db8296a82aaf1e632 Mon Sep 17 00:00:00 2001 From: Erik Scholz Date: Fri, 5 Jan 2024 16:00:00 +0100 Subject: [PATCH 14/42] fix : cuda order of synchronization when setting a buffer (ggml/679) * fix : cuda order of synchronization when setting a buffer * also sync before memcpy --------- Co-authored-by: slaren --- ggml-cuda.cu | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index e26260a35..900f7ba4a 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -10184,8 +10184,8 @@ static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, gg ggml_cuda_set_device(ctx->device); CUDA_CHECK(cudaDeviceSynchronize()); - CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice)); + CUDA_CHECK(cudaDeviceSynchronize()); } 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) { From c910e3c28a1caee8cb1398143d582dd9ab697e68 Mon Sep 17 00:00:00 2001 From: Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com> Date: Tue, 9 Jan 2024 11:16:37 -0500 Subject: [PATCH 15/42] Fix execlp call (ggml/689) NULL can be an integer constant expression with the value zero, in this case the behavior would be undefined because of an incorrect type being passed to the variable arguments. --- ggml.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml.c b/ggml.c index adb387100..4a0ec4c44 100644 --- a/ggml.c +++ b/ggml.c @@ -132,7 +132,7 @@ void ggml_print_backtrace(void) { "-ex", "bt -frame-info source-and-location", "-ex", "detach", "-ex", "quit", - NULL); + (char *) NULL); } else { waitpid(pid, NULL, 0); } From e739de790921e6abbc8c70398303cacd74913f61 Mon Sep 17 00:00:00 2001 From: leejet Date: Wed, 10 Jan 2024 21:13:42 +0800 Subject: [PATCH 16/42] ggml : change GGML_MAX_NAME at compile time (ggml/682) * change GGML_MAX_NAME to 128 * allow controlling the value of GGML_MAX_NAME through external macro definitions --- ggml.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/ggml.h b/ggml.h index c55e598b4..b6cc85952 100644 --- a/ggml.h +++ b/ggml.h @@ -218,7 +218,9 @@ #define GGML_MAX_PARAMS 2048 #define GGML_MAX_CONTEXTS 64 #define GGML_MAX_SRC 10 +#ifndef GGML_MAX_NAME #define GGML_MAX_NAME 64 +#endif #define GGML_MAX_OP_PARAMS 64 #define GGML_DEFAULT_N_THREADS 4 #define GGML_DEFAULT_GRAPH_SIZE 2048 From 5362e43962e84d61e20b91f34991d7ccaef4a7d5 Mon Sep 17 00:00:00 2001 From: Jack Mousseau Date: Wed, 10 Jan 2024 06:19:19 -0800 Subject: [PATCH 17/42] metal : wrap each operation in debug group (ggml/690) --- ggml-metal.m | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/ggml-metal.m b/ggml-metal.m index 6c2a8d04e..161906824 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -1067,6 +1067,8 @@ bool ggml_metal_graph_compute( GGML_ASSERT(!"unsupported op"); } + [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(dst)]]; + 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; @@ -2423,6 +2425,8 @@ bool ggml_metal_graph_compute( GGML_ASSERT(false); } } + + [encoder popDebugGroup]; } if (encoder != nil) { From f85a973aa139ae6f37e8b8e1966f1d278b5e0372 Mon Sep 17 00:00:00 2001 From: Timothy Cronin <40186632+4imothy@users.noreply.github.com> Date: Thu, 11 Jan 2024 02:27:48 -0500 Subject: [PATCH 18/42] ggml : remove ggml_cpy_inplace and ggml_cont_inplace (ggml/693) --- ggml.c | 30 ++++++++---------------------- ggml.h | 11 ----------- 2 files changed, 8 insertions(+), 33 deletions(-) diff --git a/ggml.c b/ggml.c index 4a0ec4c44..9c42a45e3 100644 --- a/ggml.c +++ b/ggml.c @@ -4311,13 +4311,13 @@ struct ggml_tensor * ggml_set_2d_inplace( static struct ggml_tensor * ggml_cpy_impl( struct ggml_context * ctx, struct ggml_tensor * a, - struct ggml_tensor * b, - bool inplace) { + struct ggml_tensor * b) { GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b)); bool is_node = false; - if (!inplace && (a->grad || b->grad)) { + if (a->grad || b->grad) { + // inplace is false and either one have a grad is_node = true; } @@ -4341,29 +4341,21 @@ struct ggml_tensor * ggml_cpy( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) { - return ggml_cpy_impl(ctx, a, b, false); -} - -struct ggml_tensor * ggml_cpy_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { - return ggml_cpy_impl(ctx, a, b, true); + return ggml_cpy_impl(ctx, a, b); } // ggml_cont static struct ggml_tensor * ggml_cont_impl( struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { + struct ggml_tensor * a) { bool is_node = false; - if (!inplace && a->grad) { + if (a->grad) { is_node = true; } - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); ggml_format_name(result, "%s (cont)", a->name); result->op = GGML_OP_CONT; @@ -4376,13 +4368,7 @@ static struct ggml_tensor * ggml_cont_impl( struct ggml_tensor * ggml_cont( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_cont_impl(ctx, a, false); -} - -struct ggml_tensor * ggml_cont_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a) { - return ggml_cont_impl(ctx, a, true); + return ggml_cont_impl(ctx, a); } // make contiguous, with new shape diff --git a/ggml.h b/ggml.h index b6cc85952..127dcef1d 100644 --- a/ggml.h +++ b/ggml.h @@ -1163,22 +1163,11 @@ extern "C" { struct ggml_tensor * a, struct ggml_tensor * b); - // a -> b, in-place, return view(b) - GGML_API struct ggml_tensor * ggml_cpy_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - // make contiguous GGML_API struct ggml_tensor * ggml_cont( struct ggml_context * ctx, struct ggml_tensor * a); - // make contiguous, in-place - GGML_API struct ggml_tensor * ggml_cont_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a); - // make contiguous, with new shape GGML_API struct ggml_tensor * ggml_cont_1d( struct ggml_context * ctx, From 3267c2abc72e34608224408ace3c048831050f97 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 11 Jan 2024 09:34:59 +0200 Subject: [PATCH 19/42] metal : fix deprecation warning (ggml/690) --- ggml-metal.m | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml-metal.m b/ggml-metal.m index 161906824..82d68cd1b 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -1067,7 +1067,7 @@ bool ggml_metal_graph_compute( GGML_ASSERT(!"unsupported op"); } - [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(dst)]]; + [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; From 64802ec00d6383784a9dacf616095eaced16c3c3 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 11 Jan 2024 09:39:08 +0200 Subject: [PATCH 20/42] sync : ggml --- scripts/sync-ggml.last | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index fe7f3202f..3e2c579d5 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -f96711108d55bdbbd277e6be07204dce6a94fb93 +979cc23b345006504cfc1f67c0fdf627805e3319 From 2a7c94db5fb67b2f8882d2d16a11bf5d8d12d397 Mon Sep 17 00:00:00 2001 From: Paul Tsochantaris Date: Thu, 11 Jan 2024 14:31:52 +0000 Subject: [PATCH 21/42] metal : put encoder debug group behind a define (#4873) --- ggml-metal.m | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/ggml-metal.m b/ggml-metal.m index 82d68cd1b..9698e5a79 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -1067,7 +1067,9 @@ bool ggml_metal_graph_compute( GGML_ASSERT(!"unsupported op"); } +#ifndef GGML_METAL_NDEBUG [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(dst) encoding:NSUTF8StringEncoding]]; +#endif const int64_t ne00 = src0 ? src0->ne[0] : 0; const int64_t ne01 = src0 ? src0->ne[1] : 0; @@ -2426,7 +2428,9 @@ bool ggml_metal_graph_compute( } } +#ifndef GGML_METAL_NDEBUG [encoder popDebugGroup]; +#endif } if (encoder != nil) { From 2f043328e3116724d15b915b5c6078e2df860a69 Mon Sep 17 00:00:00 2001 From: Isaac McFadyen Date: Thu, 11 Jan 2024 09:33:26 -0500 Subject: [PATCH 22/42] server : fix typo in model name (#4876) --- examples/server/server.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 4a0714997..860e4e9ae 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2515,7 +2515,7 @@ json oaicompat_completion_params_parse( // // https://platform.openai.com/docs/api-reference/chat/create llama_sampling_params default_sparams; - llama_params["model"] = json_value(body, "model", std::string("uknown")); + llama_params["model"] = json_value(body, "model", std::string("unknown")); llama_params["prompt"] = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt' llama_params["cache_prompt"] = json_value(body, "cache_prompt", false); llama_params["temperature"] = json_value(body, "temperature", 0.0); From 43f76bf1c362c067fce46bb8dcda0b64af8a9533 Mon Sep 17 00:00:00 2001 From: pudepiedj Date: Thu, 11 Jan 2024 16:14:52 +0000 Subject: [PATCH 23/42] main : print total token count and tokens consumed so far (#4874) * Token count changes * Add show token count * Updating before PR * Two requested changes * Move param def posn --- common/common.cpp | 8 ++++++++ common/common.h | 2 +- examples/main/main.cpp | 6 +++++- llama.cpp | 2 +- 4 files changed, 15 insertions(+), 3 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 4e89fe516..bfcd6d4df 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -630,6 +630,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.ppl_stride = std::stoi(argv[i]); + } else if (arg == "-stc" || arg == "--show_token_count") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.token_interval = std::stoi(argv[i]); } else if (arg == "--ppl-output-type") { if (++i >= argc) { invalid_param = true; @@ -944,6 +950,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" advanced option to override model metadata by key. may be specified multiple times.\n"); printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n"); + printf(" -stc N --show_token_count N\n"); + printf(" show consumed tokens every N tokens\n"); printf("\n"); #ifndef LOG_DISABLE_LOGS log_print_usage(); diff --git a/common/common.h b/common/common.h index e2bbfc258..a295e88b0 100644 --- a/common/common.h +++ b/common/common.h @@ -64,6 +64,7 @@ struct gpt_params { int32_t n_beams = 0; // if non-zero then use beam search of given width. int32_t grp_attn_n = 1; // group-attention factor int32_t grp_attn_w = 512; // group-attention width + int32_t token_interval = 512; // show token count every 512 tokens float rope_freq_base = 0.0f; // RoPE base frequency float rope_freq_scale = 0.0f; // RoPE frequency scaling factor float yarn_ext_factor = -1.0f; // YaRN extrapolation mix factor @@ -242,4 +243,3 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80); // Dump the KV cache view showing individual sequences in each cell (long output). void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40); - diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 5ea67051f..1f35febbd 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -500,7 +500,7 @@ int main(int argc, char ** argv) { while ((n_remain != 0 && !is_antiprompt) || params.interactive) { // predict if (!embd.empty()) { - // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via + // Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via // --prompt or --file which uses the same value. int max_embd_size = n_ctx - 4; @@ -650,6 +650,10 @@ int main(int argc, char ** argv) { n_past += n_eval; LOG("n_past = %d\n", n_past); + // Display total tokens alongside total time + if (n_past % params.token_interval == 0) { + printf("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx); + } } if (!embd.empty() && !path_session.empty()) { diff --git a/llama.cpp b/llama.cpp index e1f1932ba..aaadfa444 100644 --- a/llama.cpp +++ b/llama.cpp @@ -10921,7 +10921,7 @@ void llama_print_timings(struct llama_context * ctx) { __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval); LLAMA_LOG_INFO("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval); - LLAMA_LOG_INFO("%s: total time = %10.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms)); + LLAMA_LOG_INFO("%s: total time = %10.2f ms / %5d tokens\n", __func__, (timings.t_end_ms - timings.t_start_ms), (timings.n_p_eval + timings.n_eval)); } void llama_reset_timings(struct llama_context * ctx) { From d8d90aa343c22fe01429d3540e47ded87e9dcb9d Mon Sep 17 00:00:00 2001 From: Someone Date: Thu, 11 Jan 2024 17:22:34 +0000 Subject: [PATCH 24/42] ci: nix-flake-update: new token with pr permissions (#4879) * ci: nix-flake-update: new token with pr permissions --------- Co-authored-by: Georgi Gerganov --- .github/workflows/nix-flake-update.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/nix-flake-update.yml b/.github/workflows/nix-flake-update.yml index fa9360841..3a6a96e26 100644 --- a/.github/workflows/nix-flake-update.yml +++ b/.github/workflows/nix-flake-update.yml @@ -19,4 +19,4 @@ jobs: pr-labels: | nix pr-reviewers: philiptaron,SomeoneSerge - token: ${{ secrets.GITHUB_TOKEN }} + token: ${{ secrets.FLAKE_TOKEN }} From eab67950068e4b125007d027232c47d2a5831cd0 Mon Sep 17 00:00:00 2001 From: Behnam M <58621210+ibehnam@users.noreply.github.com> Date: Thu, 11 Jan 2024 12:41:39 -0500 Subject: [PATCH 25/42] server : add `LOG_INFO` when model is successfully loaded (#4881) * added /health endpoint to the server * added comments on the additional /health endpoint * Better handling of server state When the model is being loaded, the server state is `LOADING_MODEL`. If model-loading fails, the server state becomes `ERROR`, otherwise it becomes `READY`. The `/health` endpoint provides more granular messages now according to the server_state value. * initialized server_state * fixed a typo * starting http server before initializing the model * Update server.cpp * Update server.cpp * fixes * fixes * fixes * made ServerState atomic and turned two-line spaces into one-line * updated `server` readme to document the `/health` endpoint too * used LOG_INFO after successful model loading --- examples/server/server.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 860e4e9ae..51a4b689f 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2906,6 +2906,7 @@ int main(int argc, char **argv) } else { llama.initialize(); state.store(SERVER_STATE_READY); + LOG_INFO("model loaded", {}); } // Middleware for API key validation From 27379455c38cb13f24de92dbd6fcdd04eeb1b9d9 Mon Sep 17 00:00:00 2001 From: Michael Coppola Date: Thu, 11 Jan 2024 12:51:17 -0500 Subject: [PATCH 26/42] server : support for multiple api keys (#4864) * server: added support for multiple api keys, added loading api keys from file * minor: fix whitespace * added file error handling to --api-key-file, changed code to better reflect current style * server: update README.md for --api-key-file --------- Co-authored-by: Michael Coppola --- examples/server/README.md | 3 ++- examples/server/server.cpp | 36 ++++++++++++++++++++++++++++++------ 2 files changed, 32 insertions(+), 7 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index dc27e72b9..fd3034b99 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -23,7 +23,8 @@ Command line options: - `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`. - `--port`: Set the port to listen. Default: `8080`. - `--path`: path from which to serve static files (default examples/server/public) -- `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. +- `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys. +- `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s. - `--embedding`: Enable embedding extraction, Default: disabled. - `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1) - `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 51a4b689f..345004fa1 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -39,7 +39,7 @@ using json = nlohmann::json; struct server_params { std::string hostname = "127.0.0.1"; - std::string api_key; + std::vector api_keys; std::string public_path = "examples/server/public"; int32_t port = 8080; int32_t read_timeout = 600; @@ -2021,6 +2021,7 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf(" --port PORT port to listen (default (default: %d)\n", sparams.port); printf(" --path PUBLIC_PATH path from which to serve static files (default %s)\n", sparams.public_path.c_str()); printf(" --api-key API_KEY optional api key to enhance server security. If set, requests must include this key for access.\n"); + printf(" --api-key-file FNAME path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n"); printf(" -to N, --timeout N server read/write timeout in seconds (default: %d)\n", sparams.read_timeout); printf(" --embedding enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled"); printf(" -np N, --parallel N number of slots for process requests (default: %d)\n", params.n_parallel); @@ -2081,7 +2082,28 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, invalid_param = true; break; } - sparams.api_key = argv[i]; + sparams.api_keys.push_back(argv[i]); + } + else if (arg == "--api-key-file") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + std::ifstream key_file(argv[i]); + if (!key_file) { + fprintf(stderr, "error: failed to open file '%s'\n", argv[i]); + invalid_param = true; + break; + } + std::string key; + while (std::getline(key_file, key)) { + if (key.size() > 0) { + sparams.api_keys.push_back(key); + } + } + key_file.close(); } else if (arg == "--timeout" || arg == "-to") { @@ -2881,8 +2903,10 @@ int main(int argc, char **argv) log_data["hostname"] = sparams.hostname; log_data["port"] = std::to_string(sparams.port); - if (!sparams.api_key.empty()) { - log_data["api_key"] = "api_key: ****" + sparams.api_key.substr(sparams.api_key.length() - 4); + if (sparams.api_keys.size() == 1) { + log_data["api_key"] = "api_key: ****" + sparams.api_keys[0].substr(sparams.api_keys[0].length() - 4); + } else if (sparams.api_keys.size() > 1) { + log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded"; } LOG_INFO("HTTP server listening", log_data); @@ -2912,7 +2936,7 @@ int main(int argc, char **argv) // Middleware for API key validation auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool { // If API key is not set, skip validation - if (sparams.api_key.empty()) { + if (sparams.api_keys.empty()) { return true; } @@ -2921,7 +2945,7 @@ int main(int argc, char **argv) std::string prefix = "Bearer "; if (auth_header.substr(0, prefix.size()) == prefix) { std::string received_api_key = auth_header.substr(prefix.size()); - if (received_api_key == sparams.api_key) { + if (std::find(sparams.api_keys.begin(), sparams.api_keys.end(), received_api_key) != sparams.api_keys.end()) { return true; // API key is valid } } From 4330bd83feb39683de4bd7a34cfcf672ff8ac3e4 Mon Sep 17 00:00:00 2001 From: Laura Date: Thu, 11 Jan 2024 19:02:48 +0100 Subject: [PATCH 27/42] server : implement credentialed CORS (#4514) * Implement credentialed CORS according to MDN * Fix syntax error * Move validate_api_key up so it is defined before its first usage --- examples/server/server.cpp | 26 ++++++++++++++++++++------ 1 file changed, 20 insertions(+), 6 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 345004fa1..031824e14 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2822,9 +2822,15 @@ int main(int argc, char **argv) std::atomic state{SERVER_STATE_LOADING_MODEL}; - svr.set_default_headers({{"Server", "llama.cpp"}, - {"Access-Control-Allow-Origin", "*"}, - {"Access-Control-Allow-Headers", "content-type"}}); + svr.set_default_headers({{"Server", "llama.cpp"}}); + + // CORS preflight + svr.Options(R"(.