vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge remote-tracking branch 'origin/master' into 0cc4m/vulkan-moe

Browse source
This commit is contained in:
0cc4m 2024-05-29 22:42:59 +02:00
commit 8ecdda1e5e
72 changed files with 2766 additions and 777 deletions
Download patch file Download diff file Expand all files Collapse all files

50
.github/ISSUE_TEMPLATE/01-bug-low.yml vendored Normal file
View file

@ -0,0 +1,50 @@
name: Low Severity Bugs
description: Used to report low severity bugs in llama.cpp (e.g. cosmetic issues, non critical UI glitches)
title: "Bug: "
labels: ["bug-unconfirmed", "low severity"]
body:
- type: markdown
attributes:
value: |
Thanks for taking the time to fill out this bug report!
Please include information about your system, the steps to reproduce the bug,
and the version of llama.cpp that you are using.
If possible, please provide a minimal code example that reproduces the bug.
- type: textarea
id: what-happened
attributes:
label: What happened?
description: Also tell us, what did you expect to happen?
placeholder: Tell us what you see!
validations:
required: true
- type: textarea
id: version
attributes:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./main --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: What operating system are you seeing the problem on?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
render: shell

50
.github/ISSUE_TEMPLATE/02-bug-medium.yml vendored Normal file
View file

@ -0,0 +1,50 @@
name: Medium Severity Bug
description: Used to report medium severity bugs in llama.cpp (e.g. Malfunctioning Features but generally still useable)
title: "Bug: "
labels: ["bug-unconfirmed", "medium severity"]
body:
- type: markdown
attributes:
value: |
Thanks for taking the time to fill out this bug report!
Please include information about your system, the steps to reproduce the bug,
and the version of llama.cpp that you are using.
If possible, please provide a minimal code example that reproduces the bug.
- type: textarea
id: what-happened
attributes:
label: What happened?
description: Also tell us, what did you expect to happen?
placeholder: Tell us what you see!
validations:
required: true
- type: textarea
id: version
attributes:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./main --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: What operating system are you seeing the problem on?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
render: shell

50
.github/ISSUE_TEMPLATE/03-bug-high.yml vendored Normal file
View file

@ -0,0 +1,50 @@
name: High Severity Bug
description: Used to report high severity bugs in llama.cpp (e.g. Malfunctioning features hindering important common workflow)
title: "Bug: "
labels: ["bug-unconfirmed", "high severity"]
body:
- type: markdown
attributes:
value: |
Thanks for taking the time to fill out this bug report!
Please include information about your system, the steps to reproduce the bug,
and the version of llama.cpp that you are using.
If possible, please provide a minimal code example that reproduces the bug.
- type: textarea
id: what-happened
attributes:
label: What happened?
description: Also tell us, what did you expect to happen?
placeholder: Tell us what you see!
validations:
required: true
- type: textarea
id: version
attributes:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./main --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: What operating system are you seeing the problem on?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
render: shell

50
.github/ISSUE_TEMPLATE/04-bug-critical.yml vendored Normal file
View file

@ -0,0 +1,50 @@
name: Critical Severity Bug
description: Used to report critical severity bugs in llama.cpp (e.g. Crashing, Corrupted, Dataloss)
title: "Bug: "
labels: ["bug-unconfirmed", "critical severity"]
body:
- type: markdown
attributes:
value: |
Thanks for taking the time to fill out this bug report!
Please include information about your system, the steps to reproduce the bug,
and the version of llama.cpp that you are using.
If possible, please provide a minimal code example that reproduces the bug.
- type: textarea
id: what-happened
attributes:
label: What happened?
description: Also tell us, what did you expect to happen?
placeholder: Tell us what you see!
validations:
required: true
- type: textarea
id: version
attributes:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./main --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: What operating system are you seeing the problem on?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
render: shell

51
.github/ISSUE_TEMPLATE/05-enhancement.yml vendored Normal file
View file

@ -0,0 +1,51 @@
name: Enhancement
description: Used to request enhancements for llama.cpp
title: "Feature Request: "
labels: ["enhancement"]
body:
- type: markdown
attributes:
value: |
[Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggerganov/llama.cpp/discussions/categories/ideas)
- type: checkboxes
id: prerequisites
attributes:
label: Prerequisites
description: Please confirm the following before submitting your enhancement request.
options:
- label: I am running the latest code. Mention the version if possible as well.
required: true
- label: I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
required: true
- label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
required: true
- label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new and useful enhancement to share.
required: true
- type: textarea
id: feature-description
attributes:
label: Feature Description
description: Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
placeholder: Detailed description of the enhancement
validations:
required: true
- type: textarea
id: motivation
attributes:
label: Motivation
description: Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
placeholder: Explanation of why this feature is needed and its benefits
validations:
required: true
- type: textarea
id: possible-implementation
attributes:
label: Possible Implementation
description: If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.
placeholder: Detailed description of potential implementation
validations:
required: false

38
.github/ISSUE_TEMPLATE/06-question.yml vendored Normal file
View file

@ -0,0 +1,38 @@
name: Question
description: Used to ask questions about llama.cpp
title: "Question: "
labels: ["question"]
body:
- type: markdown
attributes:
value: |
[Please search your question first in Discussion if you got a common general question.](https://github.com/ggerganov/llama.cpp/discussions/categories/q-a)
- type: checkboxes
id: prerequisites
attributes:
label: Prerequisites
description: Please confirm the following before submitting your question.
options:
- label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
required: true
- label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new useful question to share that cannot be answered within Discussions.
required: true
- type: textarea
id: background-description
attributes:
label: Background Description
description: Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an question.
placeholder: Detailed description of your question
validations:
required: true
- type: textarea
id: possible-answer
attributes:
label: Possible Answer
description: If you have some idea of possible answers you want to confirm, that would also be appreciated.
placeholder: Your idea of possible answers
validations:
required: false

28
.github/ISSUE_TEMPLATE/07-refactor.yml vendored Normal file
View file

@ -0,0 +1,28 @@
name: Refactor (Maintainers)
description: Used to track refactoring opportunities
title: "Refactor: "
labels: ["refactor"]
body:
- type: markdown
attributes:
value: |
Don't forget to [check for existing refactor issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered.
Also you may want to check [Pull request refactor label as well](https://github.com/ggerganov/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too.
- type: textarea
id: background-description
attributes:
label: Background Description
description: Please provide a detailed written description of the pain points you are trying to solve.
placeholder: Detailed description behind your motivation to request refactor
validations:
required: true
- type: textarea
id: possible-approaches
attributes:
label: Possible Refactor Approaches
description: If you have some idea of possible approaches to solve this problem. You may want to make it a todo list.
placeholder: Your idea of possible refactoring opportunity/approaches
validations:
required: false

11
.github/ISSUE_TEMPLATE/bug.md vendored
View file

@ -1,11 +0,0 @@
---
name: Bug template
about: Used to report bugs in llama.cpp
labels: ["bug-unconfirmed"]
assignees: ''
---
Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.
If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

28
.github/ISSUE_TEMPLATE/enhancement.md vendored
View file

@ -1,28 +0,0 @@
---
name: Enhancement template
about: Used to request enhancements for llama.cpp
labels: ["enhancement"]
assignees: ''
---
# Prerequisites
Please answer the following questions for yourself before submitting an issue.
- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
# Feature Description
Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
# Motivation
Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
# Possible Implementation
If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.

14
.github/labeler.yml vendored
View file

@ -1,5 +1,16 @@
# https://github.com/actions/labeler
Kompute:
- changed-files:
- any-glob-to-any-file:
- ggml-kompute.h
- ggml-kompute.cpp
- README-kompute.md
Apple Metal:
- changed-files:
- any-glob-to-any-file:
- ggml-metal.h
- ggml-metal.cpp
- README-metal.md
SYCL:
- changed-files:
- any-glob-to-any-file:
@ -9,6 +20,7 @@ SYCL:
Nvidia GPU:
- changed-files:
- any-glob-to-any-file:
- ggml-cuda.h
- ggml-cuda/**
Vulkan:
- changed-files:

8
CMakeLists.txt
View file

@ -72,6 +72,7 @@ else()
set(INS_ENB ON)
endif()
option(LLAMA_SVE "llama: enable SVE" OFF)
option(LLAMA_AVX "llama: enable AVX" ${INS_ENB})
option(LLAMA_AVX2 "llama: enable AVX2" ${INS_ENB})
option(LLAMA_AVX512 "llama: enable AVX512" OFF)
@ -627,6 +628,10 @@ if (LLAMA_SYCL)
add_compile_definitions(GGML_SYCL_F16)
endif()
if (LLAMA_CUDA_FORCE_MMQ)
add_compile_definitions(GGML_SYCL_FORCE_MMQ)
endif()
add_compile_options(-I./) #include DPCT
add_compile_options(-I/${SYCL_INCLUDE_DIR})
@ -1040,6 +1045,9 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
# Raspberry Pi 3, 4, Zero 2 (32-bit)
list(APPEND ARCH_FLAGS -mno-unaligned-access)
endif()
if (LLAMA_SVE)
list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
endif()
endif()
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND

8
CMakePresets.json
View file

@ -1,4 +1,4 @@
{
{
"version": 4,
"configurePresets": [
{
@ -40,6 +40,10 @@
{ "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "release" ] },
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "release", "static" ] }
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "release", "static" ] },
{ "name": "x64-windows-msvc-debug" , "inherits": [ "base", "debug" ] },
{ "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] },
{ "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] }
]
}

3
Makefile
View file

@ -441,6 +441,9 @@ endif # JETSON_EOL_MODULE_DETECT
ifdef LLAMA_DEBUG
MK_NVCCFLAGS += -lineinfo
endif # LLAMA_DEBUG
ifdef LLAMA_CUDA_DEBUG
MK_NVCCFLAGS += --device-debug
endif # LLAMA_CUDA_DEBUG
ifdef LLAMA_CUDA_NVCC
NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC)
else

6
README-sycl.md
View file

@ -55,8 +55,8 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
## OS
| OS | Status | Verified |
|---------|---------|------------------------------------|
| Linux | Support | Ubuntu 22.04, Fedora Silverblue 39 |
|---------|---------|------------------------------------------------|
| Linux | Support | Ubuntu 22.04, Fedora Silverblue 39, Arch Linux |
| Windows | Support | Windows 11 |
@ -70,7 +70,7 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
|-------------------------------|---------|---------------------------------------|
| Intel Data Center Max Series | Support | Max 1550, 1100 |
| Intel Data Center Flex Series | Support | Flex 170 |
| Intel Arc Series | Support | Arc 770, 730M |
| Intel Arc Series | Support | Arc 770, 730M, Arc A750 |
| Intel built-in Arc GPU | Support | built-in Arc GPU in Meteor Lake |
| Intel iGPU | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |

5
README.md
View file

@ -203,6 +203,10 @@ Unless otherwise noted these projects are open-source with permissive licensing:
*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
**Tools:**
- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
---
Here is a typical run using LLaMA v2 13B on M2 Ultra:
@ -474,6 +478,7 @@ Building the program with BLAS support may lead to some performance improvements
| LLAMA_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
| LLAMA_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| LLAMA_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
| LLAMA_CUDA_FORCE_MMQ | Boolean | false | Force the use of dequantization + matrix multiplication kernels instead of leveraging Math libraries. | |
| LLAMA_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
| LLAMA_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
| LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |

6
common/common.cpp
View file

@ -904,6 +904,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
params.interactive_specials = true;
return true;
}
if (arg == "--special") {
params.special = true;
return true;
}
if (arg == "--embedding") {
params.embedding = true;
return true;
@ -1362,6 +1366,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
printf(" -h, --help show this help message and exit\n");
printf(" --version show version and build info\n");
printf(" -i, --interactive run in interactive mode\n");
printf(" --special special tokens output enabled\n");
printf(" --interactive-specials allow special tokens in user text, in interactive mode\n");
printf(" --interactive-first run in interactive mode and wait for input right away\n");
printf(" -cnv, --conversation run in conversation mode (does not print special tokens and suffix/prefix)\n");
@ -2844,6 +2849,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
fprintf(stream, "cpu_has_sve: %s\n", ggml_cpu_has_sve() ? "true" : "false");
fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");

1
common/common.h
View file

@ -146,6 +146,7 @@ struct gpt_params {
bool use_color = false; // use color to distinguish generations and inputs
bool interactive = false; // interactive mode
bool interactive_specials = false; // whether to allow special tokens from user, during interactive mode
bool special = false; // enable special token output
bool conversation = false; // conversation mode (does not print special tokens and suffix/prefix)
bool chatml = false; // chatml mode (used for models trained on chatml syntax)
bool prompt_cache_all = false; // save user input and generations to prompt cache

2
common/train.cpp
View file

@ -1052,7 +1052,7 @@ struct train_params_common get_default_train_params_common() {
params.custom_n_ctx = false;
params.use_flash = true;
params.use_flash = false;
params.use_checkpointing = true;
params.sample_start = "";

1
convert-hf-to-gguf-update.py
View file

@ -81,6 +81,7 @@ models = [
{"name": "jina-v2-en", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
{"name": "jina-v2-es", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
{"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
{"name": "smaug-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
]

100
convert-hf-to-gguf.py
View file

@ -473,6 +473,9 @@ class Model:
if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6":
# ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de
res = "jina-v2-de"
if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
# ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
res = "smaug-bpe"
if res is None:
logger.warning("\n")
@ -1314,6 +1317,17 @@ class LlamaModel(Model):
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
if tokenizer_config_file.is_file():
with open(tokenizer_config_file, "r", encoding="utf-8") as f:
tokenizer_config_json = json.load(f)
if "add_prefix_space" in tokenizer_config_json:
self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
# Apply to granite small models only
if self.hparams.get("vocab_size", 32000) == 49152:
self.gguf_writer.add_add_bos_token(False)
@staticmethod
def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
if n_head_kv is not None and n_head != n_head_kv:
@ -1328,9 +1342,9 @@ class LlamaModel(Model):
n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams.get("num_key_value_heads")
if name.endswith("q_proj.weight"):
if name.endswith(("q_proj.weight", "q_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith("k_proj.weight"):
if name.endswith(("k_proj.weight", "k_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
# process the experts separately
@ -2392,7 +2406,8 @@ class CommandR2Model(Model):
# max_position_embeddings = 8192 in config.json but model was actually
# trained on 128k context length
self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
# aya-23 models don't have model_max_length specified
self.hparams["max_position_embeddings"] = self.find_hparam(["model_max_length", "max_position_embeddings"])
def set_gguf_parameters(self):
super().set_gguf_parameters()
@ -2616,6 +2631,85 @@ class ArcticModel(Model):
raise ValueError(f"Unprocessed experts: {experts}")
@Model.register("DeepseekV2ForCausalLM")
class DeepseekV2Model(Model):
model_arch = gguf.MODEL_ARCH.DEEPSEEK2
def set_vocab(self):
self._set_vocab_gpt2()
def set_gguf_parameters(self):
super().set_gguf_parameters()
hparams = self.hparams
self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
self.gguf_writer.add_vocab_size(hparams["vocab_size"])
if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
self.gguf_writer.add_value_length(hparams["v_head_dim"])
self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
if self.hparams["rope_scaling"].get("type") == "yarn":
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * hparams["rope_scaling"]["mscale_all_dim"])
_experts: list[dict[str, Tensor]] | None = None
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# process the experts separately
if name.find("mlp.experts") != -1:
n_experts = self.hparams["n_routed_experts"]
assert bid is not None
if self._experts is None:
self._experts = [{} for _ in range(self.block_count)]
self._experts[bid][name] = data_torch
if len(self._experts[bid]) >= n_experts * 3:
tensors: list[tuple[str, Tensor]] = []
# merge the experts into a single 3d tensor
for w_name in ["down_proj", "gate_proj", "up_proj"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
datas.append(self._experts[bid][ename])
del self._experts[bid][ename]
data_torch = torch.stack(datas, dim=0)
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
new_name = self.map_tensor_name(merged_name)
tensors.append((new_name, data_torch))
return tensors
else:
return []
return [(self.map_tensor_name(name), data_torch)]
def write_tensors(self):
super().write_tensors()
if self._experts is not None:
# flatten `list[dict[str, Tensor]]` into `list[str]`
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
###### CONVERSION LOGIC ######

2
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -774,7 +774,7 @@ static struct train_params get_default_train_params() {
params.samples_start_after_nl = false;
params.use_adam = true;
params.use_flash = true;
params.use_flash = false;
params.use_scratch = true;
// only adam

28
examples/llama-bench/llama-bench.cpp
View file

@ -178,6 +178,7 @@ struct cmd_params {
std::vector<ggml_type> type_v;
std::vector<int> n_threads;
std::vector<int> n_gpu_layers;
std::vector<std::string> rpc_servers;
std::vector<llama_split_mode> split_mode;
std::vector<int> main_gpu;
std::vector<bool> no_kv_offload;
@ -202,6 +203,7 @@ static const cmd_params cmd_params_defaults = {
/* type_v */ {GGML_TYPE_F16},
/* n_threads */ {cpu_get_num_math()},
/* n_gpu_layers */ {99},
/* rpc_servers */ {""},
/* split_mode */ {LLAMA_SPLIT_MODE_LAYER},
/* main_gpu */ {0},
/* no_kv_offload */ {false},
@ -230,6 +232,7 @@ static void print_usage(int /* argc */, char ** argv) {
printf(" -ctv, --cache-type-v <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
printf(" -t, --threads <n> (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
printf(" -ngl, --n-gpu-layers <n> (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
printf(" -rpc, --rpc <rpc_servers> (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
printf(" -sm, --split-mode <none|layer|row> (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
printf(" -mg, --main-gpu <i> (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
@ -384,6 +387,12 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
}
auto p = split<int>(argv[i], split_delim);
params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
} else if (arg == "-rpc" || arg == "--rpc") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rpc_servers.push_back(argv[i]);
} else if (arg == "-sm" || arg == "--split-mode") {
if (++i >= argc) {
invalid_param = true;
@ -519,6 +528,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
if (params.type_k.empty()) { params.type_k = cmd_params_defaults.type_k; }
if (params.type_v.empty()) { params.type_v = cmd_params_defaults.type_v; }
if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
if (params.rpc_servers.empty()) { params.rpc_servers = cmd_params_defaults.rpc_servers; }
if (params.split_mode.empty()) { params.split_mode = cmd_params_defaults.split_mode; }
if (params.main_gpu.empty()) { params.main_gpu = cmd_params_defaults.main_gpu; }
if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
@ -541,6 +551,7 @@ struct cmd_params_instance {
ggml_type type_v;
int n_threads;
int n_gpu_layers;
std::string rpc_servers;
llama_split_mode split_mode;
int main_gpu;
bool no_kv_offload;
@ -553,6 +564,9 @@ struct cmd_params_instance {
llama_model_params mparams = llama_model_default_params();
mparams.n_gpu_layers = n_gpu_layers;
if (!rpc_servers.empty()) {
mparams.rpc_servers = rpc_servers.c_str();
}
mparams.split_mode = split_mode;
mparams.main_gpu = main_gpu;
mparams.tensor_split = tensor_split.data();
@ -564,6 +578,7 @@ struct cmd_params_instance {
bool equal_mparams(const cmd_params_instance & other) const {
return model == other.model &&
n_gpu_layers == other.n_gpu_layers &&
rpc_servers == other.rpc_servers &&
split_mode == other.split_mode &&
main_gpu == other.main_gpu &&
use_mmap == other.use_mmap &&
@ -592,6 +607,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
// this ordering minimizes the number of times that each model needs to be reloaded
for (const auto & m : params.model)
for (const auto & nl : params.n_gpu_layers)
for (const auto & rpc : params.rpc_servers)
for (const auto & sm : params.split_mode)
for (const auto & mg : params.main_gpu)
for (const auto & ts : params.tensor_split)
@ -618,6 +634,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .rpc_servers = */ rpc,
/* .split_mode = */ sm,
/* .main_gpu = */ mg,
/* .no_kv_offload= */ nkvo,
@ -643,6 +660,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .rpc_servers = */ rpc,
/* .split_mode = */ sm,
/* .main_gpu = */ mg,
/* .no_kv_offload= */ nkvo,
@ -668,6 +686,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .rpc_servers = */ rpc,
/* .split_mode = */ sm,
/* .main_gpu = */ mg,
/* .no_kv_offload= */ nkvo,
@ -692,6 +711,7 @@ struct test {
static const bool kompute;
static const bool metal;
static const bool sycl;
static const bool rpc;
static const bool gpu_blas;
static const bool blas;
static const std::string cpu_info;
@ -790,6 +810,9 @@ struct test {
if (sycl) {
return GGML_SYCL_NAME;
}
if (rpc) {
return "RPC";
}
if (gpu_blas) {
return "GPU BLAS";
}
@ -803,7 +826,7 @@ struct test {
static const std::vector<std::string> & get_fields() {
static const std::vector<std::string> fields = {
"build_commit", "build_number",
"cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas",
"cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", "blas",
"cpu_info", "gpu_info",
"model_filename", "model_type", "model_size", "model_n_params",
"n_batch", "n_ubatch",
@ -859,7 +882,7 @@ struct test {
std::vector<std::string> values = {
build_commit, std::to_string(build_number),
std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(vulkan),
std::to_string(metal), std::to_string(sycl), std::to_string(gpu_blas), std::to_string(blas),
std::to_string(metal), std::to_string(sycl), std::to_string(rpc), std::to_string(gpu_blas), std::to_string(blas),
cpu_info, gpu_info,
model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
std::to_string(n_batch), std::to_string(n_ubatch),
@ -894,6 +917,7 @@ const bool test::metal = !!ggml_cpu_has_metal();
const bool test::gpu_blas = !!ggml_cpu_has_gpublas();
const bool test::blas = !!ggml_cpu_has_blas();
const bool test::sycl = !!ggml_cpu_has_sycl();
const bool test::rpc = !!ggml_cpu_has_rpc();
const std::string test::cpu_info = get_cpu_info();
const std::string test::gpu_info = get_gpu_info();

25
examples/llama.android/app/build.gradle.kts
View file

@ -7,8 +7,6 @@ android {
namespace = "com.example.llama"
compileSdk = 34
ndkVersion = "26.1.10909125"
defaultConfig {
applicationId = "com.example.llama"
minSdk = 33
@ -20,17 +18,6 @@ android {
vectorDrawables {
useSupportLibrary = true
}
ndk {
// Add NDK properties if wanted, e.g.
// abiFilters += listOf("arm64-v8a")
}
externalNativeBuild {
cmake {
arguments += "-DCMAKE_BUILD_TYPE=Release"
cppFlags += listOf()
arguments += listOf()
}
}
}
buildTypes {
@ -55,17 +42,6 @@ android {
composeOptions {
kotlinCompilerExtensionVersion = "1.5.1"
}
packaging {
resources {
excludes += "/META-INF/{AL2.0,LGPL2.1}"
}
}
externalNativeBuild {
cmake {
path = file("src/main/cpp/CMakeLists.txt")
version = "3.22.1"
}
}
}
dependencies {
@ -78,6 +54,7 @@ dependencies {
implementation("androidx.compose.ui:ui-graphics")
implementation("androidx.compose.ui:ui-tooling-preview")
implementation("androidx.compose.material3:material3")
implementation(project(":llama"))
testImplementation("junit:junit:4.13.2")
androidTestImplementation("androidx.test.ext:junit:1.1.5")
androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")

13
examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt
View file

@ -1,5 +1,6 @@
package com.example.llama
import android.llama.cpp.LLamaAndroid
import android.util.Log
import androidx.compose.runtime.getValue
import androidx.compose.runtime.mutableStateOf
@ -9,7 +10,7 @@ import androidx.lifecycle.viewModelScope
import kotlinx.coroutines.flow.catch
import kotlinx.coroutines.launch
class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
class MainViewModel(private val llamaAndroid: LLamaAndroid = LLamaAndroid.instance()): ViewModel() {
companion object {
@JvmStatic
private val NanosPerSecond = 1_000_000_000.0
@ -28,7 +29,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
viewModelScope.launch {
try {
llm.unload()
llamaAndroid.unload()
} catch (exc: IllegalStateException) {
messages += exc.message!!
}
@ -44,7 +45,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
messages += ""
viewModelScope.launch {
llm.send(text)
llamaAndroid.send(text)
.catch {
Log.e(tag, "send() failed", it)
messages += it.message!!
@ -57,7 +58,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
viewModelScope.launch {
try {
val start = System.nanoTime()
val warmupResult = llm.bench(pp, tg, pl, nr)
val warmupResult = llamaAndroid.bench(pp, tg, pl, nr)
val end = System.nanoTime()
messages += warmupResult
@ -70,7 +71,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
return@launch
}
messages += llm.bench(512, 128, 1, 3)
messages += llamaAndroid.bench(512, 128, 1, 3)
} catch (exc: IllegalStateException) {
Log.e(tag, "bench() failed", exc)
messages += exc.message!!
@ -81,7 +82,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
fun load(pathToModel: String) {
viewModelScope.launch {
try {
llm.load(pathToModel)
llamaAndroid.load(pathToModel)
messages += "Loaded $pathToModel"
} catch (exc: IllegalStateException) {
Log.e(tag, "load() failed", exc)

1
examples/llama.android/build.gradle.kts
View file

@ -2,4 +2,5 @@
plugins {
id("com.android.application") version "8.2.0" apply false
id("org.jetbrains.kotlin.android") version "1.9.0" apply false
id("com.android.library") version "8.2.0" apply false
}

1
examples/llama.android/llama/.gitignore vendored Normal file
View file

@ -0,0 +1 @@
/build

0
examples/llama.android/app/src/main/cpp/CMakeLists.txt → examples/llama.android/llama/CMakeLists.txt
View file

68
examples/llama.android/llama/build.gradle.kts Normal file
View file

@ -0,0 +1,68 @@
plugins {
id("com.android.library")
id("org.jetbrains.kotlin.android")
}
android {
namespace = "android.llama.cpp"
compileSdk = 34
defaultConfig {
minSdk = 33
testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"
consumerProguardFiles("consumer-rules.pro")
ndk {
// Add NDK properties if wanted, e.g.
// abiFilters += listOf("arm64-v8a")
}
externalNativeBuild {
cmake {
arguments += "-DCMAKE_BUILD_TYPE=Release"
cppFlags += listOf()
arguments += listOf()
cppFlags("")
}
}
}
buildTypes {
release {
isMinifyEnabled = false
proguardFiles(
getDefaultProguardFile("proguard-android-optimize.txt"),
"proguard-rules.pro"
)
}
}
externalNativeBuild {
cmake {
path("src/main/cpp/CMakeLists.txt")
version = "3.22.1"
}
}
compileOptions {
sourceCompatibility = JavaVersion.VERSION_1_8
targetCompatibility = JavaVersion.VERSION_1_8
}
kotlinOptions {
jvmTarget = "1.8"
}
packaging {
resources {
excludes += "/META-INF/{AL2.0,LGPL2.1}"
}
}
}
dependencies {
implementation("androidx.core:core-ktx:1.12.0")
implementation("androidx.appcompat:appcompat:1.6.1")
implementation("com.google.android.material:material:1.11.0")
testImplementation("junit:junit:4.13.2")
androidTestImplementation("androidx.test.ext:junit:1.1.5")
androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
}

0
examples/llama.android/llama/consumer-rules.pro Normal file
View file

21
examples/llama.android/llama/proguard-rules.pro vendored Normal file
View file

@ -0,0 +1,21 @@
# Add project specific ProGuard rules here.
# You can control the set of applied configuration files using the
# proguardFiles setting in build.gradle.
#
# For more details, see
# http://developer.android.com/guide/developing/tools/proguard.html
# If your project uses WebView with JS, uncomment the following
# and specify the fully qualified class name to the JavaScript interface
# class:
#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
# public *;
#}
# Uncomment this to preserve the line number information for
# debugging stack traces.
#-keepattributes SourceFile,LineNumberTable
# If you keep the line number information, uncomment this to
# hide the original source file name.
#-renamesourcefileattribute SourceFile

24
examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt Normal file
View file

@ -0,0 +1,24 @@
package android.llama.cpp
import androidx.test.platform.app.InstrumentationRegistry
import androidx.test.ext.junit.runners.AndroidJUnit4
import org.junit.Test
import org.junit.runner.RunWith
import org.junit.Assert.*
/**
* Instrumented test, which will execute on an Android device.
*
* See [testing documentation](http://d.android.com/tools/testing).
*/
@RunWith(AndroidJUnit4::class)
class ExampleInstrumentedTest {
@Test
fun useAppContext() {
// Context of the app under test.
val appContext = InstrumentationRegistry.getInstrumentation().targetContext
assertEquals("android.llama.cpp.test", appContext.packageName)
}
}