*)", [](const httplib::Request &req, httplib::Response &res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + res.set_header("Access-Control-Allow-Credentials", "true"); + res.set_header("Access-Control-Allow-Methods", "POST"); + res.set_header("Access-Control-Allow-Headers", "*"); + }); svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) { server_state current_state = state.load(); @@ -2987,9 +2993,9 @@ int main(int argc, char **argv) return false; }); - svr.Get("/props", [&llama](const httplib::Request & /*req*/, httplib::Response &res) + svr.Get("/props", [&llama](const httplib::Request & req, httplib::Response &res) { - res.set_header("Access-Control-Allow-Origin", "*"); + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); json data = { { "user_name", llama.name_user.c_str() }, { "assistant_name", llama.name_assistant.c_str() } @@ -2999,6 +3005,7 @@ int main(int argc, char **argv) svr.Post("/completion", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!validate_api_key(req, res)) { return; } @@ -3066,8 +3073,9 @@ int main(int argc, char **argv) } }); - svr.Get("/v1/models", [¶ms](const httplib::Request&, httplib::Response& res) + svr.Get("/v1/models", [¶ms](const httplib::Request& req, httplib::Response& res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); std::time_t t = std::time(0); json models = { @@ -3085,9 +3093,11 @@ int main(int argc, char **argv) res.set_content(models.dump(), "application/json; charset=utf-8"); }); + // TODO: add mount point without "/v1" prefix -- how? svr.Post("/v1/chat/completions", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!validate_api_key(req, res)) { return; } @@ -3161,6 +3171,7 @@ int main(int argc, char **argv) svr.Post("/infill", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!validate_api_key(req, res)) { return; } @@ -3233,6 +3244,7 @@ int main(int argc, char **argv) svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); const json body = json::parse(req.body); std::vector tokens; if (body.count("content") != 0) @@ -3245,6 +3257,7 @@ int main(int argc, char **argv) svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); const json body = json::parse(req.body); std::string content; if (body.count("tokens") != 0) @@ -3259,6 +3272,7 @@ int main(int argc, char **argv) svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); const json body = json::parse(req.body); json prompt; if (body.count("content") != 0) From 3ba5b8ca8e6181a5c712c5b77595a29f1d3e2b97 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 11 Jan 2024 21:31:31 +0200 Subject: [PATCH 28/42] swift : pin ggml commit + remove ggml.h from spm-headers (#4878) ggml-ci --- Package.swift | 2 +- spm-headers/ggml.h | 1 - 2 files changed, 1 insertion(+), 2 deletions(-) delete mode 120000 spm-headers/ggml.h diff --git a/Package.swift b/Package.swift index 583e2e276..59191da45 100644 --- a/Package.swift +++ b/Package.swift @@ -14,7 +14,7 @@ let package = Package( .library(name: "llama", targets: ["llama"]), ], dependencies: [ - .package(url: "https://github.com/ggerganov/ggml.git", .branch("master")) + .package(url: "https://github.com/ggerganov/ggml.git", .revision("979cc23b345006504cfc1f67c0fdf627805e3319")) ], targets: [ .target( diff --git a/spm-headers/ggml.h b/spm-headers/ggml.h deleted file mode 120000 index 39215298f..000000000 --- a/spm-headers/ggml.h +++ /dev/null @@ -1 +0,0 @@ -../ggml.h \ No newline at end of file From 49662cbed3e95f5976c070b85b9fd53fd577038d Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Thu, 11 Jan 2024 20:39:39 +0100 Subject: [PATCH 29/42] ggml : SOTA 2-bit quants (add IQ2_XS) (#4856) * iq2_xs: basics * iq2_xs: this should have been in the basics * iq2_xs: CUDA and scalar CPU works * iq2_xs: WIP Metal * iq2_xs: Metal now works * iq2_xs: working, but dog slow, ARM_NEON dot product * iq2_xs: better ARM_NEON dot product We are now at 19.5 t/s for TG-128 and 61 t/s for PP-512 when running on the CPU. * iq2_xs: AVX2 dot product - 19.5 t/s * iq2_xs: faster AVX2 dit product 21.4 t/s for TG-128, 59.2 t/s for PP-512. The latter is 2x compared to the previous version. * iq2_xs: had forgotten to delete iq2-data.h * Add llama enum for IQ2_XS --------- Co-authored-by: Iwan Kawrakow --- ggml-cuda.cu | 232 +++++++++++++++++++++- ggml-metal.m | 42 +++- ggml-metal.metal | 378 +++++++++++++++++++++++++++++++++++- ggml-quants.c | 360 +++++++++++++++++++++++++++++++++- ggml-quants.h | 12 ++ ggml.c | 30 ++- ggml.h | 3 + llama.cpp | 3 + llama.h | 1 + tests/test-quantize-fns.cpp | 5 +- 10 files changed, 1038 insertions(+), 28 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 900f7ba4a..dd19699f6 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -486,6 +486,15 @@ typedef struct { } block_iq2_xxs; static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding"); +#define QR2_XS 8 +#define QI2_XS (QK_K / (4*QR2_XS)) +typedef struct { + half d; + uint16_t qs[QK_K/8]; + uint8_t scales[QK_K/32]; +} block_iq2_xs; +static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding"); + #define WARP_SIZE 32 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses @@ -1328,7 +1337,7 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t #endif } -static const __device__ uint64_t kgrid_iq2xxs[256] = { +static const __device__ uint64_t iq2xxs_grid[256] = { 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, @@ -1395,6 +1404,137 @@ static const __device__ uint64_t kgrid_iq2xxs[256] = { 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, }; +static const __device__ uint64_t iq2xs_grid[512] = { + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, + 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, + 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, + 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, + 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, + 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819, + 0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819, + 0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, + 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b, + 0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b, + 0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, + 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908, + 0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919, + 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, + 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919, + 0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908, + 0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, + 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, + 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08, + 0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, + 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808, + 0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819, + 0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, + 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819, + 0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808, + 0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, + 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819, + 0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819, + 0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, + 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908, + 0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19, + 0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, + 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b, + 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919, + 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, + 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819, + 0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819, + 0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, + 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908, + 0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808, + 0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, + 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808, + 0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, + 0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, + 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808, + 0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908, + 0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, + 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, + 0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b, + 0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, + 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, + 0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908, + 0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, + 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908, + 0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919, + 0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, + 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19, + 0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b, + 0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, + 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808, + 0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08, + 0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, + 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908, + 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b, + 0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, + 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, + 0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808, + 0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, + 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08, + 0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819, + 0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, + 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808, + 0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808, + 0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, + 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819, + 0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908, + 0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, + 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b, + 0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908, + 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, + 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908, + 0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808, + 0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, + 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819, + 0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819, + 0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, + 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b, + 0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819, + 0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, + 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08, + 0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808, + 0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, + 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919, + 0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, + 0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, + 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b, + 0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808, + 0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, + 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b, + 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, + 0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, + 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808, + 0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819, + 0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, + 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808, + 0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, + 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, + 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19, + 0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08, + 0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, + 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08, + 0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08, + 0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, + 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908, + 0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b, + 0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, + 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808, + 0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b, + 0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, + 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808, + 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19, + 0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, + 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808, + 0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, + 0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, + 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b, + 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, +}; + static const __device__ uint8_t ksigns_iq2xs[128] = { 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, @@ -1439,7 +1579,7 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds dst_t * y = yy + i*QK_K + 32*ib + 8*il; const uint16_t * q2 = x[i].qs + 4*ib; const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid = (const uint8_t *)(kgrid_iq2xxs + aux8[il]); + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]); const uint32_t aux32 = q2[2] | (q2[3] << 16); const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f; const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; @@ -1450,6 +1590,28 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds } +template +static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_iq2_xs * x = (const block_iq2_xs *) vx; + + const int tid = threadIdx.x; +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint16_t * q2 = x[i].qs + 4*ib; + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511)); + const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; + const uint8_t signs = ksigns_iq2xs[q2[il] >> 9]; + for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); +#else + assert(false); +#endif + +} + static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); @@ -3996,7 +4158,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( uint32_t aux32 = q2[2] | (q2[3] << 16); int sumi = 0; for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(kgrid_iq2xxs + aux8[l]); + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); const uint8_t signs = ksigns_iq2xs[aux32 & 127]; for (int j = 0; j < 8; ++j) { sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); @@ -4012,8 +4174,8 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( const int il = iqs%2; const uint16_t * q2 = bq2->qs + 4*ib32; const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid1 = (const uint8_t *)(kgrid_iq2xxs + aux8[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(kgrid_iq2xxs + aux8[2*il+1]); + const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); + const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); const uint32_t aux32 = q2[2] | (q2[3] << 16); const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * (float)bq8_1[ib32].ds.x * 0.25f; const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; @@ -4032,6 +4194,42 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( #endif } +static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { +#if QK_K == 256 + const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq; + + const int ib32 = iqs; + const uint16_t * q2 = bq2->qs + 4*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + const uint8_t ls1 = bq2->scales[ib32] & 0xf; + const uint8_t ls2 = bq2->scales[ib32] >> 4; + int sumi1 = 0; + for (int l = 0; l < 2; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); + const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; + for (int j = 0; j < 8; ++j) { + sumi1 += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + int sumi2 = 0; + for (int l = 2; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); + const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; + for (int j = 0; j < 8; ++j) { + sumi2 += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + const float d = (float)bq2->d * (float)bq8_1[ib32].ds.x * 0.25f; + return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); +#else + assert(false); + return 0.f; +#endif +} + template static __device__ __forceinline__ void mul_mat_q( @@ -6035,6 +6233,12 @@ static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, dequantize_block_iq2_xxs<<>>(vx, y); } +template +static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; + dequantize_block_iq2_xs<<>>(vx, y); +} + template static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) { const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE; @@ -6065,6 +6269,8 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) { return dequantize_row_q6_K_cuda; case GGML_TYPE_IQ2_XXS: return dequantize_row_iq2_xxs_cuda; + case GGML_TYPE_IQ2_XS: + return dequantize_row_iq2_xs_cuda; case GGML_TYPE_F32: return convert_unary_cuda; default: @@ -6096,6 +6302,8 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) { return dequantize_row_q6_K_cuda; case GGML_TYPE_IQ2_XXS: return dequantize_row_iq2_xxs_cuda; + case GGML_TYPE_IQ2_XS: + return dequantize_row_iq2_xs_cuda; case GGML_TYPE_F16: return convert_unary_cuda; default: @@ -6299,6 +6507,15 @@ static void mul_mat_vec_iq2_xxs_q8_1_cuda(const void * vx, const void * vy, floa <<>>(vx, vy, dst, ncols, nrows); } +static void mul_mat_vec_iq2_xs_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); + mul_mat_vec_q + <<>>(vx, vy, dst, ncols, nrows); +} + static void ggml_mul_mat_q4_0_q8_1_cuda( const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { @@ -7871,6 +8088,7 @@ static int64_t get_row_rounding(ggml_type type) { case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: return max_compute_capability >= CC_RDNA2 ? 