4
examples/llama.android/llama/src/main/AndroidManifest.xml Normal file
View file

@ -0,0 +1,4 @@
<?xml version="1.0" encoding="utf-8"?>
<manifest xmlns:android="http://schemas.android.com/apk/res/android">
</manifest>

49
examples/llama.android/llama/src/main/cpp/CMakeLists.txt Normal file
View file

@ -0,0 +1,49 @@
# For more information about using CMake with Android Studio, read the
# documentation: https://d.android.com/studio/projects/add-native-code.html.
# For more examples on how to use CMake, see https://github.com/android/ndk-samples.
# Sets the minimum CMake version required for this project.
cmake_minimum_required(VERSION 3.22.1)
# Declares the project name. The project name can be accessed via ${ PROJECT_NAME},
# Since this is the top level CMakeLists.txt, the project name is also accessible
# with ${CMAKE_PROJECT_NAME} (both CMake variables are in-sync within the top level
# build script scope).
project("llama-android")
include(FetchContent)
FetchContent_Declare(
llama
GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
GIT_TAG master
)
# Also provides "common"
FetchContent_MakeAvailable(llama)
# Creates and names a library, sets it as either STATIC
# or SHARED, and provides the relative paths to its source code.
# You can define multiple libraries, and CMake builds them for you.
# Gradle automatically packages shared libraries with your APK.
#
# In this top level CMakeLists.txt, ${CMAKE_PROJECT_NAME} is used to define
# the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME}
# is preferred for the same purpose.
#
# In order to load a library into your app from Java/Kotlin, you must call
# System.loadLibrary() and pass the name of the library defined here;
# for GameActivity/NativeActivity derived applications, the same library name must be
# used in the AndroidManifest.xml file.
add_library(${CMAKE_PROJECT_NAME} SHARED
# List C/C++ source files with relative paths to this CMakeLists.txt.
llama-android.cpp)
# Specifies libraries CMake should link to your target library. You
# can link libraries from various origins, such as libraries defined in this
# build script, prebuilt third-party libraries, or Android system libraries.
target_link_libraries(${CMAKE_PROJECT_NAME}
# List libraries link to the target library
llama
common
android
log)

28
examples/llama.android/app/src/main/cpp/llama-android.cpp → examples/llama.android/llama/src/main/cpp/llama-android.cpp
View file

@ -81,7 +81,7 @@ static void log_callback(ggml_log_level level, const char * fmt, void * data) {
extern "C"
JNIEXPORT jlong JNICALL
Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring filename) {
llama_model_params model_params = llama_model_default_params();
auto path_to_model = env->GetStringUTFChars(filename, 0);
@ -101,13 +101,13 @@ Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
extern "C"
JNIEXPORT void JNICALL
Java_com_example_llama_Llm_free_1model(JNIEnv *, jobject, jlong model) {
Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
llama_free_model(reinterpret_cast<llama_model *>(model));
}
extern "C"
JNIEXPORT jlong JNICALL
Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmodel) {
auto model = reinterpret_cast<llama_model *>(jmodel);
if (!model) {
@ -139,25 +139,25 @@ Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
extern "C"
JNIEXPORT void JNICALL
Java_com_example_llama_Llm_free_1context(JNIEnv *, jobject, jlong context) {
Java_android_llama_cpp_LLamaAndroid_free_1context(JNIEnv *, jobject, jlong context) {
llama_free(reinterpret_cast<llama_context *>(context));
}
extern "C"
JNIEXPORT void JNICALL
Java_com_example_llama_Llm_backend_1free(JNIEnv *, jobject) {
Java_android_llama_cpp_LLamaAndroid_backend_1free(JNIEnv *, jobject) {
llama_backend_free();
}
extern "C"
JNIEXPORT void JNICALL
Java_com_example_llama_Llm_log_1to_1android(JNIEnv *, jobject) {
Java_android_llama_cpp_LLamaAndroid_log_1to_1android(JNIEnv *, jobject) {
llama_log_set(log_callback, NULL);
}
extern "C"
JNIEXPORT jstring JNICALL
Java_com_example_llama_Llm_bench_1model(
Java_android_llama_cpp_LLamaAndroid_bench_1model(
JNIEnv *env,
jobject,
jlong context_pointer,
@ -271,13 +271,13 @@ Java_com_example_llama_Llm_bench_1model(
extern "C"
JNIEXPORT void JNICALL
Java_com_example_llama_Llm_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
}
extern "C"
JNIEXPORT jlong JNICALL
Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
// Source: Copy of llama.cpp:llama_batch_init but heap-allocated.
@ -313,19 +313,19 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
extern "C"
JNIEXPORT void JNICALL
Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) {
llama_backend_init();
}
extern "C"
JNIEXPORT jstring JNICALL
Java_com_example_llama_Llm_system_1info(JNIEnv *env, jobject) {
Java_android_llama_cpp_LLamaAndroid_system_1info(JNIEnv *env, jobject) {
return env->NewStringUTF(llama_print_system_info());
}
extern "C"
JNIEXPORT jint JNICALL
Java_com_example_llama_Llm_completion_1init(
Java_android_llama_cpp_LLamaAndroid_completion_1init(
JNIEnv *env,
jobject,
jlong context_pointer,
@ -376,7 +376,7 @@ Java_com_example_llama_Llm_completion_1init(
extern "C"
JNIEXPORT jstring JNICALL
Java_com_example_llama_Llm_completion_1loop(
Java_android_llama_cpp_LLamaAndroid_completion_1loop(
JNIEnv * env,
jobject,
jlong context_pointer,
@ -438,6 +438,6 @@ Java_com_example_llama_Llm_completion_1loop(
extern "C"
JNIEXPORT void JNICALL
Java_com_example_llama_Llm_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
}

8
examples/llama.android/app/src/main/java/com/example/llama/Llm.kt → examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt
View file

@ -1,4 +1,4 @@
package com.example.llama
package android.llama.cpp
import android.util.Log
import kotlinx.coroutines.CoroutineDispatcher
@ -10,7 +10,7 @@ import kotlinx.coroutines.withContext
import java.util.concurrent.Executors
import kotlin.concurrent.thread
class Llm {
class LLamaAndroid {
private val tag: String? = this::class.simpleName
private val threadLocalState: ThreadLocal<State> = ThreadLocal.withInitial { State.Idle }
@ -165,8 +165,8 @@ class Llm {
}
// Enforce only one instance of Llm.
private val _instance: Llm = Llm()
private val _instance: LLamaAndroid = LLamaAndroid()
fun instance(): Llm = _instance
fun instance(): LLamaAndroid = _instance
}
}

17
examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt Normal file
View file

@ -0,0 +1,17 @@
package android.llama.cpp
import org.junit.Test
import org.junit.Assert.*
/**
* Example local unit test, which will execute on the development machine (host).
*
* See [testing documentation](http://d.android.com/tools/testing).
*/
class ExampleUnitTest {
@Test
fun addition_isCorrect() {
assertEquals(4, 2 + 2)
}
}

1
examples/llama.android/settings.gradle.kts
View file

@ -15,3 +15,4 @@ dependencyResolutionManagement {
rootProject.name = "LlamaAndroid"
include(":app")
include(":llama")

2
examples/llava/clip.h
View file

@ -68,7 +68,7 @@ CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8
/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);

14
examples/main/main.cpp
View file

@ -740,18 +740,26 @@ int main(int argc, char ** argv) {
// display text
if (input_echo && display) {
for (auto id : embd) {
const std::string token_str = llama_token_to_piece(ctx, id, !params.conversation);
printf("%s", token_str.c_str());
const std::string token_str = llama_token_to_piece(ctx, id, params.special);
// Console/Stream Output
fprintf(stdout, "%s", token_str.c_str());
// Record Displayed Tokens To Log
// Note: Generated tokens are created one by one hence this check
if (embd.size() > 1) {
// Incoming Requested Tokens
input_tokens.push_back(id);
} else {
// Outgoing Generated Tokens
output_tokens.push_back(id);
output_ss << token_str;
}
}
fflush(stdout);
}
}
// reset color to default if there is no pending user input
if (input_echo && (int) embd_inp.size() == n_consumed) {
console::set_display(console::reset);

19
examples/server/public/index.html
View file

@ -594,7 +594,7 @@
message = html`<${Probabilities} data=${data} />`
} else {
const text = isArrayMessage ?
data.map(msg => msg.content).join('').replace(/^\s+/, '') :
data.map(msg => msg.content).join('') :
data;
message = isCompletionMode ?
text :
@ -877,7 +877,11 @@
// poor mans markdown replacement
const Markdownish = (params) => {
const md = params.text
const chunks = params.text.split('```');
for (let i = 0; i < chunks.length; i++) {
if (i % 2 === 0) { // outside code block
chunks[i] = chunks[i]
.replace(/&/g, '&amp;')
.replace(/</g, '&lt;')
.replace(/>/g, '&gt;')
@ -889,7 +893,14 @@
.replace(/```.*?\n([\s\S]*?)```/g, '<pre><code>$1</code></pre>')
.replace(/`(.*?)`/g, '<code>$1</code>')
.replace(/\n/gim, '<br />');
return html`<span dangerouslySetInnerHTML=${{ __html: md }} />`;
} else { // inside code block
chunks[i] = `<pre><code>${chunks[i]}</code></pre>`;
}
}
const restoredText = chunks.join('');
return html`<span dangerouslySetInnerHTML=${{ __html: restoredText }} />`;
};
const ModelGenerationInfo = (params) => {
@ -903,6 +914,7 @@
`
}
// simple popover impl
const Popover = (props) => {
const isOpen = useSignal(false);
@ -1054,4 +1066,3 @@
</body>
</html>

11
examples/server/public_simplechat/index.html
View file

@ -1,7 +1,7 @@
<!DOCTYPE html>
<html lang="en">
<head>
<title>SimpleChat (LlamaCPP, ...) </title>
<title>SimpleChat LlamaCppEtal </title>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<meta name="message" content="Save Nature Save Earth" />
@ -30,20 +30,17 @@
<hr>
<div class="sameline">
<label for="system-in">System</label>
<input type="text" name="system" id="system-in" class="flex-grow"/>
<input type="text" name="system" id="system-in" placeholder="e.g. you are a helpful ai assistant, who provides concise answers" class="flex-grow"/>
</div>
<hr>
<div id="chat-div">
<p> Enter the system prompt above, before entering/submitting any user query.</p>
<p> Enter your text to the ai assistant below.</p>
<p> Use shift+enter for inserting enter.</p>
<p> Refresh the page to start over fresh.</p>
<p> You need to have javascript enabled.</p>
</div>
<hr>
<div class="sameline">
<textarea id="user-in" class="flex-grow" rows="3"></textarea>
<textarea id="user-in" class="flex-grow" rows="3" placeholder="enter your query to the ai model here" ></textarea>
<button id="user-btn">submit</button>
</div>

128
examples/server/public_simplechat/readme.md
View file

@ -14,11 +14,15 @@ own system prompts.
The UI follows a responsive web design so that the layout can adapt to available display space in a usable
enough manner, in general.
NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
culling of old messages from the chat.
Allows developer/end-user to control some of the behaviour by updating gMe members from browser's devel-tool
console.
NOTE: It doesnt set any parameters other than temperature for now. However if someone wants they can update
the js file as needed.
NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
culling of old messages from the chat by default. However by enabling the sliding window chat logic, a crude
form of old messages culling can be achieved.
NOTE: It doesnt set any parameters other than temperature and max_tokens for now. However if someone wants
they can update the js file or equivalent member in gMe as needed.
## usage
@ -43,11 +47,33 @@ next run this web front end in examples/server/public_simplechat
### using the front end
Open this simple web front end from your local browser
* http://127.0.0.1:PORT/index.html
Once inside
* Select between chat and completion mode. By default it is set to chat mode.
* In completion mode
* logic by default doesnt insert any role specific "ROLE: " prefix wrt each role's message.
If the model requires any prefix wrt user role messages, then the end user has to
explicitly add the needed prefix, when they enter their chat message.
Similarly if the model requires any prefix to trigger assistant/ai-model response,
then the end user needs to enter the same.
This keeps the logic simple, while still giving flexibility to the end user to
manage any templating/tagging requirement wrt their messages to the model.
* the logic doesnt insert newline at the begining and end wrt the prompt message generated.
However if the chat being sent to /completions end point has more than one role's message,
then insert newline when moving from one role's message to the next role's message, so
that it can be clearly identified/distinguished.
* given that /completions endpoint normally doesnt add additional chat-templating of its
own, the above ensures that end user can create a custom single/multi message combo with
any tags/special-tokens related chat templating to test out model handshake. Or enduser
can use it just for normal completion related/based query.
* If you want to provide a system prompt, then ideally enter it first, before entering any user query.
Normally Completion mode doesnt need system prompt, while Chat mode can generate better/interesting
responses with a suitable system prompt.
* if chat.add_system_begin is used
* you cant change the system prompt, after it is has been submitted once along with user query.
* you cant set a system prompt, after you have submitted any user query
@ -55,27 +81,121 @@ Once inside
* one can change the system prompt any time during chat, by changing the contents of system prompt.
* inturn the updated/changed system prompt will be inserted into the chat session.
* this allows for the subsequent user chatting to be driven by the new system prompt set above.
* Enter your query and either press enter or click on the submit button.
If you want to insert enter (\n) as part of your chat/query to ai model, use shift+enter.
* Wait for the logic to communicate with the server and get the response.
* the user is not allowed to enter any fresh query during this time.
* the user input box will be disabled and a working message will be shown in it.
* just refresh the page, to reset wrt the chat history and or system prompt and start afresh.
* Using NewChat one can start independent chat sessions.
* two independent chat sessions are setup by default.
## Devel note
### Reason behind this
The idea is to be easy enough to use for basic purposes, while also being simple and easily discernable
by developers who may not be from web frontend background (so inturn may not be familiar with template /
end-use-specific-language-extensions driven flows) so that they can use it to explore/experiment things.
And given that the idea is also to help explore/experiment for developers, some flexibility is provided
to change behaviour easily using the devel-tools/console, for now. And skeletal logic has been implemented
to explore some of the end points and ideas/implications around them.
### General
Me/gMe consolidates the settings which control the behaviour into one object.
One can see the current settings, as well as change/update them using browsers devel-tool/console.
bCompletionFreshChatAlways - whether Completion mode collates complete/sliding-window history when
communicating with the server or only sends the latest user query/message.
bCompletionInsertStandardRolePrefix - whether Completion mode inserts role related prefix wrt the
messages that get inserted into prompt field wrt /Completion endpoint.
chatRequestOptions - maintains the list of options/fields to send along with chat request,
irrespective of whether /chat/completions or /completions endpoint.
If you want to add additional options/fields to send to the server/ai-model, and or
modify the existing options value or remove them, for now you can update this global var
using browser's development-tools/console.
iRecentUserMsgCnt - a simple minded SlidingWindow to limit context window load at Ai Model end.
This is disabled by default. However if enabled, then in addition to latest system message, only
the last/latest iRecentUserMsgCnt user messages after the latest system prompt and its responses
from the ai model will be sent to the ai-model, when querying for a new response. IE if enabled,
only user messages after the latest system message/prompt will be considered.
This specified sliding window user message count also includes the latest user query.
<0 : Send entire chat history to server
0 : Send only the system message if any to the server
>0 : Send the latest chat history from the latest system prompt, limited to specified cnt.
By using gMe's iRecentUserMsgCnt and chatRequestOptions.max_tokens one can try to control the
implications of loading of the ai-model's context window by chat history, wrt chat response to
some extent in a simple crude way.
Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
may not be visible. Also remember that just refreshing/reloading page in browser or for that
matter clearing site data, dont directly override site caching in all cases. Worst case you may
have to change port. Or in dev tools of browser, you may be able to disable caching fully.
Concept of multiple chat sessions with different servers, as well as saving and restoring of
those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and
its instances relatively easily, however given the current goal of keeping this simple, it has
not been added, for now.
By switching between chat.add_system_begin/anytime, one can control whether one can change
the system prompt, anytime during the conversation or only at the beginning.
read_json_early, is to experiment with reading json response data early on, if available,
so that user can be shown generated data, as and when it is being generated, rather than
at the end when full data is available.
the server flow doesnt seem to be sending back data early, atleast for request (inc options)
that is currently sent.
if able to read json data early on in future, as and when ai model is generating data, then
this helper needs to indirectly update the chat div with the recieved data, without waiting
for the overall data to be available.
### Default setup
By default things are setup to try and make the user experience a bit better, if possible.
However a developer when testing the server of ai-model may want to change these value.
Using iRecentUserMsgCnt reduce chat history context sent to the server/ai-model to be
just the system-prompt, prev-user-request-and-ai-response and cur-user-request, instead of
full chat history. This way if there is any response with garbage/repeatation, it doesnt
mess with things beyond the next question/request/query, in some ways.
Set max_tokens to 1024, so that a relatively large previous reponse doesnt eat up the space
available wrt next query-response. However dont forget that the server when started should
also be started with a model context size of 1k or more, to be on safe side.
The /completions endpoint of examples/server doesnt take max_tokens, instead it takes the
internal n_predict, for now add the same here on the client side, maybe later add max_tokens
to /completions endpoint handling code on server side.
Frequency and presence penalty fields are set to 1.2 in the set of fields sent to server
along with the user query. So that the model is partly set to try avoid repeating text in
its response.
A end-user can change these behaviour by editing gMe from browser's devel-tool/console.
## At the end
Also a thank you to all open source and open model developers, who strive for the common good.

7
examples/server/public_simplechat/simplechat.css
View file

@ -48,6 +48,13 @@ button {
flex-direction: column;
}
.ul1 {
padding-inline-start: 2vw;
}
.ul2 {
padding-inline-start: 2vw;
}
* {
margin: 0.6vmin;
}

207
examples/server/public_simplechat/simplechat.js
View file

@ -14,23 +14,86 @@ class ApiEP {
}
let gUsageMsg = `
<p> Enter the system prompt above, before entering/submitting any user query.</p>
<p> Enter your text to the ai assistant below.</p>
<p> Use shift+enter for inserting enter.</p>
<p> Refresh the page to start over fresh.</p>
<p class="role-system">Usage</p>
<ul class="ul1">
<li> Set system prompt above, to try control ai response charactersitic, if model supports same.</li>
<ul class="ul2">
<li> Completion mode normally wont have a system prompt.</li>
</ul>
<li> Enter your query to ai assistant below.</li>
<ul class="ul2">
<li> Completion mode doesnt insert user/role: prefix implicitly.</li>
<li> Use shift+enter for inserting enter/newline.</li>
</ul>
<li> Default ContextWindow = [System, Last Query+Resp, Cur Query].</li>
<ul class="ul2">
<li> experiment iRecentUserMsgCnt, max_tokens, model ctxt window to expand</li>
</ul>
</ul>
`;
/** @typedef {{role: string, content: string}[]} ChatMessages */
class SimpleChat {
constructor() {
/**
* Maintain in a form suitable for common LLM web service chat/completions' messages entry
* @type {{role: string, content: string}[]}
* @type {ChatMessages}
*/
this.xchat = [];
this.iLastSys = -1;
}
clear() {
this.xchat = [];
this.iLastSys = -1;
}
/**
* Recent chat messages.
* If iRecentUserMsgCnt < 0
* Then return the full chat history
* Else
* Return chat messages from latest going back till the last/latest system prompt.
* While keeping track that the number of user queries/messages doesnt exceed iRecentUserMsgCnt.
* @param {number} iRecentUserMsgCnt
*/
recent_chat(iRecentUserMsgCnt) {
if (iRecentUserMsgCnt < 0) {
return this.xchat;
}
if (iRecentUserMsgCnt == 0) {
console.warn("WARN:SimpleChat:SC:RecentChat:iRecentUsermsgCnt of 0 means no user message/query sent");
}
/** @type{ChatMessages} */
let rchat = [];
let sysMsg = this.get_system_latest();
if (sysMsg.length != 0) {
rchat.push({role: Roles.System, content: sysMsg});
}
let iUserCnt = 0;
let iStart = this.xchat.length;
for(let i=this.xchat.length-1; i > this.iLastSys; i--) {
if (iUserCnt >= iRecentUserMsgCnt) {
break;
}
let msg = this.xchat[i];
if (msg.role == Roles.User) {
iStart = i;
iUserCnt += 1;
}
}
for(let i = iStart; i < this.xchat.length; i++) {
let msg = this.xchat[i];
if (msg.role == Roles.System) {
continue;
}
rchat.push({role: msg.role, content: msg.content});
}
return rchat;
}
/**
* Add an entry into xchat
* @param {string} role
@ -57,7 +120,7 @@ class SimpleChat {
div.replaceChildren();
}
let last = undefined;
for(const x of this.xchat) {
for(const x of this.recent_chat(gMe.iRecentUserMsgCnt)) {
let entry = document.createElement("p");
entry.className = `role-${x.role}`;
entry.innerText = `${x.role}: ${x.content}`;
@ -69,17 +132,21 @@ class SimpleChat {
} else {
if (bClear) {
div.innerHTML = gUsageMsg;
gMe.show_info(div);
}
}
}
/**
* Add needed fields wrt json object to be sent wrt LLM web services completions endpoint
* Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
* The needed fields/options are picked from a global object.
* Convert the json into string.
* @param {Object} obj
*/
request_jsonstr(obj) {
obj["temperature"] = 0.7;
for(let k in gMe.chatRequestOptions) {
obj[k] = gMe.chatRequestOptions[k];
}
return JSON.stringify(obj);
}
@ -88,18 +155,27 @@ class SimpleChat {
*/
request_messages_jsonstr() {
let req = {
messages: this.xchat,
messages: this.recent_chat(gMe.iRecentUserMsgCnt),
}
return this.request_jsonstr(req);
}
/**
* Return a string form of json object suitable for /completions
* @param {boolean} bInsertStandardRolePrefix Insert "<THE_ROLE>: " as prefix wrt each role's message
*/
request_prompt_jsonstr() {
request_prompt_jsonstr(bInsertStandardRolePrefix) {
let prompt = "";
for(const chat of this.xchat) {
prompt += `${chat.role}: ${chat.content}\n`;
let iCnt = 0;
for(const chat of this.recent_chat(gMe.iRecentUserMsgCnt)) {
iCnt += 1;
if (iCnt > 1) {
prompt += "\n";
}
if (bInsertStandardRolePrefix) {
prompt += `${chat.role}: `;
}
prompt += `${chat.content}`;
}
let req = {
prompt: prompt,
@ -171,7 +247,6 @@ let gChatURL = {
'chat': `${gBaseURL}/chat/completions`,
'completion': `${gBaseURL}/completions`,
}
const gbCompletionFreshChatAlways = true;
/**
@ -291,6 +366,8 @@ class MultiChatUI {
// allow user to insert enter into their message using shift+enter.
// while just pressing enter key will lead to submitting.
if ((ev.key === "Enter") && (!ev.shiftKey)) {
let value = this.elInUser.value;
this.elInUser.value = value.substring(0,value.length-1);
this.elBtnUser.click();
ev.preventDefault();
}
@ -321,6 +398,29 @@ class MultiChatUI {
}
}
/**
* Try read json response early, if available.
* @param {Response} resp
*/
async read_json_early(resp) {
if (!resp.body) {
throw Error("ERRR:SimpleChat:MCUI:ReadJsonEarly:No body...");
}
let tdUtf8 = new TextDecoder("utf-8");
let rr = resp.body.getReader();
let gotBody = "";
while(true) {
let { value: cur, done: done} = await rr.read();
let curBody = tdUtf8.decode(cur);
console.debug("DBUG:SC:PART:", curBody);
gotBody += curBody;
if (done) {
break;
}
}
return JSON.parse(gotBody);
}
/**
* Handle user query submit request, wrt specified chat session.
* @param {string} chatId
@ -330,6 +430,14 @@ class MultiChatUI {
let chat = this.simpleChats[chatId];
// In completion mode, if configured, clear any previous chat history.
// So if user wants to simulate a multi-chat based completion query,
// they will have to enter the full thing, as a suitable multiline
// user input/query.
if ((apiEP == ApiEP.Completion) && (gMe.bCompletionFreshChatAlways)) {
chat.clear();
}
chat.add_system_anytime(this.elInSystem.value, chatId);
let content = this.elInUser.value;
@ -344,7 +452,7 @@ class MultiChatUI {
if (apiEP == ApiEP.Chat) {
theBody = chat.request_messages_jsonstr();
} else {
theBody = chat.request_prompt_jsonstr();
theBody = chat.request_prompt_jsonstr(gMe.bCompletionInsertStandardRolePrefix);
}
this.elInUser.value = "working...";
@ -359,6 +467,7 @@ class MultiChatUI {
});
let respBody = await resp.json();
//let respBody = await this.read_json_early(resp);
console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
let assistantMsg;
if (apiEP == ApiEP.Chat) {
@ -376,13 +485,6 @@ class MultiChatUI {
} else {
console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`);
}
// Purposefully clear at end rather than begin of this function
// so that one can switch from chat to completion mode and sequece
// in a completion mode with multiple user-assistant chat data
// from before to be sent/occur once.
if ((apiEP == ApiEP.Completion) && (gbCompletionFreshChatAlways)) {
chat.xchat.length = 0;
}
this.ui_reset_userinput();
}
@ -462,17 +564,66 @@ class MultiChatUI {
}
let gMuitChat;
const gChatIds = [ "Default", "Other" ];
class Me {
constructor() {
this.defaultChatIds = [ "Default", "Other" ];
this.multiChat = new MultiChatUI();
this.bCompletionFreshChatAlways = true;
this.bCompletionInsertStandardRolePrefix = false;
this.iRecentUserMsgCnt = 2;
// Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
this.chatRequestOptions = {
"temperature": 0.7,
"max_tokens": 1024,
"frequency_penalty": 1.2,
"presence_penalty": 1.2,
"n_predict": 1024
};
}
/**
* @param {HTMLDivElement} elDiv
*/
show_info(elDiv) {
var p = document.createElement("p");
p.innerText = "Settings (devel-tools-console gMe)";
p.className = "role-system";
elDiv.appendChild(p);
var p = document.createElement("p");
p.innerText = `bCompletionFreshChatAlways:${this.bCompletionFreshChatAlways}`;
elDiv.appendChild(p);
p = document.createElement("p");
p.innerText = `bCompletionInsertStandardRolePrefix:${this.bCompletionInsertStandardRolePrefix}`;
elDiv.appendChild(p);
p = document.createElement("p");
p.innerText = `iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`;
elDiv.appendChild(p);
p = document.createElement("p");
p.innerText = `chatRequestOptions:${JSON.stringify(this.chatRequestOptions)}`;
elDiv.appendChild(p);
}
}
/** @type {Me} */
let gMe;
function startme() {
console.log("INFO:SimpleChat:StartMe:Starting...");
gMuitChat = new MultiChatUI();
for (let cid of gChatIds) {
gMuitChat.new_chat_session(cid);
gMe = new Me();
for (let cid of gMe.defaultChatIds) {
gMe.multiChat.new_chat_session(cid);
}
gMuitChat.setup_ui(gChatIds[0]);
gMuitChat.show_sessions();
gMe.multiChat.setup_ui(gMe.defaultChatIds[0], true);
gMe.multiChat.show_sessions();
}
document.addEventListener("DOMContentLoaded", startme);

6
flake.lock generated
View file

@ -20,11 +20,11 @@
},
"nixpkgs": {
"locked": {
"lastModified": 1715961556,
"narHash": "sha256-+NpbZRCRisUHKQJZF3CT+xn14ZZQO+KjxIIanH3Pvn4=",
"lastModified": 1716509168,
"narHash": "sha256-4zSIhSRRIoEBwjbPm3YiGtbd8HDWzFxJjw5DYSDy1n8=",
"owner": "NixOS",
"repo": "nixpkgs",
"rev": "4a6b83b05df1a8bd7d99095ec4b4d271f2956b64",
"rev": "bfb7a882678e518398ce9a31a881538679f6f092",
"type": "github"
},
"original": {