128 : 64; default: GGML_ASSERT(false); @@ -7892,6 +8110,7 @@ static int64_t get_row_rounding(ggml_type type) { case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: return max_compute_capability >= CC_VOLTA ? 128 : 64; case GGML_TYPE_Q6_K: return 64; @@ -7945,6 +8164,9 @@ static void ggml_cuda_op_mul_mat_vec_q( case GGML_TYPE_IQ2_XXS: mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); break; + case GGML_TYPE_IQ2_XS: + mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; default: GGML_ASSERT(false); break; diff --git a/ggml-metal.m b/ggml-metal.m index 9698e5a79..6e5594432 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -89,6 +89,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(get_rows_q6_K); GGML_METAL_DECL_KERNEL(get_rows_i32); GGML_METAL_DECL_KERNEL(get_rows_iq2_xxs); + GGML_METAL_DECL_KERNEL(get_rows_iq2_xs); GGML_METAL_DECL_KERNEL(rms_norm); GGML_METAL_DECL_KERNEL(group_norm); GGML_METAL_DECL_KERNEL(norm); @@ -108,6 +109,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32); GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32); GGML_METAL_DECL_KERNEL(mul_mv_iq2_xxs_f32); + GGML_METAL_DECL_KERNEL(mul_mv_iq2_xs_f32); GGML_METAL_DECL_KERNEL(mul_mv_id_f32_f32); //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f16); GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32); @@ -124,6 +126,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(mul_mv_id_q5_K_f32); GGML_METAL_DECL_KERNEL(mul_mv_id_q6_K_f32); GGML_METAL_DECL_KERNEL(mul_mv_id_iq2_xxs_f32); + GGML_METAL_DECL_KERNEL(mul_mv_id_iq2_xs_f32); GGML_METAL_DECL_KERNEL(mul_mm_f32_f32); GGML_METAL_DECL_KERNEL(mul_mm_f16_f32); GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32); @@ -137,6 +140,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32); GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32); GGML_METAL_DECL_KERNEL(mul_mm_iq2_xxs_f32); + GGML_METAL_DECL_KERNEL(mul_mm_iq2_xs_f32); GGML_METAL_DECL_KERNEL(mul_mm_id_f32_f32); GGML_METAL_DECL_KERNEL(mul_mm_id_f16_f32); GGML_METAL_DECL_KERNEL(mul_mm_id_q4_0_f32); @@ -150,6 +154,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(mul_mm_id_q5_K_f32); GGML_METAL_DECL_KERNEL(mul_mm_id_q6_K_f32); GGML_METAL_DECL_KERNEL(mul_mm_id_iq2_xxs_f32); + GGML_METAL_DECL_KERNEL(mul_mm_id_iq2_xs_f32); GGML_METAL_DECL_KERNEL(rope_f32); GGML_METAL_DECL_KERNEL(rope_f16); GGML_METAL_DECL_KERNEL(alibi_f32); @@ -385,6 +390,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(get_rows_q6_K); GGML_METAL_ADD_KERNEL(get_rows_i32); GGML_METAL_ADD_KERNEL(get_rows_iq2_xxs); + GGML_METAL_ADD_KERNEL(get_rows_iq2_xs); GGML_METAL_ADD_KERNEL(rms_norm); GGML_METAL_ADD_KERNEL(group_norm); GGML_METAL_ADD_KERNEL(norm); @@ -404,6 +410,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32); GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32); GGML_METAL_ADD_KERNEL(mul_mv_iq2_xxs_f32); + GGML_METAL_ADD_KERNEL(mul_mv_iq2_xs_f32); GGML_METAL_ADD_KERNEL(mul_mv_id_f32_f32); //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f16); GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32); @@ -420,6 +427,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(mul_mv_id_q5_K_f32); GGML_METAL_ADD_KERNEL(mul_mv_id_q6_K_f32); GGML_METAL_ADD_KERNEL(mul_mv_id_iq2_xxs_f32); + GGML_METAL_ADD_KERNEL(mul_mv_id_iq2_xs_f32); if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) { GGML_METAL_ADD_KERNEL(mul_mm_f32_f32); GGML_METAL_ADD_KERNEL(mul_mm_f16_f32); @@ -434,6 +442,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32); GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32); GGML_METAL_ADD_KERNEL(mul_mm_iq2_xxs_f32); + GGML_METAL_ADD_KERNEL(mul_mm_iq2_xs_f32); GGML_METAL_ADD_KERNEL(mul_mm_id_f32_f32); GGML_METAL_ADD_KERNEL(mul_mm_id_f16_f32); GGML_METAL_ADD_KERNEL(mul_mm_id_q4_0_f32); @@ -447,6 +456,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(mul_mm_id_q5_K_f32); GGML_METAL_ADD_KERNEL(mul_mm_id_q6_K_f32); GGML_METAL_ADD_KERNEL(mul_mm_id_iq2_xxs_f32); + GGML_METAL_ADD_KERNEL(mul_mm_id_iq2_xs_f32); } GGML_METAL_ADD_KERNEL(rope_f32); GGML_METAL_ADD_KERNEL(rope_f16); @@ -513,6 +523,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(get_rows_q6_K); GGML_METAL_DEL_KERNEL(get_rows_i32); GGML_METAL_DEL_KERNEL(get_rows_iq2_xxs); + GGML_METAL_DEL_KERNEL(get_rows_iq2_xs); GGML_METAL_DEL_KERNEL(rms_norm); GGML_METAL_DEL_KERNEL(group_norm); GGML_METAL_DEL_KERNEL(norm); @@ -532,6 +543,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32); GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32); GGML_METAL_DEL_KERNEL(mul_mv_iq2_xxs_f32); + GGML_METAL_DEL_KERNEL(mul_mv_iq2_xs_f32); GGML_METAL_DEL_KERNEL(mul_mv_id_f32_f32); //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f16); GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32); @@ -548,6 +560,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(mul_mv_id_q5_K_f32); GGML_METAL_DEL_KERNEL(mul_mv_id_q6_K_f32); GGML_METAL_DEL_KERNEL(mul_mv_id_iq2_xxs_f32); + GGML_METAL_DEL_KERNEL(mul_mv_id_iq2_xs_f32); if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) { GGML_METAL_DEL_KERNEL(mul_mm_f32_f32); GGML_METAL_DEL_KERNEL(mul_mm_f16_f32); @@ -562,6 +575,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32); GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32); GGML_METAL_DEL_KERNEL(mul_mm_iq2_xxs_f32); + GGML_METAL_DEL_KERNEL(mul_mm_iq2_xs_f32); GGML_METAL_DEL_KERNEL(mul_mm_id_f32_f32); GGML_METAL_DEL_KERNEL(mul_mm_id_f16_f32); GGML_METAL_DEL_KERNEL(mul_mm_id_q4_0_f32); @@ -575,6 +589,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(mul_mm_id_q5_K_f32); GGML_METAL_DEL_KERNEL(mul_mm_id_q6_K_f32); GGML_METAL_DEL_KERNEL(mul_mm_id_iq2_xxs_f32); + GGML_METAL_DEL_KERNEL(mul_mm_id_iq2_xs_f32); } GGML_METAL_DEL_KERNEL(rope_f32); GGML_METAL_DEL_KERNEL(rope_f16); @@ -1561,6 +1576,7 @@ bool ggml_metal_graph_compute( case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break; case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break; case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_mul_mm_iq2_xxs_f32]; break; + case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_mul_mm_iq2_xs_f32]; break; default: GGML_ASSERT(false && "MUL MAT-MAT not implemented"); } [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; @@ -1679,6 +1695,12 @@ bool ggml_metal_graph_compute( nth1 = 16; [encoder setComputePipelineState:ctx->pipeline_mul_mv_iq2_xxs_f32]; } break; + case GGML_TYPE_IQ2_XS: + { + nth0 = 4; + nth1 = 16; + [encoder setComputePipelineState:ctx->pipeline_mul_mv_iq2_xs_f32]; + } break; default: { GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t); @@ -1712,12 +1734,12 @@ bool ggml_metal_graph_compute( 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_IQ2_XXS || src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } - else if (src0t == GGML_TYPE_IQ2_XXS) { - [encoder setThreadgroupMemoryLength:(256*8+128) atIndex:0]; + 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, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else if (src0t == GGML_TYPE_Q4_K) { @@ -1810,6 +1832,7 @@ bool ggml_metal_graph_compute( case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_K_f32]; break; case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q6_K_f32]; break; case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_iq2_xxs_f32]; break; + case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_iq2_xs_f32]; break; default: GGML_ASSERT(false && "MUL_MAT_ID not implemented"); } [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; @@ -1931,6 +1954,12 @@ bool ggml_metal_graph_compute( nth1 = 16; [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_iq2_xxs_f32]; } break; + case GGML_TYPE_IQ2_XS: + { + nth0 = 4; + nth1 = 16; + [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_iq2_xs_f32]; + } break; default: { GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t); @@ -1980,12 +2009,12 @@ bool ggml_metal_graph_compute( if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 || src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 || - //src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } - else if (src2t == GGML_TYPE_IQ2_XXS) { - [encoder setThreadgroupMemoryLength:(256*8+128) atIndex:0]; + else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) { + const int mem_size = src2t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128; + [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else if (src2t == GGML_TYPE_Q4_K) { @@ -2026,6 +2055,7 @@ bool ggml_metal_graph_compute( case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_K]; break; case GGML_TYPE_I32: [encoder setComputePipelineState:ctx->pipeline_get_rows_i32]; break; case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_get_rows_iq2_xxs]; break; + case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_get_rows_iq2_xs]; break; default: GGML_ASSERT(false && "not implemented"); } diff --git a/ggml-metal.metal b/ggml-metal.metal index 229efb8b6..029578dc5 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -2452,6 +2452,13 @@ typedef struct { } block_iq2_xxs; // 66 bytes / block for QK_K = 256, so 2.0625 bpw +typedef struct { + half d; + uint16_t qs[QK_K/8]; + uint8_t scales[QK_K/32]; +} block_iq2_xs; +// 74 bytes / block for QK_K = 256, so 2.3125 bpw + //====================================== dot products ========================= void kernel_mul_mv_q2_K_f32_impl( @@ -3476,7 +3483,7 @@ kernel void kernel_mul_mv_q6_K_f32( // ======================= "True" 2-bit -constexpr constant static uint64_t kgrid_iq2xxs[256] = { +constexpr constant static uint64_t iq2xxs_grid[256] = { 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, @@ -3543,6 +3550,137 @@ constexpr constant static uint64_t kgrid_iq2xxs[256] = { 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, }; +constexpr constant static uint64_t iq2xs_grid[512] = { + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, + 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, + 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, + 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, + 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, + 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819, + 0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819, + 0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, + 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b, + 0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b, + 0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, + 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908, + 0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919, + 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, + 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919, + 0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908, + 0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, + 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, + 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08, + 0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, + 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808, + 0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819, + 0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, + 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819, + 0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808, + 0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, + 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819, + 0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819, + 0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, + 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908, + 0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19, + 0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, + 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b, + 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919, + 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, + 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819, + 0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819, + 0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, + 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908, + 0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808, + 0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, + 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808, + 0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, + 0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, + 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808, + 0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908, + 0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, + 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, + 0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b, + 0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, + 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, + 0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908, + 0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, + 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908, + 0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919, + 0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, + 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19, + 0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b, + 0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, + 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808, + 0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08, + 0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, + 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908, + 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b, + 0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, + 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, + 0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808, + 0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, + 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08, + 0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819, + 0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, + 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808, + 0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808, + 0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, + 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819, + 0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908, + 0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, + 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b, + 0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908, + 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, + 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908, + 0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808, + 0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, + 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819, + 0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819, + 0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, + 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b, + 0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819, + 0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, + 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08, + 0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808, + 0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, + 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919, + 0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, + 0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, + 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b, + 0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808, + 0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, + 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b, + 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, + 0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, + 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808, + 0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819, + 0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, + 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808, + 0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, + 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, + 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19, + 0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08, + 0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, + 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08, + 0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08, + 0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, + 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908, + 0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b, + 0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, + 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808, + 0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b, + 0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, + 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808, + 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19, + 0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, + 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808, + 0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, + 0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, + 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b, + 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, +}; + constexpr constant static uint8_t ksigns_iq2xs[128] = { 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, @@ -3600,7 +3738,7 @@ void kernel_mul_mv_iq2_xxs_f32_impl( { int nval = 4; int pos = (32*sgitg + tiisg)*nval; - for (int i = 0; i < nval; ++i) values[pos + i] = kgrid_iq2xxs[pos + i]; + for (int i = 0; i < nval; ++i) values[pos + i] = iq2xxs_grid[pos + i]; nval = 2; pos = (32*sgitg + tiisg)*nval; for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i]; @@ -3689,6 +3827,149 @@ kernel void kernel_mul_mv_iq2_xxs_f32( kernel_mul_mv_iq2_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg); } +void kernel_mul_mv_iq2_xs_f32_impl( + device const void * src0, + device const float * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne10, + constant int64_t & ne12, + constant int64_t & ne0, + constant int64_t & ne1, + constant uint & r2, + constant uint & r3, + threadgroup int8_t * shared_values [[threadgroup(0)]], + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiisg[[thread_index_in_simdgroup]], + uint