37
ggml-cuda.cu
View file

@ -119,6 +119,20 @@ int ggml_cuda_get_device() {
return id;
}
static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
ggml_cuda_set_device(device);
#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
auto res = hipMallocManaged(ptr, size);
if (res == hipSuccess) {
// if error we "need" to know why...
CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
}
return res;
#else
return cudaMalloc(ptr, size);
#endif
}
static ggml_cuda_device_info ggml_cuda_init() {
#ifdef __HIP_PLATFORM_AMD__
// Workaround for a rocBLAS bug when using multiple graphics cards:
@ -271,7 +285,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
size_t look_ahead_size = (size_t) (1.05 * size);
look_ahead_size = 256 * ((look_ahead_size + 255)/256);
ggml_cuda_set_device(device);
CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
*actual_size = look_ahead_size;
pool_size += look_ahead_size;
#ifdef DEBUG_CUDA_MALLOC
@ -537,7 +551,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe
size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
void * dev_ptr;
cudaError_t err = cudaMalloc(&dev_ptr, size);
cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device);
if (err != cudaSuccess) {
// clear the error
cudaGetLastError();
@ -798,7 +812,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_bu
// currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
ggml_cuda_set_device(id);
char * buf;
CUDA_CHECK(cudaMalloc(&buf, size));
CUDA_CHECK(ggml_cuda_device_malloc((void**)&buf, size, id));
// set padding to 0 to avoid possible NaN values
if (size > original_size) {
@ -1856,7 +1870,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
}
}
#else
if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
// use cublasGemmStridedBatchedEx
CUBLAS_CHECK(
@ -2510,9 +2524,9 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
bool use_cuda_graph = true;
bool cuda_graph_update_required = false;
// pointer to CUDA cpy kernel, which is required to identify
// vector of pointers to CUDA cpy kernels, which are required to identify
// kernel parameters which need updated in the graph for each token
void * ggml_cuda_cpy_fn_ptr = nullptr;
std::vector<void *> ggml_cuda_cpy_fn_ptrs;
if (cuda_ctx->cuda_graph->graph == nullptr) {
if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
@ -2588,9 +2602,10 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
if (node->op == GGML_OP_CPY) {
// store the copy op parameter which changes with each token.
cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
if (ggml_cuda_cpy_fn_ptr == nullptr) {
// store a pointer to the copy op CUDA kernel to identify it later
ggml_cuda_cpy_fn_ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
// store a pointer to each copy op CUDA kernel to identify it later
void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
ggml_cuda_cpy_fn_ptrs.push_back(ptr);
}
}
@ -2720,7 +2735,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
if (!cuda_graph_update_required) { // on update steps, the live parameters will already be captured
int k = 0;
for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
if (cuda_ctx->cuda_graph->params[i].func == ggml_cuda_cpy_fn_ptr) {
if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
@ -2871,7 +2886,9 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
case GGML_OP_CONT:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
return true;
case GGML_OP_ROPE:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_IM2COL:
case GGML_OP_POOL_2D:
case GGML_OP_SUM_ROWS:

5
ggml-cuda/common.cuh
View file

@ -79,13 +79,8 @@
#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
#define cudaHostUnregister hipHostUnregister
#define cudaLaunchHostFunc hipLaunchHostFunc
#ifdef GGML_HIP_UMA
#define cudaMalloc hipMallocManaged
#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
#else
#define cudaMalloc hipMalloc
#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
#endif
#define cudaMemcpy hipMemcpy
#define cudaMemcpyAsync hipMemcpyAsync
#define cudaMemcpyPeerAsync hipMemcpyPeerAsync

163
ggml-cuda/concat.cu
View file

@ -1,15 +1,69 @@
#include "concat.cuh"
static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) {
// contiguous kernels
static __global__ void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) {
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
if (nidx >= ne0) {
return;
}
// operation
int offset_dst =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * gridDim.y;
if (nidx < ne00) { // src0
int offset_src =
nidx +
blockIdx.y * ne00 +
blockIdx.z * ne00 * gridDim.y;
dst[offset_dst] = x[offset_src];
} else {
int offset_src =
(nidx - ne00) +
blockIdx.y * (ne0 - ne00) +
blockIdx.z * (ne0 - ne00) * gridDim.y;
dst[offset_dst] = y[offset_src];
}
}
static __global__ void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) {
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
if (nidx >= ne0) {
return;
}
int offset_dst =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * gridDim.y;
if (blockIdx.y < ne01) { // src0
int offset_src =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * ne01;
dst[offset_dst] = x[offset_src];
} else {
int offset_src =
nidx +
(blockIdx.y - ne01) * ne0 +
blockIdx.z * ne0 * (gridDim.y - ne01);
dst[offset_dst] = y[offset_src];
}
}
static __global__ void concat_f32_dim2(const float * x, const float * y, float * dst, const int ne0, const int ne02) {
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
if (nidx >= ne0) {
return;
}
int offset_dst =
nidx +
blockIdx.y * ne0 +
blockIdx.z * ne0 * gridDim.y;
if (blockIdx.z < ne02) { // src0
int offset_src =
nidx +
@ -25,25 +79,118 @@ static __global__ void concat_f32(const float * x,const float * y, float * dst,
}
}
static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) {
static void concat_f32_cuda(const float * x, const float * y, float * dst, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) {
int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
dim3 gridDim(num_blocks, ne1, ne2);
concat_f32<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
if (dim == 0) {
concat_f32_dim0<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne00);
return;
}
if (dim == 1) {
concat_f32_dim1<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne01);
return;
}
concat_f32_dim2<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
}
// non-contiguous kernel (slow)
static __global__ void concat_f32_non_cont(
const char * src0,
const char * src1,
char * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne03,
uint64_t nb00,
uint64_t nb01,
uint64_t nb02,
uint64_t nb03,
int64_t /*ne10*/,
int64_t /*ne11*/,
int64_t /*ne12*/,
int64_t /*ne13*/,
uint64_t nb10,
uint64_t nb11,
uint64_t nb12,
uint64_t nb13,
int64_t ne0,
int64_t /*ne1*/,
int64_t /*ne2*/,
int64_t /*ne3*/,
uint64_t nb0,
uint64_t nb1,
uint64_t nb2,
uint64_t nb3,
int32_t dim) {
const int64_t i3 = blockIdx.z;
const int64_t i2 = blockIdx.y;
const int64_t i1 = blockIdx.x;
int64_t o[4] = {0, 0, 0, 0};
o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
const float * x;
for (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
x = (const float *)(src0 + (i3 )*nb03 + (i2 )*nb02 + (i1 )*nb01 + (i0 )*nb00);
} else {
x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
}
float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
*y = *x;
}
}
void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const float * src0_d = (const float *)src0->data;
const float * src1_d = (const float *)src1->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
const int32_t dim = ((int32_t *) dst->op_params)[0];
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
const float * src0_d = (const float *)src0->data;
const float * src1_d = (const float *)src1->data;
float * dst_d = (float *)dst->data;
if (dim != 3) {
for (int i3 = 0; i3 < dst->ne[3]; i3++) {
concat_f32_cuda(src0_d + i3 * (src0->nb[3] / 4), src1_d + i3 * (src1->nb[3] / 4), dst_d + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], stream);
concat_f32_cuda(
src0_d + i3 * (src0->nb[3] / 4),
src1_d + i3 * (src1->nb[3] / 4),
dst_d + i3 * ( dst->nb[3] / 4),
src0->ne[0], src0->ne[1], src0->ne[2],
dst->ne[0], dst->ne[1], dst->ne[2], dim, stream);
}
} else {
const size_t size0 = ggml_nbytes(src0);
const size_t size1 = ggml_nbytes(src1);
CUDA_CHECK(cudaMemcpyAsync(dst_d, src0_d, size0, cudaMemcpyDeviceToDevice, stream));
CUDA_CHECK(cudaMemcpyAsync(dst_d + size0/4, src1_d, size1, cudaMemcpyDeviceToDevice, stream));
}
} else {
dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
concat_f32_non_cont<<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
(const char *)src0->data,
(const char *)src1->data,
( char *)dst->data,
src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
}
}

6
ggml-cuda/norm.cu
View file

@ -170,6 +170,8 @@ void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
@ -188,6 +190,8 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
@ -202,6 +206,8 @@ void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);

18
ggml-cuda/rope.cu
View file

@ -61,7 +61,7 @@ static __global__ void rope(
template<typename T, bool has_pos, bool has_freq_facs>
static __global__ void rope_neox(
const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, const float * freq_factors
float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors
) {
const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
@ -85,15 +85,13 @@ static __global__ void rope_neox(
const int i = row*ncols + ib*n_dims + ic/2;
const int i2 = row/p_delta_rows;
float cur_rot = inv_ndims * ic - ib;
const int p = has_pos ? pos[i2] : 0;
const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f;
const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f)/freq_factor;
const float theta_base = p*powf(theta_scale, col/2.0f)/freq_factor;
float cos_theta, sin_theta;
rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
rope_yarn(theta_base, freq_scale, corr_dims, ic, ext_factor, attn_factor, &cos_theta, &sin_theta);
const float x0 = x[i + 0];
const float x1 = x[i + n_dims/2];
@ -174,30 +172,29 @@ static void rope_neox_cuda(
const dim3 block_nums(nrows, num_blocks_x, 1);
const float theta_scale = powf(freq_base, -2.0f/n_dims);
const float inv_ndims = -1.0f / n_dims;
if (pos == nullptr) {
if (freq_factors == nullptr) {
rope_neox<T, false, false><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
theta_scale, freq_factors
);
} else {
rope_neox<T, false, true><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
theta_scale, freq_factors
);
}
} else {
if (freq_factors == nullptr) {
rope_neox<T, true, false><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
theta_scale, freq_factors
);
} else {
rope_neox<T, true, true><<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
theta_scale, inv_ndims, freq_factors
theta_scale, freq_factors
);
}
}
@ -254,6 +251,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
GGML_ASSERT(src0->type == dst->type);

4
ggml-impl.h
View file

@ -144,6 +144,10 @@ extern "C" {
#endif
#endif
#if defined(__ARM_FEATURE_SVE)
#include <arm_sve.h>
#endif
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t

1
ggml-kompute.cpp
View file

@ -1597,7 +1597,6 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
{
GGML_ASSERT(ne00 == ne10);
// TODO: assert that dim2 and dim3 are contiguous
GGML_ASSERT(ne12 % ne02 == 0);
GGML_ASSERT(ne13 % ne03 == 0);

76
ggml-metal.m
View file

@ -35,6 +35,10 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_ROW,
GGML_METAL_KERNEL_TYPE_DIV,
GGML_METAL_KERNEL_TYPE_DIV_ROW,
GGML_METAL_KERNEL_TYPE_REPEAT_F32,
GGML_METAL_KERNEL_TYPE_REPEAT_F16,
GGML_METAL_KERNEL_TYPE_REPEAT_I32,
GGML_METAL_KERNEL_TYPE_REPEAT_I16,
GGML_METAL_KERNEL_TYPE_SCALE,
GGML_METAL_KERNEL_TYPE_SCALE_4,
GGML_METAL_KERNEL_TYPE_CLAMP,
@ -184,9 +188,9 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
//GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,
//GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
@ -485,6 +489,10 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32, repeat_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16, repeat_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32, repeat_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16, repeat_i16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true);
@ -634,9 +642,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true);
@ -746,6 +754,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
case GGML_OP_ACC:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_REPEAT:
case GGML_OP_SCALE:
case GGML_OP_CLAMP:
case GGML_OP_SQR:
@ -770,6 +779,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
case GGML_OP_LEAKY_RELU:
return true;
case GGML_OP_FLASH_ATTN_EXT:
if (op->src[0]->ne[0] == 256) {
return false;
}
return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
@ -976,10 +988,10 @@ static enum ggml_status ggml_metal_graph_compute(
switch (dst->op) {
case GGML_OP_CONCAT:
{
const int64_t nb = ne00;
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline;
const int32_t dim = ((int32_t *) dst->op_params)[0];
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
@ -1008,7 +1020,7 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26];
[encoder setBytes:&nb length:sizeof(nb) atIndex:27];
[encoder setBytes:&dim length:sizeof(dim) atIndex:27];
const int nth = MIN(1024, ne0);
@ -1018,11 +1030,14 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_OP_MUL:
case GGML_OP_DIV:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
GGML_ASSERT(src1t == GGML_TYPE_F32);
const size_t offs = 0;
bool bcast_row = false;
int64_t nb = ne00;
int64_t nb = ne00; // used by the "row" kernels
id<MTLComputePipelineState> pipeline = nil;
@ -1091,6 +1106,42 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
}
} break;
case GGML_OP_REPEAT:
{
id<MTLComputePipelineState> pipeline;
switch (src0t) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break;
case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break;
default: GGML_ASSERT(false);
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ACC:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
@ -1468,7 +1519,6 @@ static enum ggml_status ggml_metal_graph_compute(
{
GGML_ASSERT(ne00 == ne10);
// TODO: assert that dim2 and dim3 are contiguous
GGML_ASSERT(ne12 % ne02 == 0);
GGML_ASSERT(ne13 % ne03 == 0);
@ -2136,6 +2186,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_OP_RMS_NORM:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(ggml_is_contiguous_1(src0));
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
@ -2163,6 +2214,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_OP_GROUP_NORM:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(ggml_is_contiguous(src0));
//float eps;
//memcpy(&eps, dst->op_params, sizeof(float));
@ -2196,6 +2248,8 @@ static enum ggml_status ggml_metal_graph_compute(
} break;
case GGML_OP_NORM:
{
GGML_ASSERT(ggml_is_contiguous_1(src0));
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
@ -2573,7 +2627,7 @@ static enum ggml_status ggml_metal_graph_compute(
case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
//case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
@ -2586,7 +2640,7 @@ static enum ggml_status ggml_metal_graph_compute(
switch (ne00) {
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break;
case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
//case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);

96
ggml-metal.metal
View file

@ -168,6 +168,53 @@ kernel void kernel_div(
}
}
template<typename T>
kernel void kernel_repeat(
device const char * src0,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i3 = tgpig.z;
const int64_t i2 = tgpig.y;
const int64_t i1 = tgpig.x;
const int64_t i03 = i3 % ne03;
const int64_t i02 = i2 % ne02;
const int64_t i01 = i1 % ne01;
device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
device char * dst_ptr = dst + i3*nb3 + i2*nb2 + i1*nb1 ;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
const int i00 = i0 % ne00;
*((device T *)(dst_ptr + i0*nb0)) = *((device T *)(src0_ptr + i00*nb00));
}
}
typedef decltype(kernel_repeat<float>) kernel_repeat_t;
template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;
// assumption: src1 is a row
// broadcast src1 into src0
kernel void kernel_add_row(
@ -1720,13 +1767,13 @@ kernel void kernel_rope(
const int64_t p = pos[i2];
const float theta_0 = (float)p;
const float theta_base = (float)p;
const float inv_ndims = -1.f/n_dims;
if (!is_neox) {
for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
const float theta = theta_base * pow(freq_base, inv_ndims*i0);
const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
float cos_theta, sin_theta;
rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
@ -1742,18 +1789,14 @@ kernel void kernel_rope(
} else {
for (int64_t ic = 2*tiitg; ic < ne0; ic += 2*tptg.x) {
if (ic < n_dims) {
const int64_t ib = 0;
const int64_t i0 = ic/2;
// simplified from `(ib * n_dims + ic) * inv_ndims`
const float cur_rot = inv_ndims*ic - ib;
const float freq_factor = src2 != src0 ? src2[ic/2] : 1.0f;
const float freq_factor = src2 != src0 ? src2[i0] : 1.0f;
const float theta = theta_0 * pow(freq_base, cur_rot) / freq_factor;
const float theta = theta_base * pow(freq_base, inv_ndims*ic);
float cos_theta, sin_theta;
rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
const int64_t i0 = ib*n_dims + ic/2;
rope_yarn(theta/freq_factor, freq_scale, corr_dims, ic, ext_factor, attn_factor, &cos_theta, &sin_theta);
device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
device T * dst_data = (device T *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
@ -2418,7 +2461,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f
template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<96>;
template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<112>;
template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128>;
template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>;
//template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>;
template<int64_t D, int64_t Q = 1, int64_t C = 32> // head size, queries per threadgroup, cache items per threadgroup
kernel void kernel_flash_attn_ext_vec_f16(
@ -2696,7 +2739,7 @@ kernel void kernel_flash_attn_ext_vec_f16(
}
template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<128>;
template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>;
//template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>;
kernel void kernel_cpy_f16_f16(
device const half * src0,
@ -3319,31 +3362,30 @@ kernel void kernel_concat(
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int32_t & dim,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x;
const int64_t i3 = tgpig.z;
const int64_t i2 = tgpig.y;
const int64_t i1 = tgpig.x;
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
int64_t o[4] = {0, 0, 0, 0};
o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1 + tpitg.x*nb0;
device const float * x;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
if (i02 < ne02) {
((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0];
src0_ptr += ntg.x*nb00;
if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
x = (device const float *)(src0 + (i3 )*nb03 + (i2 )*nb02 + (i1 )*nb01 + (i0 )*nb00);
} else {
((device float *)dst_ptr)[0] = ((device float *)src1_ptr)[0];
src1_ptr += ntg.x*nb10;
x = (device const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
}
dst_ptr += ntg.x*nb0;
device float * y = (device float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
*y = *x;
}
}

66
ggml-quants.c
View file

@ -3813,7 +3813,44 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
return;
}
#endif
#if defined(__ARM_NEON)
#if defined(__ARM_FEATURE_SVE)
const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
svfloat32_t sumv0 = svdup_n_f32(0.0f);
svfloat32_t sumv1 = svdup_n_f32(0.0f);
assert(nb % 2 == 0); // TODO: handle odd nb
for (int i = 0; i < nb; i += 2) {
const block_q4_0 * restrict x0 = &x[i + 0];
const block_q4_0 * restrict x1 = &x[i + 1];
const block_q8_0 * restrict y0 = &y[i + 0];
const block_q8_0 * restrict y1 = &y[i + 1];
// load x
const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
// 4-bit -> 8-bit
const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04));
const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04));
// sub 8
const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
// load y
const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
// dot product
sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
*s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
#elif defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
float32x4_t sumv1 = vdupq_n_f32(0.0f);
@ -5384,7 +5421,32 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
return;
}
#endif
#if defined(__ARM_NEON)
#if defined(__ARM_FEATURE_SVE)
svfloat32_t sumv0 = svdup_n_f32(0.0f);
svfloat32_t sumv1 = svdup_n_f32(0.0f);
assert(nb % 2 == 0); // TODO: handle odd nb
for (int i = 0; i < nb; i += 2) {
const block_q8_0 * restrict x0 = &x[i + 0];
const block_q8_0 * restrict x1 = &x[i + 1];
const block_q8_0 * restrict y0 = &y[i + 0];
const block_q8_0 * restrict y1 = &y[i + 1];
// load x
const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
// load y
const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
*s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
#elif defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
float32x4_t sumv1 = vdupq_n_f32(0.0f);

127
ggml-rpc.cpp
View file

@ -6,6 +6,7 @@
#include <string>
#include <vector>
#include <memory>
#include <mutex>
#include <unordered_map>
#include <unordered_set>
#ifdef _WIN32
@ -47,6 +48,7 @@ struct socket_t {
sockfd_t fd;
socket_t(sockfd_t fd) : fd(fd) {}
~socket_t() {
GGML_PRINT_DEBUG("[%s] closing socket %d\n", __func__, this->fd);
#ifdef _WIN32
closesocket(this->fd);
#else
@ -97,7 +99,7 @@ static ggml_guid_t ggml_backend_rpc_guid() {
}
struct ggml_backend_rpc_buffer_type_context {
std::shared_ptr<socket_t> sock;
std::string endpoint;
std::string name;
size_t alignment;
size_t max_size;
@ -106,8 +108,6 @@ struct ggml_backend_rpc_buffer_type_context {
struct ggml_backend_rpc_context {
std::string endpoint;
std::string name;
std::shared_ptr<socket_t> sock;
ggml_backend_buffer_type_t buft;
};
struct ggml_backend_rpc_buffer_context {
@ -231,14 +231,13 @@ static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
return true;
}
static bool parse_endpoint(const char * endpoint, std::string & host, int & port) {
std::string str(endpoint);
size_t pos = str.find(':');
static bool parse_endpoint(const std::string & endpoint, std::string & host, int & port) {
size_t pos = endpoint.find(':');
if (pos == std::string::npos) {
return false;
}
host = str.substr(0, pos);
port = std::stoi(str.substr(pos + 1));
host = endpoint.substr(0, pos);
port = std::stoi(endpoint.substr(pos + 1));
return true;
}
@ -273,6 +272,44 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
// RPC client-side implementation
static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
static std::unordered_map<std::string, std::weak_ptr<socket_t>> sockets;
static bool initialized = false;
auto it = sockets.find(endpoint);
if (it != sockets.end()) {
if (auto sock = it->second.lock()) {
return sock;
}
}
std::string host;
int port;
if (!parse_endpoint(endpoint, host, port)) {
return nullptr;
}
#ifdef _WIN32
if (!initialized) {
WSADATA wsaData;
int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
if (res != 0) {
return nullptr;
}
initialized = true;
}
#else
UNUSED(initialized);
#endif
auto sock = socket_connect(host.c_str(), port);
if (sock == nullptr) {
return nullptr;
}
GGML_PRINT_DEBUG("[%s] connected to %s, sockfd=%d\n", __func__, endpoint.c_str(), sock->fd);
sockets[endpoint] = sock;
return sock;
}
GGML_CALL static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
return ctx->name.c_str();
@ -442,7 +479,8 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer
std::vector<uint8_t> input(input_size, 0);
memcpy(input.data(), &size, sizeof(size));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(buft_ctx->sock, ALLOC_BUFFER, input, output);
auto sock = get_socket(buft_ctx->endpoint);
bool status = send_rpc_cmd(sock, ALLOC_BUFFER, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == 2*sizeof(uint64_t));
// output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) |
@ -453,7 +491,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer
if (remote_ptr != 0) {
ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
ggml_backend_rpc_buffer_interface,
new ggml_backend_rpc_buffer_context{buft_ctx->sock, {}, remote_ptr, "RPC"},
new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC"},
remote_size);
return buffer;
} else {
@ -508,7 +546,7 @@ GGML_CALL static bool ggml_backend_rpc_buffer_type_supports_backend(ggml_backend
}
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
return buft_ctx->sock == rpc_ctx->sock;
return buft_ctx->endpoint == rpc_ctx->endpoint;
}
static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = {
@ -521,7 +559,6 @@ static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = {
/* .is_host = */ NULL,
};
GGML_CALL static const char * ggml_backend_rpc_name(ggml_backend_t backend) {
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
@ -530,16 +567,13 @@ GGML_CALL static const char * ggml_backend_rpc_name(ggml_backend_t backend) {
GGML_CALL static void ggml_backend_rpc_free(ggml_backend_t backend) {
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)rpc_ctx->buft->context;
delete buft_ctx;
delete rpc_ctx->buft;
delete rpc_ctx;
delete backend;
}
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) {
ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;
return ctx->buft;
return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str());
}
GGML_CALL static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
@ -590,7 +624,8 @@ GGML_CALL static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t
std::vector<uint8_t> input;
serialize_graph(cgraph, input);
std::vector<uint8_t> output;
bool status = send_rpc_cmd(rpc_ctx->sock, GRAPH_COMPUTE, input, output);
auto sock = get_socket(rpc_ctx->endpoint);
bool status = send_rpc_cmd(sock, GRAPH_COMPUTE, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == 1);
return (enum ggml_status)output[0];
@ -624,42 +659,24 @@ static ggml_backend_i ggml_backend_rpc_interface = {
/* .event_synchronize = */ NULL,
};
static std::unordered_map<std::string, ggml_backend_t> instances;
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
ggml_backend_t backend = ggml_backend_rpc_init(endpoint);
return backend != nullptr ? ggml_backend_rpc_get_default_buffer_type(backend) : nullptr;
}
GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
std::string endpoint_str(endpoint);
if (instances.find(endpoint_str) != instances.end()) {
return instances[endpoint_str];
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
// NOTE: buffer types are allocated and never freed; this is by design
static std::unordered_map<std::string, ggml_backend_buffer_type_t> buft_map;
auto it = buft_map.find(endpoint);
if (it != buft_map.end()) {
return it->second;
}
#ifdef _WIN32
{
WSADATA wsaData;
int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
if (res != 0) {
return nullptr;
}
}
#endif
fprintf(stderr, "Connecting to %s\n", endpoint);
std::string host;
int port;
if (!parse_endpoint(endpoint, host, port)) {
return nullptr;
}
auto sock = socket_connect(host.c_str(), port);
auto sock = get_socket(endpoint);
if (sock == nullptr) {
return nullptr;
}
size_t alignment = get_alignment(sock);
size_t max_size = get_max_size(sock);
ggml_backend_rpc_buffer_type_context * buft_ctx = new ggml_backend_rpc_buffer_type_context {
/* .sock = */ sock,
/* .name = */ "RPC" + std::to_string(sock->fd),
/* .endpoint = */ endpoint,
/* .name = */ "RPC[" + std::string(endpoint) + "]",
/* .alignment = */ alignment,
/* .max_size = */ max_size
};
@ -668,21 +685,22 @@ GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
/* .iface = */ ggml_backend_rpc_buffer_type_interface,
/* .context = */ buft_ctx
};
buft_map[endpoint] = buft;
return buft;
}
GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
/* .endpoint = */ endpoint,
/* .name = */ "RPC" + std::to_string(sock->fd),
/* .sock = */ sock,
/* .buft = */ buft
/* .name = */ "RPC",
};
instances[endpoint] = new ggml_backend {
ggml_backend_t backend = new ggml_backend {
/* .guid = */ ggml_backend_rpc_guid(),
/* .interface = */ ggml_backend_rpc_interface,
/* .context = */ ctx
};
return instances[endpoint];
return backend;
}
GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend) {
@ -706,14 +724,13 @@ static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * f
}
GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
ggml_backend_t backend = ggml_backend_rpc_init(endpoint);
if (backend == nullptr) {
auto sock = get_socket(endpoint);
if (sock == nullptr) {
*free = 0;
*total = 0;
return;
}
ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;
get_device_memory(ctx->sock, free, total);
get_device_memory(sock, free, total);
}
// RPC server-side implementation