sgitg[[simdgroup_index_in_threadgroup]]) { + + const int nb = ne00/QK_K; + const int r0 = tgpig.x; + const int r1 = tgpig.y; + const int im = tgpig.z; + + const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int ib_row = first_row * nb; + + const uint i12 = im%ne12; + const uint i13 = im/ne12; + + const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + + device const block_iq2_xs * x = (device const block_iq2_xs *) src0 + ib_row + offset0; + device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1; + + float yl[32]; + float sumf[N_DST]={0.f}, all_sum; + + const int nb32 = nb * (QK_K / 32); + + threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values; + threadgroup uint8_t * shared_signs = (threadgroup uint8_t *)(values + 512); + { + int nval = 8; + int pos = (32*sgitg + tiisg)*nval; + for (int i = 0; i < nval; ++i) values[pos + i] = iq2xs_grid[pos + i]; + nval = 2; + pos = (32*sgitg + tiisg)*nval; + for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i]; + threadgroup_barrier(mem_flags::mem_threadgroup); + } + +#if QK_K == 256 + const int ix = tiisg; + + device const float * y4 = y + 32 * ix; + + for (int ib32 = ix; ib32 < nb32; ib32 += 32) { + + for (int i = 0; i < 32; ++i) { + yl[i] = y4[i]; + } + + const int ibl = ib32 / (QK_K / 32); + const int ib = ib32 % (QK_K / 32); + + device const block_iq2_xs * xr = x + ibl; + device const uint16_t * q2 = xr->qs + 4 * ib; + device const uint8_t * sc = xr->scales + ib; + device const half * dh = &xr->d; + + for (int row = 0; row < N_DST; row++) { + + const float db = dh[0]; + const uint8_t ls1 = sc[0] & 0xf; + const uint8_t ls2 = sc[0] >> 4; + const float d1 = db * (0.5f + ls1); + const float d2 = db * (0.5f + ls2); + + float sum1 = 0, sum2 = 0; + for (int l = 0; l < 2; ++l) { + const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511)); + const uint8_t signs = shared_signs[(q2[l] >> 9)]; + for (int j = 0; j < 8; ++j) { + sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); + } + } + for (int l = 2; l < 4; ++l) { + const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511)); + const uint8_t signs = shared_signs[(q2[l] >> 9)]; + for (int j = 0; j < 8; ++j) { + sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); + } + } + sumf[row] += d1 * sum1 + d2 * sum2; + + dh += nb*sizeof(block_iq2_xs)/2; + q2 += nb*sizeof(block_iq2_xs)/2; + sc += nb*sizeof(block_iq2_xs); + } + + y4 += 32 * 32; + } +#else + // TODO +#endif + + for (int row = 0; row < N_DST; ++row) { + all_sum = simd_sum(sumf[row]); + if (tiisg == 0) { + dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f; + } + } +} + +[[host_name("kernel_mul_mv_iq2_xs_f32")]] +kernel void kernel_mul_mv_iq2_xs_f32( + device const void * src0, + device const float * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + constant uint & r2, + constant uint & r3, + threadgroup int8_t * shared_values [[threadgroup(0)]], + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiisg[[thread_index_in_simdgroup]], + uint sgitg[[simdgroup_index_in_threadgroup]]) { + + kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg); +} + //============================= templates and their specializations ============================= // NOTE: this is not dequantizing - we are simply fitting the template @@ -3973,18 +4254,39 @@ void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x const uint32_t aux32_s = q2[2] | (q2[3] << 16); thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g; const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f; - constant uint8_t * grid = (constant uint8_t *)(kgrid_iq2xxs + aux8[2*il+0]); + constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]); uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127]; for (int i = 0; i < 8; ++i) { reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f); } - grid = (constant uint8_t *)(kgrid_iq2xxs + aux8[2*il+1]); + grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]); signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127]; for (int i = 0; i < 8; ++i) { reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f); } } +template +void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) { + // il is 0...15 for QK_K = 256 => index of block of 32 is il/2 + const float d = xb->d; + const int ib32 = il/2; + il = il%2; + // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16 + device const uint16_t * q2 = xb->qs + 4*ib32; + const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f; + constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511)); + uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9]; + for (int i = 0; i < 8; ++i) { + reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f); + } + grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511)); + signs = ksigns_iq2xs[q2[2*il+1] >> 9]; + for (int i = 0; i < 8; ++i) { + reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f); + } +} + template kernel void kernel_get_rows( device const void * src0, @@ -4525,6 +4827,7 @@ template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows; +template [[host_name("kernel_get_rows_iq2_xs")]] kernel get_rows_t kernel_get_rows; // // matrix-matrix multiplication @@ -4562,6 +4865,7 @@ template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mat_mm_t kernel_mul_mm; // // indirect matrix-matrix multiplication @@ -4611,6 +4915,7 @@ template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mu template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; // // matrix-vector multiplication @@ -5448,3 +5753,68 @@ kernel void kernel_mul_mv_id_iq2_xxs_f32( tiisg, sgitg); } + +[[host_name("kernel_mul_mv_id_iq2_xs_f32")]] +kernel void kernel_mul_mv_id_iq2_xs_f32( + device const char * ids, + device const char * src1, + device float * dst, + constant uint64_t & nbi1, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant int64_t & ne13, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + constant uint64_t & nb1, + constant uint & r2, + constant uint & r3, + constant int & idx, + device const char * src00, + device const char * src01, + device const char * src02, + device const char * src03, + device const char * src04, + device const char * src05, + device const char * src06, + device const char * src07, + threadgroup int8_t * shared_values [[threadgroup(0)]], + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiitg[[thread_index_in_threadgroup]], + uint tiisg[[thread_index_in_simdgroup]], + uint sgitg[[simdgroup_index_in_threadgroup]]) { + device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07}; + + const int64_t bid = tgpig.z/(ne12*ne13); + + tgpig.z = tgpig.z%(ne12*ne13); + + const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx]; + + kernel_mul_mv_iq2_xs_f32_impl( + src0[id], + (device const float *) (src1 + bid*nb11), + dst + bid*ne0, + ne00, + ne01, + ne02, + ne10, + ne12, + ne0, + ne1, + r2, + r3, + shared_values, + tgpig, + tiisg, + sgitg); +} diff --git a/ggml-quants.c b/ggml-quants.c index d497e6de9..a24b4b244 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -2342,15 +2342,7 @@ size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * // ====================== "True" 2-bit (de)-quantization -void quantize_row_iq2_xxs_reference(const float * restrict x, block_iq2_xxs * restrict y, int k) { - (void)x; - (void)y; - (void)k; - assert(k % QK_K == 0); - //fprintf(stderr, "=========================== %s: not implemented\n", __func__); -} - -static const uint64_t iq2xxs_grid[256] = { +static const uint64_t iq2xxs_grid[256] = { 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, @@ -2417,6 +2409,137 @@ static const uint64_t iq2xxs_grid[256] = { 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, }; +static const uint64_t iq2xs_grid[512] = { + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, + 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, + 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, + 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, + 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, + 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819, + 0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819, + 0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, + 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b, + 0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b, + 0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, + 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908, + 0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919, + 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, + 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919, + 0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908, + 0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, + 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, + 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08, + 0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, + 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808, + 0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819, + 0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, + 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819, + 0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808, + 0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, + 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819, + 0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819, + 0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, + 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908, + 0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19, + 0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, + 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b, + 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919, + 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, + 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819, + 0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819, + 0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, + 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908, + 0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808, + 0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, + 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808, + 0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, + 0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, + 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808, + 0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908, + 0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, + 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, + 0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b, + 0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, + 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, + 0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908, + 0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, + 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908, + 0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919, + 0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, + 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19, + 0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b, + 0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, + 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808, + 0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08, + 0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, + 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908, + 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b, + 0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, + 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, + 0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808, + 0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, + 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08, + 0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819, + 0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, + 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808, + 0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808, + 0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, + 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819, + 0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908, + 0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, + 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b, + 0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908, + 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, + 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908, + 0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808, + 0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, + 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819, + 0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819, + 0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, + 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b, + 0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819, + 0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, + 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08, + 0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808, + 0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, + 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919, + 0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, + 0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, + 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b, + 0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808, + 0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, + 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b, + 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, + 0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, + 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808, + 0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819, + 0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, + 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808, + 0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, + 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, + 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19, + 0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08, + 0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, + 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08, + 0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08, + 0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, + 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908, + 0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b, + 0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, + 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808, + 0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b, + 0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, + 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808, + 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19, + 0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, + 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808, + 0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, + 0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, + 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b, + 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, +}; + static const uint8_t ksigns_iq2xs[128] = { 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, @@ -2427,8 +2550,17 @@ static const uint8_t ksigns_iq2xs[128] = { 96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111, 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, }; + static const uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128}; +void quantize_row_iq2_xxs_reference(const float * restrict x, block_iq2_xxs * restrict y, int k) { + (void)x; + (void)y; + (void)k; + assert(k % QK_K == 0); + //fprintf(stderr, "=========================== %s: not implemented\n", __func__); +} + void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -2472,6 +2604,58 @@ size_t ggml_quantize_iq2_xxs(const float * src, void * dst, int n, int k, int64_ return (n/QK_K*sizeof(block_iq2_xxs)); } +// ====================== 2.3125 bpw (de)-quantization + +void quantize_row_iq2_xs_reference(const float * restrict x, block_iq2_xs * restrict y, int k) { + (void)x; + (void)y; + (void)k; + assert(k % QK_K == 0); + //fprintf(stderr, "=========================== %s: not implemented\n", __func__); +} + +void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y, int k) { + assert(k % QK_K == 0); + const int nb = k / QK_K; + + float db[2]; + + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d); + + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + db[0] = d * (0.5f + (x[i].scales[ib32] & 0xf)) * 0.25f; + db[1] = d * (0.5f + (x[i].scales[ib32] >> 4)) * 0.25f; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (x[i].qs[4*ib32 + l] & 511)); + const uint8_t signs = ksigns_iq2xs[x[i].qs[4*ib32 + l] >> 9]; + for (int j = 0; j < 8; ++j) { + y[j] = db[l/2] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); + } + y += 8; + } + } + } +} + +void quantize_row_iq2_xs(const float * restrict x, void * restrict vy, int k) { + assert(k % QK_K == 0); + block_iq2_xs * restrict y = vy; + quantize_row_iq2_xs_reference(x, y, k); +} + +size_t ggml_quantize_iq2_xs(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % QK_K == 0); + (void)hist; // TODO: collect histograms + + for (int j = 0; j < n; j += k) { + block_iq2_xs * restrict y = (block_iq2_xs *)dst + j/QK_K; + quantize_row_iq2_xs_reference(src + j, y, k); + } + return (n/QK_K*sizeof(block_iq2_xs)); +} + //===================================== Q8_K ============================================== void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k) { @@ -7357,3 +7541,161 @@ void ggml_vec_dot_iq2_xxs_q8_K(const int n, float * restrict s, const void * res *s = 0.125f * sumf; #endif } + +void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + assert(n % QK_K == 0); + + const block_iq2_xs * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if defined(__ARM_NEON) + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + int8x16x4_t q2u; + int8x16x4_t q2s; + int8x16x4_t q8b; + + int32x4x4_t scales32; + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + const uint8x8_t scales8 = vld1_u8(x[i].scales); + const uint8x8_t scales_l = vand_u8(scales8, vdup_n_u8(0xf)); + const uint8x8_t scales_h = vshr_n_u8(scales8, 4); + uint8x16_t scales = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h)); + scales = vaddq_u8(vshlq_n_u8(scales, 1), vdupq_n_u8(1)); + const uint16x8_t scales1 = vmovl_u8(vget_low_u8(scales)); + const uint16x8_t scales2 = vmovl_u8(vget_high_u8(scales)); + scales32.val[0] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales1))); + scales32.val[1] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales1))); + scales32.val[2] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales2))); + scales32.val[3] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales2))); + int32x4_t sumi = vdupq_n_s32(0); + for (int ib64 = 0; ib64 < QK_K/64; ++ib64) { + q8b = vld1q_s8_x4(q8); q8 += 64; + q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[0] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[1] & 511)))); + q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[2] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[3] & 511)))); + q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[4] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[5] & 511)))); + q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[6] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[7] & 511)))); + q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[0] >> 9))), vld1_s8((const void *)(signs64 + (q2[1] >> 9)))); + q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[2] >> 9))), vld1_s8((const void *)(signs64 + (q2[3] >> 9)))); + q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[4] >> 9))), vld1_s8((const void *)(signs64 + (q2[5] >> 9)))); + q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[6] >> 9))), vld1_s8((const void *)(signs64 + (q2[7] >> 9)))); + q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]); + q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]); + q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]); + q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]); + const int32x4_t p1 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]); + const int32x4_t p2 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[1], q8b.val[1]); + const int32x4_t p3 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]); + const int32x4_t p4 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[3], q8b.val[3]); + const int32x4_t p = vpaddq_s32(vpaddq_s32(p1, p2), vpaddq_s32(p3, p4)); + sumi = vmlaq_s32(sumi, p, scales32.val[ib64]); + q2 += 8; + } + sumf += d*vaddvq_s32(sumi); + } + *s = 0.125f * sumf; + +#elif defined(__AVX2__) + + const __m128i m4 = _mm_set1_epi8(0xf); + const __m128i m1 = _mm_set1_epi8(1); + const __m128i m511 = _mm_set1_epi16(511); + const __m128i m127 = _mm_set1_epi16(127); + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + uint64_t aux64; + + // somewhat hacky, but gives a significant boost in performance + __m128i aux_gindex, aux_sindex; + const uint16_t * gindex = (const uint16_t *)&aux_gindex; + const uint16_t * sindex = (const uint16_t *)&aux_sindex; + + __m256 accumf = _mm256_setzero_ps(); + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + memcpy(&aux64, x[i].scales, 8); + __m128i stmp = _mm_set1_epi64x(aux64); + stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4)); + const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1); + + __m256i sumi1 = _mm256_setzero_si256(); + __m256i sumi2 = _mm256_setzero_si256(); + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m128i q2_data = _mm_loadu_si128((const __m128i*)q2); q2 += 8; + aux_gindex = _mm_and_si128(q2_data, m511); + aux_sindex = _mm_and_si128(_mm_srli_epi16(q2_data, 9), m127); + const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]], iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]); + const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]], iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]); + const __m256i s2_1 = _mm256_set_epi64x(signs64[sindex[3]], signs64[sindex[2]], signs64[sindex[1]], signs64[sindex[0]]); + const __m256i s2_2 = _mm256_set_epi64x(signs64[sindex[7]], signs64[sindex[6]], signs64[sindex[5]], signs64[sindex[4]]); + const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1); + const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2); + const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); + const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); + + const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0))); + const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1))); + + sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1)); + sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2)); + } + + accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf); + + } + + *s = 0.125f * hsum_float_8(accumf); + +#else + + float sumf = 0.f; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const uint8_t * restrict sc = x[i].scales; + const int8_t * restrict q8 = y[i].qs; + int32_t bsum = 0; + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1; + const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1; + int32_t sumi = 0; + for (int l = 0; l < 2; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); + const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; + for (int j = 0; j < 8; ++j) { + sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + bsum += sumi * ls1; + sumi = 0; + for (int l = 2; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); + const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; + for (int j = 0; j < 8; ++j) { + sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + bsum += sumi * ls2; + q2 += 4; + } + sumf += d * bsum; + } + *s = 0.125f * sumf; +#endif +} diff --git a/ggml-quants.h b/ggml-quants.h index 8dd911d41..df5e7ae80 100644 --- a/ggml-quants.h +++ b/ggml-quants.h @@ -174,6 +174,14 @@ typedef struct { } block_iq2_xxs; static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding"); +// 2.3125 bpw quants +typedef struct { + ggml_fp16_t d; + uint16_t qs[QK_K/8]; + uint8_t scales[QK_K/32]; +} block_iq2_xs; +static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding"); + // Quantization void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k); void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k); @@ -189,6 +197,7 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k); void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k); void quantize_row_iq2_xxs_reference(const float * restrict x, block_iq2_xxs * restrict y, int k); +void quantize_row_iq2_xs_reference (const float * restrict x, block_iq2_xs * restrict y, int k); void quantize_row_q4_0(const float * restrict x, void * restrict y, int k); void quantize_row_q4_1(const float * restrict x, void * restrict y, int k); @@ -204,6 +213,7 @@ void quantize_row_q5_K(const float * restrict x, void * restrict y, int k); void quantize_row_q6_K(const float * restrict x, void * restrict y, int k); void quantize_row_q8_K(const float * restrict x, void * restrict y, int k); void quantize_row_iq2_xxs(const float * restrict x, void * restrict y, int k); +void quantize_row_iq2_xs (const float * restrict x, void * restrict y, int k); // Dequantization void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k); @@ -220,6 +230,7 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k); void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k); void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k); +void dequantize_row_iq2_xs (const block_iq2_xs * restrict x, float * restrict y, int k); // Dot product void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy); @@ -234,3 +245,4 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); +void ggml_vec_dot_iq2_xs_q8_K (int n, float * restrict s, const void * restrict vx, const void * restrict vy); diff --git a/ggml.c b/ggml.c index 9c42a45e3..d2a8c0478 100644 --- a/ggml.c +++ b/ggml.c @@ -584,6 +584,17 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .vec_dot = ggml_vec_dot_iq2_xxs_q8_K, .vec_dot_type = GGML_TYPE_Q8_K, }, + [GGML_TYPE_IQ2_XS] = { + .type_name = "iq2_xs", + .blck_size = QK_K, + .type_size = sizeof(block_iq2_xs), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_iq2_xs, + .from_float = quantize_row_iq2_xs, + .from_float_reference = (ggml_from_float_t) quantize_row_iq2_xs_reference, + .vec_dot = ggml_vec_dot_iq2_xs_q8_K, + .vec_dot_type = GGML_TYPE_Q8_K, + }, [GGML_TYPE_Q8_K] = { .type_name = "q8_K", .blck_size = QK_K, @@ -2123,6 +2134,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_Q5_K: wtype = GGML_TYPE_Q5_K; break; case GGML_FTYPE_MOSTLY_Q6_K: wtype = GGML_TYPE_Q6_K; break; case GGML_FTYPE_MOSTLY_IQ2_XXS: wtype = GGML_TYPE_IQ2_XXS; break; + case GGML_FTYPE_MOSTLY_IQ2_XS: wtype = GGML_TYPE_IQ2_XS; break; case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break; case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break; } @@ -7435,6 +7447,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: { ggml_compute_forward_add_q_f32(params, src0, src1, dst); } break; @@ -7700,6 +7713,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: { ggml_compute_forward_add1_q_f32(params, src0, src1, dst); } break; @@ -7815,6 +7829,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: default: { GGML_ASSERT(false); @@ -10457,6 +10472,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: { ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst); } break; @@ -10632,6 +10648,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: default: { GGML_ASSERT(false); @@ -10827,6 +10844,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: { ggml_compute_forward_get_rows_q(params, src0, src1, dst); } break; @@ -11464,6 +11482,7 @@ static void ggml_compute_forward_alibi( case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: case GGML_TYPE_Q8_K: case GGML_TYPE_I8: case GGML_TYPE_I16: @@ -11539,6 +11558,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: case GGML_TYPE_Q8_K: case GGML_TYPE_I8: case GGML_TYPE_I16: @@ -18660,6 +18680,12 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i block_iq2_xxs * block = (block_iq2_xxs*)dst + start / QK_K; result = ggml_quantize_iq2_xxs(src + start, block, n, n, hist); } break; + case GGML_TYPE_IQ2_XS: + { + GGML_ASSERT(start % QK_K == 0); + block_iq2_xs * block = (block_iq2_xs*)dst + start / QK_K; + result = ggml_quantize_iq2_xs(src + start, block, n, n, hist); + } break; case GGML_TYPE_F16: { int elemsize = sizeof(ggml_fp16_t); @@ -19015,8 +19041,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p (int64_t) info->ne[3]; if (ne % ggml_blck_size(info->type) != 0) { - fprintf(stderr, "%s: tensor '%s' number of elements (%" PRId64 ") is not a multiple of block size (%d)\n", - __func__, info->name.data, ne, ggml_blck_size(info->type)); + fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%d)\n", + __func__, info->name.data, (int)info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type)); fclose(file); gguf_free(ctx); return NULL; diff --git a/ggml.h b/ggml.h index 127dcef1d..93b42a27d 100644 --- a/ggml.h +++ b/ggml.h @@ -342,6 +342,7 @@ extern "C" { GGML_TYPE_Q6_K = 14, GGML_TYPE_Q8_K = 15, GGML_TYPE_IQ2_XXS = 16, + GGML_TYPE_IQ2_XS = 17, GGML_TYPE_I8, GGML_TYPE_I16, GGML_TYPE_I32, @@ -377,6 +378,7 @@ extern "C" { GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors + GGML_FTYPE_MOSTLY_IQ2_XS = 16, // except 1d tensors }; // available tensor operations: @@ -2061,6 +2063,7 @@ extern "C" { GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_iq2_xxs(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_iq2_xs (const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist); diff --git a/llama.cpp b/llama.cpp index aaadfa444..bd219d49c 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2223,6 +2223,7 @@ struct llama_model_loader { case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break; + case GGML_TYPE_IQ2_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS; break; default: { LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); @@ -2595,6 +2596,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium"; case LLAMA_FTYPE_MOSTLY_Q6_K: return "Q6_K"; case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw"; default: return "unknown, may not work"; } @@ -9050,6 +9052,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break; case LLAMA_FTYPE_MOSTLY_Q6_K: quantized_type = GGML_TYPE_Q6_K; break; case LLAMA_FTYPE_MOSTLY_IQ2_XXS:quantized_type = GGML_TYPE_IQ2_XXS; break; + case LLAMA_FTYPE_MOSTLY_IQ2_XS :quantized_type = GGML_TYPE_IQ2_XS; break; default: throw std::runtime_error(format("invalid output file type %d\n", ftype)); } diff --git a/llama.h b/llama.h index c11075bbc..6fde113ff 100644 --- a/llama.h +++ b/llama.h @@ -104,6 +104,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q5_K_M = 17, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors + LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file }; diff --git a/tests/test-quantize-fns.cpp b/tests/test-quantize-fns.cpp index cee712618..31a78c632 100644 --- a/tests/test-quantize-fns.cpp +++ b/tests/test-quantize-fns.cpp @@ -134,8 +134,9 @@ int main(int argc, char * argv[]) { continue; } - if ((ggml_type)i == GGML_TYPE_IQ2_XXS) { - printf("Skip %s due to missing quantization functionality\n", ggml_type_name((ggml_type) i)); + const ggml_type ei = (ggml_type)i; + if (ei == GGML_TYPE_IQ2_XXS || ei == GGML_TYPE_IQ2_XS) { + printf("Skip %s due to missing quantization functionality\n", ggml_type_name(ei)); continue; } From 469e75d0a35b08de549a4fd87f082ca7a8a539ba Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Thu, 11 Jan 2024 20:43:15 +0100 Subject: [PATCH 30/42] llama : restore intended k-quants mixes for MoE models (#4872) * Restore intended k-quants quantization mixes for MoE models * Update Q2_K_S values in the quantize tool Still using LLaMA-v1 PPL values in the quant description today does not make much sense. But let's leave this update for another PR. --------- Co-authored-by: Iwan Kawrakow Co-authored-by: Georgi Gerganov --- examples/quantize/quantize.cpp | 1 + llama.cpp | 24 +++++++++++++++--------- llama.h | 1 + 3 files changed, 17 insertions(+), 9 deletions(-) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index d27ea5e91..f878f6911 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -18,6 +18,7 @@ static const std::vector QUANT_OPTIONS = { { "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", }, { "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", }, { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", }, + { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", }, { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", }, { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", }, diff --git a/llama.cpp b/llama.cpp index bd219d49c..d39ff94c7 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2586,7 +2586,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0"; // K-quants - case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K"; + case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K - Medium"; + case LLAMA_FTYPE_MOSTLY_Q2_K_S: return "Q2_K - Small"; case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small"; case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium"; case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large"; @@ -8955,10 +8956,13 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty // TODO: explore better strategies new_type = GGML_TYPE_Q8_0; } - } else if (name.find("ffn_down.weight") != std::string::npos) { + } else if (name.find("ffn_down") != std::string::npos) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) { + if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q4_K; + } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { - new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K + new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q5_K : arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K; } @@ -8967,14 +8971,14 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty } else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) { if (arch == LLM_ARCH_FALCON) { - new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K : + new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q6_K : use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; } else { if (use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; } } else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < 4) { + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) { new_type = GGML_TYPE_Q5_K; } ++qs.i_feed_forward_w2; @@ -8992,9 +8996,10 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K; } - else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; - } + // IK: let's remove this, else Q2_K is almost the same as Q3_K_S + //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) { + // if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + //} // This can be used to reduce the size of the Q5_K_S model. // The associated PPL increase is fully in line with the size reduction //else { @@ -9043,6 +9048,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // K-quants case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break; + case LLAMA_FTYPE_MOSTLY_Q2_K_S: quantized_type = GGML_TYPE_Q2_K; break; case LLAMA_FTYPE_MOSTLY_Q3_K_S: case LLAMA_FTYPE_MOSTLY_Q3_K_M: case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break; @@ -9101,7 +9107,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) { ++qs.n_attention_wv; } - else if (name.find("ffn_down.weight") != std::string::npos) { + else if (name.find("ffn_down") != std::string::npos) { ++qs.n_feed_forward_w2; } } diff --git a/llama.h b/llama.h index 6fde113ff..43d41b8f6 100644 --- a/llama.h +++ b/llama.h @@ -105,6 +105,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q6_K = 18, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file }; From b0377875488b33f7114138687d828da1de61775d Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 11 Jan 2024 21:58:28 +0200 Subject: [PATCH 31/42] swift : track ggml release branch (#4867) --- Package.swift | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Package.swift b/Package.swift index 59191da45..37524edee 100644 --- a/Package.swift +++ b/Package.swift @@ -14,7 +14,7 @@ let package = Package( .library(name: "llama", targets: ["llama"]), ], dependencies: [ - .package(url: "https://github.com/ggerganov/ggml.git", .revision("979cc23b345006504cfc1f67c0fdf627805e3319")) + .package(url: "https://github.com/ggerganov/ggml.git", .branch("release")) ], targets: [ .target( From 3ca63b4538dfc78aaec88cd2c3e3f8417c1924e3 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 11 Jan 2024 22:43:05 +0200 Subject: [PATCH 32/42] main : disable token count by default (#4874) --- common/common.cpp | 6 +++--- common/common.h | 2 +- examples/main/main.cpp | 4 ++-- 3 files changed, 6 insertions(+), 6 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index bfcd6d4df..287e8bd5a 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -630,7 +630,7 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.ppl_stride = std::stoi(argv[i]); - } else if (arg == "-stc" || arg == "--show_token_count") { + } else if (arg == "-stc" || arg == "--show-token-count") { if (++i >= argc) { invalid_param = true; break; @@ -950,8 +950,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" advanced option to override model metadata by key. may be specified multiple times.