466
ggml-sycl.cpp
View file

@ -2944,6 +2944,57 @@ namespace dpct
using shared_memory = detail::device_memory<T, shared, Dimension>;
template <typename T,
sycl::access::address_space addressSpace =
sycl::access::address_space::global_space,
sycl::memory_order memoryOrder = sycl::memory_order::relaxed,
sycl::memory_scope memoryScope = sycl::memory_scope::device>
inline T atomic_fetch_add(T *addr, T operand) {
auto atm =
sycl::atomic_ref<T, memoryOrder, memoryScope, addressSpace>(addr[0]);
return atm.fetch_add(operand);
}
template <sycl::access::address_space addressSpace =
sycl::access::address_space::global_space,
sycl::memory_order memoryOrder = sycl::memory_order::relaxed,
sycl::memory_scope memoryScope = sycl::memory_scope::device,
typename T1, typename T2>
inline T1 atomic_fetch_add(T1 *addr, T2 operand) {
auto atm =
sycl::atomic_ref<T1, memoryOrder, memoryScope, addressSpace>(addr[0]);
return atm.fetch_add(operand);
}
template <typename T, sycl::access::address_space addressSpace =
sycl::access::address_space::global_space>
inline T atomic_fetch_add(T *addr, T operand,
sycl::memory_order memoryOrder) {
switch (memoryOrder) {
case sycl::memory_order::relaxed:
return atomic_fetch_add<T, addressSpace, sycl::memory_order::relaxed,
sycl::memory_scope::device>(addr, operand);
case sycl::memory_order::acq_rel:
return atomic_fetch_add<T, addressSpace, sycl::memory_order::acq_rel,
sycl::memory_scope::device>(addr, operand);
case sycl::memory_order::seq_cst:
return atomic_fetch_add<T, addressSpace, sycl::memory_order::seq_cst,
sycl::memory_scope::device>(addr, operand);
default:
assert(false && "Invalid memory_order for atomics. Valid memory_order for "
"atomics are: sycl::memory_order::relaxed, "
"sycl::memory_order::acq_rel, sycl::memory_order::seq_cst!");
}
}
template <sycl::access::address_space addressSpace =
sycl::access::address_space::global_space,
typename T1, typename T2>
inline T1 atomic_fetch_add(T1 *addr, T2 operand,
sycl::memory_order memoryOrder) {
atomic_fetch_add<T1, addressSpace>(addr, operand, memoryOrder);
}
} // COPY from DPCT head files
#define GGML_COMMON_DECL_SYCL
@ -2971,20 +3022,19 @@ static int g_work_group_size = 0;
// typedef sycl::half ggml_fp16_t;
#define __SYCL_ARCH__ DPCT_COMPATIBILITY_TEMP
#define VER_4VEC 610 //todo for hardward optimize.
#define VER_4VEC 130 //todo for hardward optimize.
#define VER_GEN9 700 //todo for hardward optimize.
#define VER_GEN12 1000000 //todo for hardward optimize.
#define VER_GEN13 (VER_GEN12 + 1030) //todo for hardward optimize.
#define GGML_SYCL_MAX_NODES 8192 //TODO: adapt to hardwares
//define for XMX in Intel GPU
//TODO: currently, it's not used for XMX really.
#define SYCL_USE_XMX
#if !defined(GGML_SYCL_FORCE_MMQ)
#define SYCL_USE_XMX
#endif
// max batch size to use MMQ kernels when tensor cores are available
#define XMX_MAX_BATCH_SIZE 32
#define MMQ_MAX_BATCH_SIZE 32
#if defined(_MSC_VER)
@ -3060,6 +3110,7 @@ void ggml_sycl_get_device_description(int device, char * description, size_t d
bool ggml_backend_is_sycl(ggml_backend_t backend);
int ggml_backend_sycl_get_device(ggml_backend_t backend);
int get_main_device();
static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer);
void print_ggml_tensor(const char*name, struct ggml_tensor *src);
void log_tensor_with_cnt(const char* name, struct ggml_tensor * src, int stop_cnt);
@ -8830,12 +8881,11 @@ static void rope(
dst[i + 1] = x0*sin_theta + x1*cos_theta;
}
template<typename T, bool has_pos>
template<typename T, bool has_pos, bool has_freq_facs>
static void rope_neox(
const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims
,
const sycl::nd_item<3> &item_ct1) {
float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims,
const float * freq_factors, const sycl::nd_item<3> &item_ct1) {
const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
item_ct1.get_local_id(1));
@ -8863,8 +8913,10 @@ static void rope_neox(
float cur_rot = inv_ndims * ic - ib;
const int p = has_pos ? pos[i2] : 0;
const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f;
const float theta_base =
p * freq_scale * dpct::pow(theta_scale, col / 2.0f);
p * freq_scale * dpct::pow(theta_scale, col / 2.0f)/freq_factor;
float cos_theta, sin_theta;
rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
@ -12413,7 +12465,7 @@ static void rope_neox_sycl(const T *x, T *dst, int ncols, int n_dims, int nrows,
const int32_t *pos, float freq_scale,
int p_delta_rows, float freq_base, float ext_factor,
float attn_factor, rope_corr_dims corr_dims,
dpct::queue_ptr stream) {
const float * freq_factors, dpct::queue_ptr stream) {
GGML_ASSERT(ncols % 2 == 0);
const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
const int num_blocks_x = (ncols + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE);
@ -12423,39 +12475,49 @@ static void rope_neox_sycl(const T *x, T *dst, int ncols, int n_dims, int nrows,
const float inv_ndims = -1.0f / n_dims;
if (pos == nullptr) {
/*
DPCT1049:42: The work-group size passed to the SYCL kernel may exceed
the limit. To get the device limit, query
info::device::max_work_group_size. Adjust the work-group size if needed.
*/
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
if (freq_factors == nullptr) {
stream->parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
rope_neox<T, false>(x, dst, ncols, n_dims, pos, freq_scale,
rope_neox<T, false, false>(x, dst, ncols, n_dims, pos, freq_scale,
p_delta_rows, ext_factor, attn_factor,
corr_dims, theta_scale, inv_ndims,
corr_dims, theta_scale, inv_ndims, freq_factors,
item_ct1);
});
} else {
/*
DPCT1049:43: The work-group size passed to the SYCL kernel may exceed
the limit. To get the device limit, query
info::device::max_work_group_size. Adjust the work-group size if needed.
*/
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
stream->parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
rope_neox<T, true>(x, dst, ncols, n_dims, pos, freq_scale,
rope_neox<T, false, true>(x, dst, ncols, n_dims, pos, freq_scale,
p_delta_rows, ext_factor, attn_factor,
corr_dims, theta_scale, inv_ndims, item_ct1);
corr_dims, theta_scale, inv_ndims, freq_factors,
item_ct1);
});
}
} else {
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
if (freq_factors == nullptr) {
stream->parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
rope_neox<T, true, false>(x, dst, ncols, n_dims, pos, freq_scale,
p_delta_rows, ext_factor, attn_factor,
corr_dims, theta_scale, inv_ndims, freq_factors, item_ct1);
});
} else {
stream->parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
rope_neox<T, true, true>(x, dst, ncols, n_dims, pos, freq_scale,
p_delta_rows, ext_factor, attn_factor,
corr_dims, theta_scale, inv_ndims, freq_factors, item_ct1);
});
}
}
}
static void rope_glm_f32_sycl(const float *x, float *dst, int ncols, int nrows,
@ -13501,6 +13563,10 @@ inline void ggml_sycl_op_concat(const ggml_tensor *src0,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const dpct::queue_ptr &main_stream) {
#pragma message("TODO: generalize concat kernel for dim != 2")
#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7563")
int dim = dst->op_params[0];
GGML_ASSERT(dim == 2);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
@ -13986,9 +14052,7 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const dpct::queue_ptr &main_stream) {
#pragma message("TODO: implement phi3 frequency factors support")
#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225")
GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
const ggml_tensor * src2 = dst->src[2];
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
@ -14014,6 +14078,7 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const float * freq_factors = nullptr;
const int32_t * pos = nullptr;
if ((mode & 1) == 0) {
GGML_ASSERT(src1->type == GGML_TYPE_I32);
@ -14024,6 +14089,16 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
const bool is_neox = mode & 2;
const bool is_glm = mode & 4;
if (is_neox) {
pos = (const int32_t *) src1_dd;
if (src2 != nullptr) {
freq_factors = (const float *) src2->data;
}
} else {
GGML_ASSERT(src2 == nullptr && "TODO: freq_factors not implemented for !is_neox");
}
rope_corr_dims corr_dims;
ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
@ -14035,13 +14110,13 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
if (src0->type == GGML_TYPE_F32) {
rope_neox_sycl(
(const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, main_stream
attn_factor, corr_dims, freq_factors, main_stream
);
} else if (src0->type == GGML_TYPE_F16) {
rope_neox_sycl((const sycl::half *)src0_dd, (sycl::half *)dst_dd,
ne00, n_dims, nrows, pos, freq_scale, ne01,
freq_base, ext_factor, attn_factor, corr_dims,
main_stream);
freq_factors, main_stream);
} else {
GGML_ASSERT(false);
}
@ -15108,7 +15183,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
const int64_t r2 = ne12/ne02;
const int64_t r3 = ne13/ne03;
if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
*g_sycl_handles[g_main_device], oneapi::mkl::transpose::trans,
@ -15173,6 +15248,29 @@ catch (sycl::exception const &exc) {
std::exit(1);
}
inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
// TODO: accuracy issues in MMQ
return false;
}
bool ggml_sycl_supports_dmmv(enum ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
case GGML_TYPE_F16:
return true;
default:
return false;
}
}
static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const bool all_on_device =
@ -15189,76 +15287,43 @@ static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
}
}
// check data types and tensor shapes for custom matrix multiplication kernels:
bool use_dequantize_mul_mat_vec = ggml_sycl_supports_dmmv(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src0->ne[0] % GGML_SYCL_DMMV_X == 0 && src1->ne[1] == 1;
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
bool use_mul_mat_q = ggml_sycl_supports_mmq(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
// mmvq and mmq need the __dp4a instruction which is available for gen12+
// Workaround in https://github.com/ggerganov/llama.cpp/commit/95f84d5ce8b449a9b16009434aca800df504a02e
use_mul_mat_q = use_mul_mat_q && (src0->type != GGML_TYPE_IQ2_XXS);
#ifdef SYCL_USE_XMX
const bool use_xmx = true;
#else
const bool use_xmx = false;
#endif
use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
#endif // SYCL_USE_XMX
// debug helpers
//printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
//printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
//printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
//printf(" %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
//printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
//printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
if (!split && all_on_device && !use_xmx && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
// KQ single-batch
// GGML_SYCL_DEBUG("ggml_sycl_mul_mat_vec_p021\n");
ggml_sycl_mul_mat_vec_p021(src0, src1, dst);
} else if (!split && all_on_device && !use_xmx && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
} else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
// KQV single-batch
// GGML_SYCL_DEBUG("ggml_sycl_mul_mat_vec_nc\n");
ggml_sycl_mul_mat_vec_nc(src0, src1, dst);
} else if (!split && all_on_device && use_xmx && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16) && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
// KQ + KQV multi-batch
// GGML_SYCL_DEBUG("ggml_sycl_mul_mat_batched_sycl\n");
ggml_sycl_mul_mat_batched_sycl(src0, src1, dst);
} else if (src0->type == GGML_TYPE_F32) {
// GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat\n");
ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false);
} else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
// GGML_SYCL_DEBUG("ggml_is_quantized or GGML_TYPE_F16\n");
if (src1->ne[1] == 1 && src0->ne[0] % GGML_SYCL_DMMV_X == 0) {
#ifdef GGML_SYCL_FORCE_DMMV
const bool use_mul_mat_vec_q = false;
#else
bool use_mul_mat_vec_q = min_compute_capability >= VER_4VEC && ggml_is_quantized(src0->type);
use_mul_mat_vec_q = use_mul_mat_vec_q ||
(src0->type == GGML_TYPE_IQ2_XXS) || (src0->type == GGML_TYPE_IQ2_XS) || (src0->type == GGML_TYPE_IQ2_S) ||
(src0->type == GGML_TYPE_IQ3_XXS) || (src0->type == GGML_TYPE_IQ3_S) ||
(src0->type == GGML_TYPE_IQ4_NL) || (src0->type == GGML_TYPE_IQ4_XS) ||
(src0->type == GGML_TYPE_IQ1_S) || (src0->type == GGML_TYPE_IQ1_M);
#endif // GGML_SYCL_FORCE_DMMV
if (use_mul_mat_vec_q) {
// GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_vec_q path\n");
ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true);
} else {
// GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_dequantize_mul_mat_vec path\n");
} else if (use_dequantize_mul_mat_vec) {
ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false);
}
} else {
bool use_mul_mat_q = min_compute_capability >= VER_4VEC && ggml_is_quantized(src0->type);
if (use_xmx && min_compute_capability >= VER_GEN9 && src1->ne[1] > XMX_MAX_BATCH_SIZE) {
use_mul_mat_q = false;
}
if (use_mul_mat_q) {
// GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_q path\n");
} else if (use_mul_mat_vec_q) {
ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true);
} else if (use_mul_mat_q) {
ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_q, true);
} else {
// GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_sycl path\n");
ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false);
}
}
} else {
GGML_ASSERT(false);
}
}
#if 0
@ -15434,22 +15499,86 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) {
}
#endif
struct mmid_row_mapping {
int32_t i1;
int32_t i2;
};
__dpct_inline__ static void k_copy_src1_to_contiguous(
const char *__restrict__ src1_original, char *__restrict__ src1_contiguous,
int *__restrict__ cur_src1_row, mmid_row_mapping *__restrict__ row_mapping,
const char *__restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
int64_t ne11, int64_t ne10, size_t nb11, size_t nb12,
const sycl::nd_item<3> &item_ct1, int &src1_row) {
int32_t iid1 = item_ct1.get_group(2);
int32_t id = item_ct1.get_group(1);
const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
if (row_id_i != i02) {
return;
}
const int64_t i11 = id % ne11;
const int64_t i12 = iid1;
if (item_ct1.get_local_id(2) == 0) {
src1_row =
dpct::atomic_fetch_add<sycl::access::address_space::generic_space>(
cur_src1_row, 1);
row_mapping[src1_row] = {id, iid1};
}
/*
DPCT1065:194: Consider replacing sycl::nd_item::barrier() with
sycl::nd_item::barrier(sycl::access::fence_space::local_space) for better
performance if there is no access to global memory.
*/
item_ct1.barrier();
const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
#pragma unroll
for (int i = item_ct1.get_local_id(2); i < ne10;
i += item_ct1.get_local_range(2)) {
src1_row_contiguous[i] = src1_row_original[i];
}
}
__dpct_inline__ static void k_copy_dst_from_contiguous(
char *__restrict__ dst_original, const char *__restrict__ dst_contiguous,
const mmid_row_mapping *__restrict__ row_mapping, int64_t ne0, size_t nb1,
size_t nb2, const sycl::nd_item<3> &item_ct1) {
int32_t i = item_ct1.get_group(2);
const int32_t i1 = row_mapping[i].i1;
const int32_t i2 = row_mapping[i].i2;
const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
#pragma unroll
for (int j = item_ct1.get_local_id(2); j < ne0;
j += item_ct1.get_local_range(2)) {
dst_row_original[j] = dst_row_contiguous[j];
}
}
static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
const ggml_tensor *src1,
ggml_tensor *dst) try {
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT &&
"mul_mat_id does not support split buffers");
GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer) && "mul_mat_id does not support split buffers");
const ggml_tensor *ids = dst->src[2];
GGML_TENSOR_BINARY_OP_LOCALS
const dpct::queue_ptr stream = g_syclStreams[g_main_device][0];
const size_t nb11 = src1->nb[1];
const size_t nb1 = dst->nb[1];
const int32_t id = ((int32_t *)dst->op_params)[0];
const int32_t n_as = src0->ne[2];
const int64_t n_as = ne02;
const int64_t n_ids = ids->ne[0];
std::vector<char> ids_host(ggml_nbytes(ids));
const char *ids_dev = (const char *)ids->data;
const char * ids_dev = (const char *) ids->data;
SYCL_CHECK(CHECK_TRY_ERROR(
stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids))));
@ -15489,25 +15618,41 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
src0_row.ne[2] = 1;
src0_row.ne[3] = 1;
src0_row.nb[3] = src0->nb[2];
src0_row.nb[3] = nb02;
if (src1->ne[1] == 1) {
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
const int32_t row_id =
*(const int32_t *)(ids_host.data() + i01 * ids->nb[1] +
id * ids->nb[0]);
src1_row.ne[1] = 1;
src1_row.ne[2] = 1;
src1_row.ne[3] = 1;
src1_row.nb[2] = nb11;
src1_row.nb[3] = nb11;
GGML_ASSERT(row_id >= 0 && row_id < n_as);
dst_row.ne[1] = 1;
dst_row.ne[2] = 1;
dst_row.ne[3] = 1;
dst_row.nb[2] = nb1;
dst_row.nb[3] = nb1;
if (ne12 == 1) {
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
for (int64_t id = 0; id < n_ids; id++) {
const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
GGML_ASSERT(i02 >= 0 && i02 < n_as);
const int64_t i11 = id % ne11;
const int64_t i12 = iid1;
const int64_t i1 = id;
const int64_t i2 = i12;
src0_row_extra.data_device[g_main_device] =
src0_original + row_id * src0->nb[2];
src0_original + i02*nb02;
src1_row_extra.data_device[g_main_device] =
src1_original + i01 * src1->nb[1];
src1_original + + i11*nb11 + i12*nb12;
dst_row_extra.data_device[g_main_device] =
dst_original + i01 * dst->nb[1];
dst_original + i1*nb1 + i2*nb2;
ggml_sycl_mul_mat(&src0_row, &src1_row, &dst_row);
}
}
} else {
sycl_pool_alloc<char> src1_contiguous(sizeof(float)*ggml_nelements(src1));
sycl_pool_alloc<char> dst_contiguous(sizeof(float)*ggml_nelements(dst));
@ -15515,64 +15660,98 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
src1_row_extra.data_device[g_main_device] = src1_contiguous.get();
dst_row_extra.data_device[g_main_device] = dst_contiguous.get();
for (int32_t row_id = 0; row_id < n_as; ++row_id) {
for (int64_t i02 = 0; i02 < n_as; i02++) {
int64_t num_src1_rows = 0;
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
for (int64_t id = 0; id < n_ids; id++) {
const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
if (row_id_i != row_id) {
GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
if (row_id_i != i02) {
continue;
}
GGML_ASSERT(row_id >= 0 && row_id < n_as);
SYCL_CHECK(CHECK_TRY_ERROR(
stream->memcpy(src1_contiguous.get() + num_src1_rows * nb11,
src1_original + i01 * nb11, nb11)));
num_src1_rows++;
}
}
if (num_src1_rows == 0) {
continue;
}
src0_row_extra.data_device[g_main_device] =
src0_original + row_id * src0->nb[2];
sycl_pool_alloc<int> dev_cur_src1_row(1);
sycl_pool_alloc<mmid_row_mapping> dev_row_mapping(num_src1_rows);
SYCL_CHECK(CHECK_TRY_ERROR(
stream->memset(dev_cur_src1_row.get(), 0, sizeof(int))));
{
sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u));
sycl::range<3> grid_dims(1, n_ids, ids->ne[1]);
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 0> src1_row_acc(cgh);
char *__restrict src1_contiguous_get =
src1_contiguous.get();
int *__restrict dev_cur_src1_row_get =
dev_cur_src1_row.get();
mmid_row_mapping *__restrict dev_row_mapping_get =
dev_row_mapping.get();
size_t ids_nb_ct6 = ids->nb[1];
size_t ids_nb_ct7 = ids->nb[0];
cgh.parallel_for(
sycl::nd_range<3>(grid_dims * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
k_copy_src1_to_contiguous(
src1_original, src1_contiguous_get,
dev_cur_src1_row_get,
dev_row_mapping_get, ids_dev, i02,
ids_nb_ct6, ids_nb_ct7, ne11, ne10, nb11, nb12,
item_ct1, src1_row_acc);
});
});
}
src0_row_extra.data_device[g_main_device] = src0_original + i02*nb02;
GGML_ASSERT(nb11 == sizeof(float)*ne10);
GGML_ASSERT(nb1 == sizeof(float)*ne0);
src1_row.ne[1] = num_src1_rows;
dst_row.ne[1] = num_src1_rows;
src1_row.nb[1] = nb11;
src1_row.nb[2] = num_src1_rows*nb11;
src1_row.nb[3] = num_src1_rows*nb11;
dst_row.ne[1] = num_src1_rows;
dst_row.nb[1] = nb1;
dst_row.nb[2] = num_src1_rows*nb1;
dst_row.nb[3] = num_src1_rows*nb1;
ggml_sycl_mul_mat(&src0_row, &src1_row, &dst_row);
num_src1_rows = 0;
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
{
sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u));
sycl::range<3> grid_dims(1, 1, num_src1_rows);
stream->submit([&](sycl::handler &cgh) {
const char *__restrict dst_contiguous_get =
dst_contiguous.get();
const mmid_row_mapping *__restrict dev_row_mapping_get =
dev_row_mapping.get();
if (row_id_i != row_id) {
continue;
}
GGML_ASSERT(row_id >= 0 && row_id < n_as);
SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(
dst_original + i01 * nb1,
dst_contiguous.get() + num_src1_rows * nb1, nb1)));
num_src1_rows++;
cgh.parallel_for(
sycl::nd_range<3>(grid_dims * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
k_copy_dst_from_contiguous(dst_original,
dst_contiguous_get,
dev_row_mapping_get,
ne0, nb1, nb2, item_ct1);
});
});
}
}
}
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
SYCL_CHECK(CHECK_TRY_ERROR(stream->wait()));
}
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@ -16555,10 +16734,9 @@ GGML_CALL static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backe
UNUSED(buffer);
}
// unused at the moment
//static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) {
// return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name;
//}
static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name;
}
GGML_CALL static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;