\n"); printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n"); - printf(" -stc N --show_token_count N\n"); - printf(" show consumed tokens every N tokens\n"); + printf(" -stc N --show-token-count N\n"); + printf(" show consumed tokens every N tokens (default: %d)\n", params.token_interval); printf("\n"); #ifndef LOG_DISABLE_LOGS log_print_usage(); diff --git a/common/common.h b/common/common.h index a295e88b0..82d23cf54 100644 --- a/common/common.h +++ b/common/common.h @@ -64,7 +64,7 @@ struct gpt_params { int32_t n_beams = 0; // if non-zero then use beam search of given width. int32_t grp_attn_n = 1; // group-attention factor int32_t grp_attn_w = 512; // group-attention width - int32_t token_interval = 512; // show token count every 512 tokens + int32_t token_interval = -1; // show token count every 512 tokens (-1 = disabled) float rope_freq_base = 0.0f; // RoPE base frequency float rope_freq_scale = 0.0f; // RoPE frequency scaling factor float yarn_ext_factor = -1.0f; // YaRN extrapolation mix factor diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 1f35febbd..6953d107c 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -651,8 +651,8 @@ int main(int argc, char ** argv) { LOG("n_past = %d\n", n_past); // Display total tokens alongside total time - if (n_past % params.token_interval == 0) { - printf("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx); + if (params.token_interval > 0 && n_past % params.token_interval == 0) { + LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx); } } From 7edefbd79cc6dea96640edc54c6b94b2b2496d8b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 11 Jan 2024 22:46:26 +0200 Subject: [PATCH 33/42] main : better name for variable n_print (#4874) --- common/common.cpp | 8 ++++---- common/common.h | 2 +- examples/main/main.cpp | 2 +- 3 files changed, 6 insertions(+), 6 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 287e8bd5a..b2cb0e257 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -630,12 +630,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.ppl_stride = std::stoi(argv[i]); - } else if (arg == "-stc" || arg == "--show-token-count") { + } else if (arg == "-ptc" || arg == "--print-token-count") { if (++i >= argc) { invalid_param = true; break; } - params.token_interval = std::stoi(argv[i]); + params.n_print = std::stoi(argv[i]); } else if (arg == "--ppl-output-type") { if (++i >= argc) { invalid_param = true; @@ -950,8 +950,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" advanced option to override model metadata by key. may be specified multiple times.\n"); printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n"); - printf(" -stc N --show-token-count N\n"); - printf(" show consumed tokens every N tokens (default: %d)\n", params.token_interval); + printf(" -stc N --print-token-count N\n"); + printf(" print token count every N tokens (default: %d)\n", params.n_print); printf("\n"); #ifndef LOG_DISABLE_LOGS log_print_usage(); diff --git a/common/common.h b/common/common.h index 82d23cf54..1359e76ab 100644 --- a/common/common.h +++ b/common/common.h @@ -64,7 +64,7 @@ struct gpt_params { int32_t n_beams = 0; // if non-zero then use beam search of given width. int32_t grp_attn_n = 1; // group-attention factor int32_t grp_attn_w = 512; // group-attention width - int32_t token_interval = -1; // show token count every 512 tokens (-1 = disabled) + int32_t n_print = -1; // print token count every n tokens (-1 = disabled) float rope_freq_base = 0.0f; // RoPE base frequency float rope_freq_scale = 0.0f; // RoPE frequency scaling factor float yarn_ext_factor = -1.0f; // YaRN extrapolation mix factor diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 6953d107c..c53b29978 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -651,7 +651,7 @@ int main(int argc, char ** argv) { LOG("n_past = %d\n", n_past); // Display total tokens alongside total time - if (params.token_interval > 0 && n_past % params.token_interval == 0) { + if (params.n_print > 0 && n_past % params.n_print == 0) { LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx); } } From 1d118386fea031f01550f8cd47a5c86296e5333f Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 11 Jan 2024 23:23:49 +0200 Subject: [PATCH 34/42] server : fix infill when prompt is empty (#4833) --- examples/server/server.cpp | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 031824e14..1d30a15a6 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1406,7 +1406,7 @@ struct llama_server_context task.multitask_id = multitask_id; // when a completion task's prompt array is not a singleton, we split it into multiple requests - if (task.data.at("prompt").size() > 1) + if (task.data.count("prompt") && task.data.at("prompt").size() > 1) { lock.unlock(); // entering new func scope return split_multiprompt_task(task); @@ -1577,9 +1577,9 @@ struct llama_server_context slot->reset(); - slot->infill = task.infill_mode; - slot->embedding = task.embedding_mode; - slot->task_id = task.id; + slot->infill = task.infill_mode; + slot->embedding = task.embedding_mode; + slot->task_id = task.id; slot->multitask_id = task.multitask_id; if (!launch_slot_with_data(slot, task.data)) @@ -1731,7 +1731,8 @@ struct llama_server_context const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get().empty()) || !slot.images.empty(); // empty prompt passed -> release the slot and send empty response - if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt) + // note: infill mode allows empty prompt + if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill) { slot.release(); slot.print_timings(); @@ -2609,8 +2610,8 @@ static json format_final_response_oaicompat(const json &request, const task_resu {"object", streaming ? "chat.completion.chunk" : "chat.completion"}, {"usage", json{{"completion_tokens", num_tokens_predicted}, - {"prompt_tokens", num_prompt_tokens}, - {"total_tokens", num_tokens_predicted + num_prompt_tokens}}}, + {"prompt_tokens", num_prompt_tokens}, + {"total_tokens", num_tokens_predicted + num_prompt_tokens}}}, {"id", gen_chatcmplid()}}; if (server_verbose) { From 326b418b59b6d48d854c4461a2303e8ac0a311e6 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Fri, 12 Jan 2024 06:59:57 +0100 Subject: [PATCH 35/42] Importance Matrix calculation (#4861) * imatrix: 1st version * imatrix: WIP * Cleanup * Update examples/imatrix/imatrix.cpp Co-authored-by: Georgi Gerganov --------- Co-authored-by: Iwan Kawrakow Co-authored-by: Georgi Gerganov --- Makefile | 5 +- examples/CMakeLists.txt | 1 + examples/imatrix/CMakeLists.txt | 5 + examples/imatrix/imatrix.cpp | 380 ++++++++++++++++++++++++++++++++ ggml.c | 14 ++ ggml.h | 6 + 6 files changed, 410 insertions(+), 1 deletion(-) create mode 100644 examples/imatrix/CMakeLists.txt create mode 100644 examples/imatrix/imatrix.cpp diff --git a/Makefile b/Makefile index 4c7e175bf..05fe9a0f6 100644 --- a/Makefile +++ b/Makefile @@ -1,6 +1,6 @@ # Define the default target now so that it is always the first target BUILD_TARGETS = \ - main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ + main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search \ speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey tests/test-c.o @@ -614,6 +614,9 @@ quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.o ggml. perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) +imatrix: examples/imatrix/imatrix.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + embedding: examples/embedding/embedding.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 0c71cbdf7..fa127a3aa 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -36,6 +36,7 @@ else() add_subdirectory(lookahead) add_subdirectory(lookup) add_subdirectory(train-text-from-scratch) + add_subdirectory(imatrix) if (LLAMA_METAL) add_subdirectory(metal) endif() diff --git a/examples/imatrix/CMakeLists.txt b/examples/imatrix/CMakeLists.txt new file mode 100644 index 000000000..d688a1620 --- /dev/null +++ b/examples/imatrix/CMakeLists.txt @@ -0,0 +1,5 @@ +set(TARGET imatrix) +add_executable(${TARGET} imatrix.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp new file mode 100644 index 000000000..1461bc963 --- /dev/null +++ b/examples/imatrix/imatrix.cpp @@ -0,0 +1,380 @@ +#include "common.h" +#include "llama.h" + +#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 + +struct Stats { + std::vector values; + int ncall = 0; +}; + +struct StatParams { + std::string ofile = "imatrix.dat"; + int n_output_frequency = 10; + int verbosity = 1; + bool collect_output_weight = false; +}; + +class IMatrixCollector { +public: + IMatrixCollector() = default; + void set_parameters(StatParams&& params) { m_params = std::move(params); } + void collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1); + void save_imatrix() const; +private: + std::unordered_map m_stats; + StatParams m_params; + std::mutex m_mutex; + int m_last_call = 0; +}; + +void IMatrixCollector::collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) { + if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return; + if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return; + std::lock_guard lock(m_mutex); + auto& e = m_stats[src0->name]; + if (e.values.empty()) { + e.values.resize(src1->ne[0], 0); + } + else if (e.values.size() != (size_t)src1->ne[0]) { + fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]); + exit(1); //GGML_ASSERT(false); + } + ++e.ncall; + if (m_params.verbosity > 1) { + printf("%s[%d]: %s, %d x %d, %d\n",__func__,m_last_call,src0->name,(int)src1->ne[0],(int)src1->ne[1],(int)src1->type); + } + for (int row = 0; row < (int)src1->ne[1]; ++row) { + const float * x = (const float *)src1->data + row * src1->ne[0]; + for (int j = 0; j < (int)src1->ne[0]; ++j) { + e.values[j] += x[j]*x[j]; + } + } + if (e.ncall > m_last_call) { + m_last_call = e.ncall; + if (m_last_call % m_params.n_output_frequency == 0) { + save_imatrix(); + } + } +} + +void IMatrixCollector::save_imatrix() const { + const char * fname = m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str(); + std::ofstream out(fname, std::ios::binary); + int n_entries = m_stats.size(); + out.write((const char*)&n_entries, sizeof(n_entries)); + for (auto& p : m_stats) { + int len = p.first.size(); + out.write((const char*)&len, sizeof(len)); + out.write(p.first.c_str(), len); + out.write((const char*)&p.second.ncall, sizeof(p.second.ncall)); + int nval = p.second.values.size(); + out.write((const char*)&nval, sizeof(nval)); + if (nval > 0) out.write((const char*)p.second.values.data(), nval*sizeof(float)); + } + if (m_params.verbosity > 0) { + fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n",__func__,m_last_call,fname); + } +} + +static IMatrixCollector g_collector; + +static void ik_collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) { + g_collector.collect_imatrix(src0, src1); +} + + +struct results_log_softmax { + double log_softmax; + float logit; + float prob; +}; + +static std::vector softmax(const std::vector& logits) { + std::vector probs(logits.size()); + float max_logit = logits[0]; + for (float v : logits) { + max_logit = std::max(max_logit, v); + } + double sum_exp = 0.0; + for (size_t i = 0; i < logits.size(); i++) { + // Subtract the maximum logit value from the current logit value for numerical stability + const float logit = logits[i] - max_logit; + const float exp_logit = expf(logit); + sum_exp += exp_logit; + probs[i] = exp_logit; + } + for (size_t i = 0; i < probs.size(); i++) { + probs[i] /= sum_exp; + } + return probs; +} + +static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) { + float max_logit = logits[0]; + for (int i = 1; i < n_vocab; ++i) { + max_logit = std::max(max_logit, logits[i]); + } + double sum_exp = 0.0; + for (int i = 0; i < n_vocab; ++i) { + sum_exp += expf(logits[i] - max_logit); + } + return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp}; +} + +static void process_logits( + int n_vocab, const float * logits, const int * tokens, int n_token, std::vector & workers, + double & nll, double & nll2, float * logit_history, float * prob_history +) { + std::mutex mutex; + int counter = 0; + auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () { + double local_nll = 0; + double local_nll2 = 0; + while (true) { + std::unique_lock lock(mutex); + int i = counter++; + if (i >= n_token) { + nll += local_nll; nll2 += local_nll2; + break; + } + lock.unlock(); + const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]); + const double v = -results.log_softmax; + local_nll += v; + local_nll2 += v*v; + + logit_history[i] = results.logit; + prob_history[i] = results.prob; + } + }; + for (auto & w : workers) { + w = std::thread(compute); + } + compute(); + for (auto & w : workers) { + w.join(); + } +} + +static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { + + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + const int n_ctx = llama_n_ctx(ctx); + + auto tim1 = std::chrono::high_resolution_clock::now(); + fprintf(stderr, "%s: tokenizing the input ..\n", __func__); + + std::vector tokens = ::llama_tokenize(ctx, params.prompt, add_bos); + + auto tim2 = std::chrono::high_resolution_clock::now(); + fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast(tim2-tim1).count()); + + if (int(tokens.size()) < 2*n_ctx) { + fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx, + n_ctx); + fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size()); + return false; + } + + std::vector logit_history; + logit_history.resize(tokens.size()); + + std::vector prob_history; + prob_history.resize(tokens.size()); + + const int n_chunk_max = tokens.size() / n_ctx; + + const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max); + const int n_vocab = llama_n_vocab(llama_get_model(ctx)); + const int n_batch = params.n_batch; + + int count = 0; + double nll = 0.0; + double nll2 = 0.0; + + fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch); + + std::vector workers(std::thread::hardware_concurrency() - 1); + + for (int i = 0; i < n_chunk; ++i) { + const int start = i * n_ctx; + const int end = start + n_ctx; + + const int num_batches = (n_ctx + n_batch - 1) / n_batch; + + std::vector logits; + + const auto t_start = std::chrono::high_resolution_clock::now(); + + // clear the KV cache + llama_kv_cache_clear(ctx); + + for (int j = 0; j < num_batches; ++j) { + const int batch_start = start + j * n_batch; + const int batch_size = std::min(end - batch_start, n_batch); + + // save original token and restore it after eval + const auto token_org = tokens[batch_start]; + + // add BOS token for the first batch of each chunk + if (add_bos && j == 0) { + tokens[batch_start] = llama_token_bos(llama_get_model(ctx)); + } + + if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) { + fprintf(stderr, "%s : failed to eval\n", __func__); + return false; + } + + // restore the original token in case it was set to BOS + tokens[batch_start] = token_org; + + const auto * batch_logits = llama_get_logits(ctx); + logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + } + + const auto t_end = std::chrono::high_resolution_clock::now(); + + if (i == 0) { + const float t_total = std::chrono::duration(t_end - t_start).count(); + fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total); + int total_seconds = (int)(t_total * n_chunk); + if (total_seconds >= 60*60) { + fprintf(stderr, "%d hours ", total_seconds / (60*60)); + total_seconds = total_seconds % (60*60); + } + fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); + } + + const int first = n_ctx/2; + process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); + count += n_ctx - first - 1; + + printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); + fflush(stdout); + } + printf("\n"); + + nll2 /= count; + nll /= count; + const double ppl = exp(nll); + nll2 -= nll * nll; + if (nll2 > 0) { + nll2 = sqrt(nll2/(count-1)); + printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl); + } else { + printf("Unexpected negative standard deviation of log(prob)\n"); + } + + return true; +} + +int main(int argc, char ** argv) { + + StatParams sparams; + std::vector args; + args.push_back(argv[0]); + int iarg = 1; + for (; iarg < argc-1; ++iarg) { + std::string arg{argv[iarg]}; + if (arg == "-o" || arg == "--output-file") { + sparams.ofile = argv[++iarg]; + } + else if (arg == "-ofreq" || arg == "--output-frequency") { + sparams.n_output_frequency = std::stoi(argv[++iarg]); + } + else if (arg == "-ow" || arg == "--output-weight") { + sparams.collect_output_weight = std::stoi(argv[++iarg]); + } + else if (arg == "--verbosity") { + sparams.verbosity = std::stoi(argv[++iarg]); + } else { + args.push_back(argv[iarg]); + } + } + if (iarg < argc) { + args.push_back(argv[iarg]); + } + + gpt_params params; + params.n_batch = 512; + if (!gpt_params_parse(args.size(), args.data(), params)) { + return 1; + } + + g_collector.set_parameters(std::move(sparams)); + + ggml_set_imatrix_collection(ik_collect_imatrix); + + params.logits_all = true; + params.n_batch = std::min(params.n_batch, params.n_ctx); + + print_build_info(); + + if (params.seed == LLAMA_DEFAULT_SEED) { + params.seed = time(NULL); + } + + fprintf(stderr, "%s: seed = %u\n", __func__, params.seed); + + std::mt19937 rng(params.seed); + if (params.random_prompt) { + params.prompt = gpt_random_prompt(rng); + } + + llama_backend_init(params.numa); + + llama_model * model; + llama_context * ctx; + + // load the model and apply lora adapter, if any + std::tie(model, ctx) = llama_init_from_gpt_params(params); + if (model == NULL) { + fprintf(stderr, "%s: error: unable to load model\n", __func__); + return 1; + } + + const int n_ctx_train = llama_n_ctx_train(model); + if (params.n_ctx > n_ctx_train) { + fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n", + __func__, n_ctx_train, params.n_ctx); + } + + // print system information + { + fprintf(stderr, "\n"); + fprintf(stderr, "%s\n", get_system_info(params).c_str()); + } + + bool OK = compute_imatrix(ctx, params); + if (!OK) { + return 1; + } + + g_collector.save_imatrix(); + + llama_print_timings(ctx); + + llama_free(ctx); + llama_free_model(model); + + llama_backend_free(); + + return 0; +} diff --git a/ggml.c b/ggml.c index d2a8c0478..f5caeba08 100644 --- a/ggml.c +++ b/ggml.c @@ -394,6 +394,12 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y); static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y); +ggml_collect_imatrix_t g_imatrix_collect = NULL; + +void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect) { + g_imatrix_collect = imatrix_collect; +} + static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { [GGML_TYPE_I8] = { .type_name = "i8", @@ -9763,6 +9769,10 @@ static void ggml_compute_forward_mul_mat( const int ith = params->ith; const int nth = params->nth; + if (ith == 1 && g_imatrix_collect) { + g_imatrix_collect(src0, src1); + } + const enum ggml_type type = src0->type; const bool src1_cont = ggml_is_contiguous(src1); @@ -10066,6 +10076,10 @@ static void ggml_compute_forward_mul_mat_id( const struct ggml_tensor * src0_cur = dst->src[cur_a + 2]; + if (ith == 1 && g_imatrix_collect) { + g_imatrix_collect(src0_cur, src1); + } + const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; const size_t row_size = ggml_row_size(vec_dot_type, ne10); diff --git a/ggml.h b/ggml.h index 93b42a27d..4c2ff6c66 100644 --- a/ggml.h +++ b/ggml.h @@ -2067,6 +2067,12 @@ extern "C" { GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist); + // + // Importance matrix + // + typedef void(*ggml_collect_imatrix_t)(const struct ggml_tensor * src0, const struct ggml_tensor * src1); + GGML_API void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect); + // // gguf // From f445c0e68cf8e1faca0b2aa8dfb9d48231cec301 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 12 Jan 2024 13:01:56 +0200 Subject: [PATCH 36/42] llama : fix llm_build_k_shift to use correct n_rot (#4889) * llama : fix llm_build_k_shift to use correct n_rot ggml-ci * llama : always use hparams.n_rot for ggml_rope_custom ggml-ci * convert : fix persimmon conversion to write correct n_rot --- common/common.cpp | 3 ++ convert-hf-to-gguf.py | 9 ++++- gguf-py/gguf/tensor_mapping.py | 7 ++++ llama.cpp | 65 +++++++++++++++++----------------- 4 files changed, 51 insertions(+), 33 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index b2cb0e257..3aefed01d 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1055,6 +1055,9 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & } static ggml_type kv_cache_type_from_str(const std::string & s) { + if (s == "f32") { + return GGML_TYPE_F32; + } if (s == "f16") { return GGML_TYPE_F16; } diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 203eaf64b..813aeeed6 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -817,10 +817,17 @@ class PersimmonModel(Model): hidden_size = self.hparams["hidden_size"] self.gguf_writer.add_name('persimmon-8b-chat') + self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hidden_size) self.gguf_writer.add_block_count(block_count) self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) - self.gguf_writer.add_rope_dimension_count(hidden_size // head_count) + + # NOTE: not sure about this change - why does the model not have a rope dimension count when it is smaller + # than the head size? + # ref: https://github.com/ggerganov/llama.cpp/pull/4889 + #self.gguf_writer.add_rope_dimension_count(hidden_size // head_count) + self.gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2) + self.gguf_writer.add_head_count(head_count) self.gguf_writer.add_head_count_kv(head_count_kv) self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"]) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 80c1d5449..24a089037 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -57,6 +57,7 @@ class TensorNameMap: "transformer.norm_f", # mpt "ln_f", # refact bloom qwen gpt2 "language_model.encoder.final_layernorm", # persimmon + "model.final_layernorm", # persimmon "lm_head.ln", # phi2 ), @@ -98,6 +99,7 @@ class TensorNameMap: "transformer.h.{bid}.self_attention.query_key_value", # falcon "h.{bid}.self_attention.query_key_value", # bloom "language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon + "model.layers.{bid}.self_attn.query_key_value", # persimmon "h.{bid}.attn.c_attn", # gpt2 "transformer.h.{bid}.mixer.Wqkv", # phi2 ), @@ -141,6 +143,7 @@ class TensorNameMap: "encoder.layer.{bid}.attention.output.dense", # bert "transformer.h.{bid}.attn.out_proj", # gpt-j "language_model.encoder.layers.{bid}.self_attention.dense", # persimmon + "model.layers.{bid}.self_attn.dense", # persimmon "h.{bid}.attn.c_proj", # gpt2 "transformer.h.{bid}.mixer.out_proj", # phi2 "model.layers.layers.{bid}.self_attn.o_proj", # plamo @@ -184,6 +187,7 @@ class TensorNameMap: "encoder.layer.{bid}.intermediate.dense", # bert "transformer.h.{bid}.mlp.fc_in", # gpt-j "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon + "model.layers.{bid}.mlp.dense_h_to_4h", # persimmon "transformer.h.{bid}.mlp.w1", # qwen "h.{bid}.mlp.c_fc", # gpt2 "transformer.h.{bid}.mlp.fc1", # phi2 @@ -225,6 +229,7 @@ class TensorNameMap: "encoder.layer.{bid}.output.dense", # bert "transformer.h.{bid}.mlp.fc_out", # gpt-j "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon + "model.layers.{bid}.mlp.dense_4h_to_h", # persimmon "h.{bid}.mlp.c_proj", # gpt2 "transformer.h.{bid}.mlp.fc2", # phi2 "model.layers.layers.{bid}.mlp.down_proj", # plamo @@ -237,10 +242,12 @@ class TensorNameMap: MODEL_TENSOR.ATTN_Q_NORM: ( "language_model.encoder.layers.{bid}.self_attention.q_layernorm", + "model.layers.{bid}.self_attn.q_layernorm", # persimmon ), MODEL_TENSOR.ATTN_K_NORM: ( "language_model.encoder.layers.{bid}.self_attention.k_layernorm", + "model.layers.{bid}.self_attn.k_layernorm", # persimmon ), MODEL_TENSOR.ROPE_FREQS: ( diff --git a/llama.cpp b/llama.cpp index d39ff94c7..0bab95563 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4104,7 +4104,6 @@ static void llm_build_k_shift( struct ggml_cgraph * graph, llm_rope_type type, int64_t n_ctx, - int n_rot, float freq_base, float freq_scale, const llm_build_cb & cb) { @@ -4112,14 +4111,13 @@ static void llm_build_k_shift( const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head_k = hparams.n_embd_head_k; const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); + const int32_t n_rot = hparams.n_rot; const int32_t n_orig_ctx = cparams.n_yarn_orig_ctx; const float ext_factor = cparams.yarn_ext_factor; const float attn_factor = cparams.yarn_attn_factor; const float beta_fast = cparams.yarn_beta_fast; const float beta_slow = cparams.yarn_beta_slow; - GGML_ASSERT(n_embd_head_k % n_rot == 0); - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx); cb(K_shift, "K_shift", -1); @@ -4523,7 +4521,7 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4561,14 +4559,14 @@ struct llm_build_context { Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, - n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); Kcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, - n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); @@ -4691,6 +4689,7 @@ struct llm_build_context { const int64_t n_embd_head = hparams.n_embd_head_v; GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -4708,7 +4707,7 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4734,12 +4733,12 @@ struct llm_build_context { case MODEL_7B: Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, - n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); Kcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, - n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); break; @@ -4812,6 +4811,7 @@ struct llm_build_context { const int64_t n_embd_head = hparams.n_embd_head_v; const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -4829,7 +4829,7 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4870,13 +4870,13 @@ struct llm_build_context { // using mode = 2 for neox mode Qcur = ggml_rope_custom( - ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + ctx0, Qcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); Kcur = ggml_rope_custom( - ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); @@ -5033,9 +5033,8 @@ struct llm_build_context { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - const int64_t n_rot = n_embd_head_k / 2; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head/2 == hparams.n_rot); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -5052,7 +5051,7 @@ struct llm_build_context { cb(KQ_mask, "KQ_mask", -1); if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5112,7 +5111,7 @@ struct llm_build_context { // RoPE the first n_rot of q/k, pass the other half, and concat. struct ggml_tensor * qrot = ggml_view_3d( - ctx0, tmpq, n_rot, n_head, n_tokens, + ctx0, tmpq, hparams.n_rot, n_head, n_tokens, ggml_element_size(tmpq) * n_embd_head, ggml_element_size(tmpq) * n_embd_head * n_head, 0 @@ -5120,7 +5119,7 @@ struct llm_build_context { cb(qrot, "qrot", il); struct ggml_tensor * krot = ggml_view_3d( - ctx0, tmpk, n_rot, n_head, n_tokens, + ctx0, tmpk, hparams.n_rot, n_head, n_tokens, ggml_element_size(tmpk) * n_embd_head, ggml_element_size(tmpk) * n_embd_head * n_head, 0 @@ -5129,29 +5128,29 @@ struct llm_build_context { // get the second half of tmpq, e.g tmpq[n_rot:, :, :] struct ggml_tensor * qpass = ggml_view_3d( - ctx0, tmpq, n_rot, n_head, n_tokens, + ctx0, tmpq, hparams.n_rot, n_head, n_tokens, ggml_element_size(tmpq) * n_embd_head, ggml_element_size(tmpq) * n_embd_head * n_head, - ggml_element_size(tmpq) * n_rot + ggml_element_size(tmpq) * hparams.n_rot ); cb(qpass, "qpass", il); struct ggml_tensor * kpass = ggml_view_3d( - ctx0, tmpk, n_rot, n_head, n_tokens, + ctx0, tmpk, hparams.n_rot, n_head, n_tokens, ggml_element_size(tmpk) * n_embd_head, ggml_element_size(tmpk) * n_embd_head * n_head, - ggml_element_size(tmpk) * n_rot + ggml_element_size(tmpk) * hparams.n_rot ); cb(kpass, "kpass", il); struct ggml_tensor * qrotated = ggml_rope_custom( - ctx0, qrot, inp_pos, n_rot, 2, 0, n_orig_ctx, + ctx0, qrot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(qrotated, "qrotated", il); struct ggml_tensor * krotated = ggml_rope_custom( - ctx0, krot, inp_pos, n_rot, 2, 0, n_orig_ctx, + ctx0, krot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(krotated, "krotated", il); @@ -5531,6 +5530,7 @@ struct llm_build_context { const int64_t n_embd_head = hparams.n_embd_head_v; GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -5548,7 +5548,7 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, hparams.n_rot, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5661,7 +5661,7 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5693,13 +5693,13 @@ struct llm_build_context { // using mode = 2 for neox mode Qcur = ggml_rope_custom( - ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + ctx0, Qcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); Kcur = ggml_rope_custom( - ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); @@ -5778,7 +5778,7 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5874,6 +5874,7 @@ struct llm_build_context { const int64_t n_embd_head = hparams.n_embd_head_v; GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -5891,7 +5892,7 @@ struct llm_build_context { // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5917,13 +5918,13 @@ struct llm_build_context { cb(Vcur, "Vcur", il); Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + ctx0, ggml_reshape_3d(ctx0, Qcur, hparams.n_rot, n_head, n_tokens), inp_pos, n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cb(Qcur, "Qcur", il); Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + ctx0, ggml_reshape_3d(ctx0, Kcur, hparams.n_rot, n_head_kv, n_tokens), inp_pos, n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); From 2d00741e12c5db4a33dfccd1125f5de4adec9a5b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 12 Jan 2024 13:03:38 +0200 Subject: [PATCH 37/42] py : fix lint (#4889) --- convert-hf-to-gguf.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 813aeeed6..a1c79fd47 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -825,7 +825,7 @@ class PersimmonModel(Model): # NOTE: not sure about this change - why does the model not have a rope dimension count when it is smaller # than the head size? # ref: https://github.com/ggerganov/llama.cpp/pull/4889 - #self.gguf_writer.add_rope_dimension_count(hidden_size // head_count) + # self.gguf_writer.add_rope_dimension_count(hidden_size // head_count) self.gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2) self.gguf_writer.add_head_count(head_count) From 4315a94366708828f949f9db89d2a8d99b634459 Mon Sep 17 00:00:00 2001 From: howlger Date: Fri, 12 Jan 2024 12:05:32 +0100 Subject: [PATCH 38/42] common : streamline the formatting of help (#4890) * common : streamline the formatting of help - Separate alternative parameters by a comma - Do not indent `--version` differently * Update common/common.cpp --------- Co-authored-by: Georgi Gerganov --- common/common.cpp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 3aefed01d..062a8b4de 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -818,7 +818,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf("\n"); printf("options:\n"); printf(" -h, --help show this help message and exit\n"); - printf(" --version show version and build info\n"); + printf(" --version show version and build info\n"); printf(" -i, --interactive run in interactive mode\n"); printf(" --interactive-first run in interactive mode and wait for input right away\n"); printf(" -ins, --instruct run in instruction mode (use with Alpaca models)\n"); @@ -915,7 +915,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" number of layers to store in VRAM\n"); printf(" -ngld N, --n-gpu-layers-draft N\n"); printf(" number of layers to store in VRAM for the draft model\n"); - printf(" -ts SPLIT --tensor-split SPLIT\n"); + printf(" -ts SPLIT, --tensor-split SPLIT\n"); printf(" how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n"); printf(" -mg i, --main-gpu i the GPU to use for scratch and small tensors\n"); #ifdef GGML_USE_CUBLAS @@ -950,7 +950,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" advanced option to override model metadata by key. may be specified multiple times.\n"); printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n"); - printf(" -stc N --print-token-count N\n"); + printf(" -ptc N, --print-token-count N\n"); printf(" print token count every N tokens (default: %d)\n", params.n_print); printf("\n"); #ifndef LOG_DISABLE_LOGS From 3cabe80630c7eeb57713cd02249053a8cf6894fa Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 12 Jan 2024 13:10:19 +0200 Subject: [PATCH 39/42] llama : fix typo "imp_embd" -> "inp_embd" --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 0bab95563..29f8873f6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -5040,7 +5040,7 @@ struct llm_build_context { struct ggml_tensor * inpL; inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); - cb(inpL, "imp_embd", -1); + cb(inpL, "inp_embd", -1); // inp_pos - contains the positions struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); From 1b280c9fffd682b6924010a4437f0275f2921fa9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Fri, 12 Jan 2024 12:30:41 +0100 Subject: [PATCH 40/42] CUDA: fix softmax compile for old CUDA versions (#4862) --- ggml-cuda.cu | 34 ++++++++++++++++++---------------- 1 file changed, 18 insertions(+), 16 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index dd19699f6..a345b0c4a 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -116,6 +116,8 @@ #include "ggml.h" #include "ggml-backend-impl.h" +#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed) + #define CC_PASCAL 600 #define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products #define CC_VOLTA 700 @@ -605,16 +607,16 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) { } static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { -#if __CUDA_ARCH__ < CC_PASCAL || (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) - (void) a; - bad_arch(); -#else +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL #pragma unroll for (int mask = 16; mask > 0; mask >>= 1) { a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32)); } return a; -#endif // __CUDA_ARCH__ < CC_PASCAL || (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) +#else + (void) a; + bad_arch(); +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL } static __device__ __forceinline__ float warp_reduce_max(float x) { @@ -626,16 +628,16 @@ static __device__ __forceinline__ float warp_reduce_max(float x) { } static __device__ __forceinline__ half2 warp_reduce_max(half2 x) { -#if __CUDA_ARCH__ < CC_PASCAL || (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) - (void) x; - bad_arch(); -#else +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX #pragma unroll for (int mask = 16; mask > 0; mask >>= 1) { x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); } return x; -#endif // __CUDA_ARCH__ < CC_PASCAL || (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) +#else + (void) x; + bad_arch(); +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX } static __device__ __forceinline__ float op_repeat(const float a, const float b) { @@ -5613,7 +5615,7 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int template static __global__ void soft_max_f16(const float * x, const float * y, float * dst, const int ncols_par, const int nrows_y, const float scale) { -#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX const int ncols_data = ncols_template == 0 ? ncols_par : ncols_template; const int ncols_smem = GGML_PAD(ncols_data, 2*WARP_SIZE)/2; @@ -5738,7 +5740,7 @@ static __global__ void soft_max_f16(const float * x, const float * y, float * ds #else (void) x; (void) y; (void) dst; (void) ncols_par; (void) nrows_y; (void) scale; bad_arch(); -#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX } template @@ -8574,15 +8576,15 @@ static void ggml_cuda_op_soft_max( float scale = 1.0f; memcpy(&scale, dst->op_params, sizeof(float)); -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - const bool use_f16_soft_max = false; -#else +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION >= CUDART_HMAX #ifdef GGML_CUDA_F16 const bool use_f16_soft_max = true; #else const bool use_f16_soft_max = false; #endif // GGML_CUDA_F16 -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#else + const bool use_f16_soft_max = false; +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && CUDART_VERSION >= CUDART_HMAX if (use_f16_soft_max) { soft_max_f16_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream); From 5537d9d36bfdb4379555431f574d3d78ce6e7955 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 12 Jan 2024 14:33:21 +0200 Subject: [PATCH 41/42] gitignore : imatrix --- .gitignore | 1 + 1 file changed, 1 insertion(+) diff --git a/.gitignore b/.gitignore index cf1b692e9..fba207045 100644 --- a/.gitignore +++ b/.gitignore @@ -43,6 +43,7 @@ models-mnt /embedding /gguf /gguf-llama-simple +/imatrix /infill /libllama.so /llama-bench From e790eef21ce659f5c16d59f8a5c8dcf6cde0692a Mon Sep 17 00:00:00 2001 From: Zay <95888118+isaiahbjork@users.noreply.github.com> Date: Fri, 12 Jan 2024 05:48:00 -0700 Subject: [PATCH 42/42] llama.swiftui : update models layout (#4826) * Updated Models Layout - Added a models drawer - Added downloading directly from Hugging Face - Load custom models from local folder - Delete models by swiping left * trimmed trailing white space * Updated Models Layout --- .../llama.swiftui.xcodeproj/project.pbxproj | 8 +- .../llama.swiftui/Models/LlamaState.swift | 89 ++++++++ .../llama.swiftui/UI/ContentView.swift | 213 +++++++++--------- .../llama.swiftui/UI/DownloadButton.swift | 2 + .../llama.swiftui/UI/InputButton.swift | 131 +++++++++++ 5 files changed, 338 insertions(+), 105 deletions(-) create mode 100644 examples/llama.swiftui/llama.swiftui/UI/InputButton.swift diff --git a/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj b/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj index a8848a49f..3950b9e9d 100644 --- a/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj +++ b/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj @@ -8,6 +8,7 @@ /* Begin PBXBuildFile section */ 549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; }; + 79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */; }; 7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; }; 8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */; }; 8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83782AC328BD0096AF73 /* ContentView.swift */; }; @@ -22,6 +23,7 @@ /* Begin PBXFileReference section */ 549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; }; + 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = InputButton.swift; sourceTree = ""; }; 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = ""; }; 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = llama.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; }; 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = llama_swiftuiApp.swift; sourceTree = ""; }; @@ -119,6 +121,7 @@ 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */, 8A1C83782AC328BD0096AF73 /* ContentView.swift */, F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */, + 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */, ); path = UI; sourceTree = ""; @@ -213,6 +216,7 @@ 8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */, 8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */, 7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */, + 79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */, ); runOnlyForDeploymentPostprocessing = 0; }; @@ -345,7 +349,7 @@ CLANG_ENABLE_MODULES = YES; CODE_SIGN_STYLE = Automatic; CURRENT_PROJECT_VERSION = 1; - DEVELOPMENT_TEAM = STLSG3FG8Q; + DEVELOPMENT_TEAM = K5UQJPP73A; ENABLE_PREVIEWS = YES; GENERATE_INFOPLIST_FILE = YES; INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES; @@ -377,7 +381,7 @@ CLANG_ENABLE_MODULES = YES; CODE_SIGN_STYLE = Automatic; CURRENT_PROJECT_VERSION = 1; - DEVELOPMENT_TEAM = STLSG3FG8Q; + DEVELOPMENT_TEAM = K5UQJPP73A; ENABLE_PREVIEWS = YES; GENERATE_INFOPLIST_FILE = YES; INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES; diff --git a/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift b/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift index 17cb5b9dd..5bde18917 100644 --- a/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift +++ b/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift @@ -1,9 +1,19 @@ import Foundation +struct Model: Identifiable { + var id = UUID() + var name: String + var url: String + var filename: String + var status: String? +} + @MainActor class LlamaState: ObservableObject { @Published var messageLog = "" @Published var cacheCleared = false + @Published var downloadedModels: [Model] = [] + @Published var undownloadedModels: [Model] = [] let NS_PER_S = 1_000_000_000.0 private var llamaContext: LlamaContext? @@ -13,23 +23,102 @@ class LlamaState: ObservableObject { } init() { + loadModelsFromDisk() + loadDefaultModels() + } + + private func loadModelsFromDisk() { + do { + let documentsURL = getDocumentsDirectory() + let modelURLs = try FileManager.default.contentsOfDirectory(at: documentsURL, includingPropertiesForKeys: nil, options: [.skipsHiddenFiles, .skipsSubdirectoryDescendants]) + for modelURL in modelURLs { + let modelName = modelURL.deletingPathExtension().lastPathComponent + downloadedModels.append(Model(name: modelName, url: "", filename: modelURL.lastPathComponent, status: "downloaded")) + } + } catch { + print("Error loading models from disk: \(error)") + } + } + + private func loadDefaultModels() { do { try loadModel(modelUrl: defaultModelUrl) } catch { messageLog += "Error!\n" } + + for model in defaultModels { + let fileURL = getDocumentsDirectory().appendingPathComponent(model.filename) + if FileManager.default.fileExists(atPath: fileURL.path) { + + } else { + var undownloadedModel = model + undownloadedModel.status = "download" + undownloadedModels.append(undownloadedModel) + } + } } + func getDocumentsDirectory() -> URL { + let paths = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask) + return paths[0] + } + private let defaultModels: [Model] = [ + Model(name: "TinyLlama-1.1B (Q4_0, 0.6 GiB)",url: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf", status: "download"), + Model( + name: "TinyLlama-1.1B Chat (Q8_0, 1.1 GiB)", + url: "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/tinyllama-1.1b-chat-v1.0.Q8_0.gguf?download=true", + filename: "tinyllama-1.1b-chat-v1.0.Q8_0.gguf", status: "download" + ), + + Model( + name: "TinyLlama-1.1B (F16, 2.2 GiB)", + url: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true", + filename: "tinyllama-1.1b-f16.gguf", status: "download" + ), + + Model( + name: "Phi-2.7B (Q4_0, 1.6 GiB)", + url: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true", + filename: "phi-2-q4_0.gguf", status: "download" + ), + + Model( + name: "Phi-2.7B (Q8_0, 2.8 GiB)", + url: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true", + filename: "phi-2-q8_0.gguf", status: "download" + ), + + Model( + name: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)", + url: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true", + filename: "mistral-7b-v0.1.Q4_0.gguf", status: "download" + ), + Model( + name: "OpenHermes-2.5-Mistral-7B (Q3_K_M, 3.52 GiB)", + url: "https://huggingface.co/TheBloke/OpenHermes-2.5-Mistral-7B-GGUF/resolve/main/openhermes-2.5-mistral-7b.Q3_K_M.gguf?download=true", + filename: "openhermes-2.5-mistral-7b.Q3_K_M.gguf", status: "download" + ) + ] func loadModel(modelUrl: URL?) throws { if let modelUrl { messageLog += "Loading model...\n" llamaContext = try LlamaContext.create_context(path: modelUrl.path()) messageLog += "Loaded model \(modelUrl.lastPathComponent)\n" + + // Assuming that the model is successfully loaded, update the downloaded models + updateDownloadedModels(modelName: modelUrl.lastPathComponent, status: "downloaded") } else { messageLog += "Load a model from the list below\n" } } + + private func updateDownloadedModels(modelName: String, status: String) { + undownloadedModels.removeAll { $0.name == modelName } + } + + func complete(text: String) async { guard let llamaContext else { return diff --git a/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift b/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift index 7c81ea256..30c2dc431 100644 --- a/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift +++ b/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift @@ -2,115 +2,57 @@ import SwiftUI struct ContentView: View { @StateObject var llamaState = LlamaState() - @State private var multiLineText = "" - - private static func cleanupModelCaches() { - // Delete all models (*.gguf) - let fileManager = FileManager.default - let documentsUrl = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0] - do { - let fileURLs = try fileManager.contentsOfDirectory(at: documentsUrl, includingPropertiesForKeys: nil) - for fileURL in fileURLs { - if fileURL.pathExtension == "gguf" { - try fileManager.removeItem(at: fileURL) - } - } - } catch { - print("Error while enumerating files \(documentsUrl.path): \(error.localizedDescription)") - } - } + @State private var showingHelp = false // To track if Help Sheet should be shown var body: some View { - VStack { - ScrollView(.vertical, showsIndicators: true) { - Text(llamaState.messageLog) - .font(.system(size: 12)) - .frame(maxWidth: .infinity, alignment: .leading) + NavigationView { + VStack { + ScrollView(.vertical, showsIndicators: true) { + Text(llamaState.messageLog) + .font(.system(size: 12)) + .frame(maxWidth: .infinity, alignment: .leading) + .padding() + .onTapGesture { + UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil) + } + } + + TextEditor(text: $multiLineText) + .frame(height: 80) + .padding() + .border(Color.gray, width: 0.5) + + HStack { + Button("Send") { + sendText() + } + + Button("Bench") { + bench() + } + + Button("Clear") { + clear() + } + + Button("Copy") { + UIPasteboard.general.string = llamaState.messageLog + } + } + .buttonStyle(.bordered) .padding() - .onTapGesture { - UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil) - } - } - TextEditor(text: $multiLineText) - .frame(height: 80) + NavigationLink(destination: DrawerView(llamaState: llamaState)) { + Text("View Models") + } .padding() - .border(Color.gray, width: 0.5) - HStack { - Button("Send") { - sendText() - } - - Button("Bench") { - bench() - } - - Button("Clear") { - clear() - } - - Button("Copy") { - UIPasteboard.general.string = llamaState.messageLog - } - }.buttonStyle(.bordered) - - VStack(alignment: .leading) { - DownloadButton( - llamaState: llamaState, - modelName: "TinyLlama-1.1B (Q4_0, 0.6 GiB)", - modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true", - filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf" - ) - - DownloadButton( - llamaState: llamaState, - modelName: "TinyLlama-1.1B (Q8_0, 1.1 GiB)", - modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q8_0.gguf?download=true", - filename: "tinyllama-1.1b-1t-openorca.Q8_0.gguf" - ) - - DownloadButton( - llamaState: llamaState, - modelName: "TinyLlama-1.1B (F16, 2.2 GiB)", - modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true", - filename: "tinyllama-1.1b-f16.gguf" - ) - - DownloadButton( - llamaState: llamaState, - modelName: "Phi-2.7B (Q4_0, 1.6 GiB)", - modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true", - filename: "phi-2-q4_0.gguf" - ) - - DownloadButton( - llamaState: llamaState, - modelName: "Phi-2.7B (Q8_0, 2.8 GiB)", - modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true", - filename: "phi-2-q8_0.gguf" - ) - - DownloadButton( - llamaState: llamaState, - modelName: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)", - modelUrl: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true", - filename: "mistral-7b-v0.1.Q4_0.gguf" - ) - - Button("Clear downloaded models") { - ContentView.cleanupModelCaches() - llamaState.cacheCleared = true - } - - LoadCustomButton(llamaState: llamaState) } - .padding(.top, 4) - .font(.system(size: 12)) - .frame(maxWidth: .infinity, alignment: .leading) + .padding() + .navigationBarTitle("Model Settings", displayMode: .inline) + } - .padding() } func sendText() { @@ -131,8 +73,73 @@ struct ContentView: View { await llamaState.clear() } } + struct DrawerView: View { + + @ObservedObject var llamaState: LlamaState + @State private var showingHelp = false + func delete(at offsets: IndexSet) { + offsets.forEach { offset in + let model = llamaState.downloadedModels[offset] + let fileURL = getDocumentsDirectory().appendingPathComponent(model.filename) + do { + try FileManager.default.removeItem(at: fileURL) + } catch { + print("Error deleting file: \(error)") + } + } + + // Remove models from downloadedModels array + llamaState.downloadedModels.remove(atOffsets: offsets) + } + + func getDocumentsDirectory() -> URL { + let paths = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask) + return paths[0] + } + var body: some View { + List { + Section(header: Text("Download Models From Hugging Face")) { + HStack { + InputButton(llamaState: llamaState) + } + } + Section(header: Text("Downloaded Models")) { + ForEach(llamaState.downloadedModels) { model in + DownloadButton(llamaState: llamaState, modelName: model.name, modelUrl: model.url, filename: model.filename) + } + .onDelete(perform: delete) + } + Section(header: Text("Default Models")) { + ForEach(llamaState.undownloadedModels) { model in + DownloadButton(llamaState: llamaState, modelName: model.name, modelUrl: model.url, filename: model.filename) + } + } + + } + .listStyle(GroupedListStyle()) + .navigationBarTitle("Model Settings", displayMode: .inline).toolbar { + ToolbarItem(placement: .navigationBarTrailing) { + Button("Help") { + showingHelp = true + } + } + }.sheet(isPresented: $showingHelp) { // Sheet for help modal + VStack(alignment: .leading) { + VStack(alignment: .leading) { + Text("1. Make sure the model is in GGUF Format") + .padding() + Text("2. Copy the download link of the quantized model") + .padding() + } + Spacer() + } + } + } + } } -//#Preview { -// ContentView() -//} +struct ContentView_Previews: PreviewProvider { + static var previews: some View { + ContentView() + } +} diff --git a/examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift b/examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift index c9f322ca1..4584d6eaa 100644 --- a/examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift +++ b/examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift @@ -53,6 +53,8 @@ struct DownloadButton: View { llamaState.cacheCleared = false + let model = Model(name: modelName, url: modelUrl, filename: filename, status: "downloaded") + llamaState.downloadedModels.append(model) status = "downloaded" } } catch let err { diff --git a/examples/llama.swiftui/llama.swiftui/UI/InputButton.swift b/examples/llama.swiftui/llama.swiftui/UI/InputButton.swift new file mode 100644 index 000000000..c5ffbad4e --- /dev/null +++ b/examples/llama.swiftui/llama.swiftui/UI/InputButton.swift @@ -0,0 +1,131 @@ +import SwiftUI + +struct InputButton: View { + @ObservedObject var llamaState: LlamaState + @State private var inputLink: String = "" + @State private var status: String = "download" + @State private var filename: String = "" + + @State private var downloadTask: URLSessionDownloadTask? + @State private var progress = 0.0 + @State private var observation: NSKeyValueObservation? + + private static func extractModelInfo(from link: String) -> (modelName: String, filename: String)? { + guard let url = URL(string: link), + let lastPathComponent = url.lastPathComponent.components(separatedBy: ".").first, + let modelName = lastPathComponent.components(separatedBy: "-").dropLast().joined(separator: "-").removingPercentEncoding, + let filename = lastPathComponent.removingPercentEncoding else { + return nil + } + + return (modelName, filename) + } + + private static func getFileURL(filename: String) -> URL { + FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename) + } + + private func download() { + guard let extractedInfo = InputButton.extractModelInfo(from: inputLink) else { + // Handle invalid link or extraction failure + return + } + + let (modelName, filename) = extractedInfo + self.filename = filename // Set the state variable + + status = "downloading" + print("Downloading model \(modelName) from \(inputLink)") + guard let url = URL(string: inputLink) else { return } + let fileURL = InputButton.getFileURL(filename: filename) + + downloadTask = URLSession.shared.downloadTask(with: url) { temporaryURL, response, error in + if let error = error { + print("Error: \(error.localizedDescription)") + return + } + + guard let response = response as? HTTPURLResponse, (200...299).contains(response.statusCode) else { + print("Server error!") + return + } + + do { + if let temporaryURL = temporaryURL { + try FileManager.default.copyItem(at: temporaryURL, to: fileURL) + print("Writing to \(filename) completed") + + llamaState.cacheCleared = false + + let model = Model(name: modelName, url: self.inputLink, filename: filename, status: "downloaded") + llamaState.downloadedModels.append(model) + status = "downloaded" + } + } catch let err { + print("Error: \(err.localizedDescription)") + } + } + + observation = downloadTask?.progress.observe(\.fractionCompleted) { progress, _ in + self.progress = progress.fractionCompleted + } + + downloadTask?.resume() + } + + var body: some View { + VStack { + HStack { + TextField("Paste Quantized Download Link", text: $inputLink) + .textFieldStyle(RoundedBorderTextFieldStyle()) + + Button(action: { + downloadTask?.cancel() + status = "download" + }) { + Text("Cancel") + } + } + + if status == "download" { + Button(action: download) { + Text("Download Custom Model") + } + } else if status == "downloading" { + Button(action: { + downloadTask?.cancel() + status = "download" + }) { + Text("Downloading \(Int(progress * 100))%") + } + } else if status == "downloaded" { + Button(action: { + let fileURL = InputButton.getFileURL(filename: self.filename) + if !FileManager.default.fileExists(atPath: fileURL.path) { + download() + return + } + do { + try llamaState.loadModel(modelUrl: fileURL) + } catch let err { + print("Error: \(err.localizedDescription)") + } + }) { + Text("Load Custom Model") + } + } else { + Text("Unknown status") + } + } + .onDisappear() { + downloadTask?.cancel() + } + .onChange(of: llamaState.cacheCleared) { newValue in + if newValue { + downloadTask?.cancel() + let fileURL = InputButton.getFileURL(filename: self.filename) + status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download" + } + } + } +}