2
ggml-vulkan.cpp
View file

@ -6139,6 +6139,8 @@ static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
};
GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num) {
ggml_vk_instance_init();
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_backend_vk_buffer_type(" << dev_num << ")" << std::endl;
#endif

178
ggml.c
View file

@ -60,6 +60,9 @@
typedef volatile LONG atomic_int;
typedef atomic_int atomic_bool;
typedef atomic_int atomic_flag;
#define ATOMIC_FLAG_INIT 0
static void atomic_store(atomic_int * ptr, LONG val) {
InterlockedExchange(ptr, val);
@ -73,6 +76,12 @@ static LONG atomic_fetch_add(atomic_int * ptr, LONG inc) {
static LONG atomic_fetch_sub(atomic_int * ptr, LONG dec) {
return atomic_fetch_add(ptr, -(dec));
}
static atomic_bool atomic_flag_test_and_set(atomic_flag * ptr) {
return InterlockedExchange(ptr, 1);
}
static void atomic_flag_clear(atomic_flag * ptr) {
InterlockedExchange(ptr, 0);
}
typedef HANDLE pthread_t;
@ -2883,24 +2892,20 @@ struct ggml_state {
// global state
static struct ggml_state g_state;
static atomic_int g_state_barrier = 0;
static atomic_flag g_state_critical = ATOMIC_FLAG_INIT;
// barrier via spin lock
inline static void ggml_critical_section_start(void) {
int processing = atomic_fetch_add(&g_state_barrier, 1);
while (processing > 0) {
// wait for other threads to finish
atomic_fetch_sub(&g_state_barrier, 1);
sched_yield(); // TODO: reconsider this
processing = atomic_fetch_add(&g_state_barrier, 1);
while (atomic_flag_test_and_set(&g_state_critical)) {
// spin
sched_yield();
}
}
// TODO: make this somehow automatically executed
// some sort of "sentry" mechanism
inline static void ggml_critical_section_end(void) {
atomic_fetch_sub(&g_state_barrier, 1);
atomic_flag_clear(&g_state_critical);
}
#if defined(__gnu_linux__)
@ -3216,7 +3221,11 @@ GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
}
static inline bool ggml_is_contiguous_except_dim_1(const struct ggml_tensor * tensor) {
GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) {
return ggml_is_contiguous(tensor);
}
GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return
@ -3225,6 +3234,14 @@ static inline bool ggml_is_contiguous_except_dim_1(const struct ggml_tensor * te
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
}
GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return
tensor->nb[0] == ggml_type_size(tensor->type) &&
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
}
GGML_CALL bool ggml_is_permuted(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
@ -4882,10 +4899,21 @@ struct ggml_tensor * ggml_repeat_back(
// ggml_concat
struct ggml_tensor * ggml_concat(
struct ggml_context* ctx,
struct ggml_tensor* a,
struct ggml_tensor* b) {
GGML_ASSERT(a->ne[0] == b->ne[0] && a->ne[1] == b->ne[1] && a->ne[3] == b->ne[3]);
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int dim) {
GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS);
int64_t ne[GGML_MAX_DIMS];
for (int d = 0; d < GGML_MAX_DIMS; ++d) {
if (d == dim) {
ne[d] = a->ne[d] + b->ne[d];
continue;
}
GGML_ASSERT(a->ne[d] == b->ne[d]);
ne[d] = a->ne[d];
}
bool is_node = false;
@ -4893,7 +4921,9 @@ struct ggml_tensor * ggml_concat(
is_node = true;
}
struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, a->ne[0], a->ne[1], a->ne[2] + b->ne[2], a->ne[3]);
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne);
ggml_set_op_params_i32(result, 0, dim);
result->op = GGML_OP_CONCAT;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -5013,6 +5043,7 @@ struct ggml_tensor * ggml_leaky_relu(
}
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params(result, &negative_slope, sizeof(negative_slope));
result->op = GGML_OP_LEAKY_RELU;
@ -6378,6 +6409,16 @@ struct ggml_tensor * ggml_rope_custom_inplace(
);
}
struct ggml_tensor * ggml_rope_xpos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int n_dims,
float base,
bool down) {
return ggml_rope_impl(ctx, a, b, NULL, n_dims, 0, 0, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, base, down, true);
}
// ggml_rope_back
struct ggml_tensor * ggml_rope_back(
@ -10967,35 +11008,38 @@ static void ggml_compute_forward_concat_f32(
GGML_ASSERT(nb00 == sizeof(float));
GGML_ASSERT(nb10 == sizeof(float));
const int32_t dim = ggml_get_op_params_i32(dst, 0);
GGML_ASSERT(dim >= 0 && dim < 4);
int64_t o[4] = {0, 0, 0, 0};
o[dim] = src0->ne[dim];
const float * x;
// TODO: smarter multi-theading
for (int i3 = 0; i3 < ne3; i3++) {
for (int i2 = ith; i2 < ne2; i2 += nth) {
if (i2 < ne02) { // src0
for (int i1 = 0; i1 < ne1; i1++) {
for (int i0 = 0; i0 < ne0; i0++) {
const float * x = (float *)((char *) src0->data + i0 * nb00 + i1 * nb01 + i2 * nb02 + i3 * nb03);
if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
x = (const float *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03);
} else {
x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
}
float * y = (float *)((char *)dst->data + i0 * nb0 + i1 * nb1 + i2 * nb2 + i3 * nb3);
*y = *x;
}
}
} // src1
else {
for (int i1 = 0; i1 < ne1; i1++) {
for (int i0 = 0; i0 < ne0; i0++) {
const float * x = (float *)((char *) src1->data + i0 * nb10 + i1 * nb11 + (i2 - ne02) * nb12 + i3 * nb13);
float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
float * y = (float *)((char *)dst->data + i0 * nb0 + i1 * nb1 + i2 * nb2 + i3 * nb3);
*y = *x;
}
}
}
}
}
}
static void ggml_compute_forward_concat(
const struct ggml_compute_params* params,
struct ggml_tensor* dst) {
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0];
@ -11388,8 +11432,8 @@ static void ggml_compute_forward_gelu_f32(
const struct ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@ -11451,8 +11495,8 @@ static void ggml_compute_forward_gelu_quick_f32(
const struct ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@ -11514,8 +11558,8 @@ static void ggml_compute_forward_silu_f32(
const struct ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
@ -11626,9 +11670,9 @@ static void ggml_compute_forward_silu_back_f32(
const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * grad = dst->src[1];
GGML_ASSERT(ggml_is_contiguous_except_dim_1(grad));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
GGML_ASSERT(ggml_is_contiguous_1(grad));
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
GGML_ASSERT(ggml_are_same_shape(src0, grad));
@ -14326,7 +14370,7 @@ static void ggml_compute_forward_rope_f32(
int ir = 0;
const float theta_scale = powf(freq_base, -2.0f/n_dims);
const float inv_ndims = -1.f/n_dims;
float corr_dims[2];
ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
@ -14410,29 +14454,22 @@ static void ggml_compute_forward_rope_f32(
dst_data[1] = x0*sin_theta*zeta + x1*cos_theta*zeta;
}
} else {
// TODO: this might be wrong for ne0 != n_dims - need double check
// it seems we have to rope just the first n_dims elements and do nothing with the rest
// ref: https://github.com/ml-explore/mlx/blob/dc2edc762c797e3b8de50b1dad4dc0a131691033/benchmarks/python/llama_jax_bench.py#L11-L26
theta_base *= freq_scale;
// ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
for (int64_t ic = 0; ic < ne0; ic += 2) {
if (ic < n_dims) {
const int64_t ib = 0;
const int64_t i0 = ic/2;
// simplified from `(ib * n_dims + ic) * inv_ndims`
float cur_rot = inv_ndims * ic - ib;
float freq_factor = freq_factors ? freq_factors[ic/2] : 1.0f;
const float freq_factor = freq_factors ? freq_factors[i0] : 1.0f;
float cos_theta, sin_theta;
rope_yarn(
theta_base/freq_factor, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor,
theta_base/freq_factor, freq_scale, corr_dims, ic, ext_factor, attn_factor,
&cos_theta, &sin_theta
);
sin_theta *= sin_sign;
theta_base *= theta_scale;
const int64_t i0 = ib*n_dims + ic/2;
const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
@ -14511,7 +14548,7 @@ static void ggml_compute_forward_rope_f16(
int ir = 0;
const float theta_scale = powf(freq_base, -2.0f/n_dims);
const float inv_ndims = -1.f/n_dims;
float corr_dims[2];
ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
@ -14591,29 +14628,22 @@ static void ggml_compute_forward_rope_f16(
dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
}
} else {
// TODO: this might be wrong for ne0 != n_dims - need double check
// it seems we have to rope just the first n_dims elements and do nothing with the rest
// ref: https://github.com/ml-explore/mlx/blob/dc2edc762c797e3b8de50b1dad4dc0a131691033/benchmarks/python/llama_jax_bench.py#L11-L26
theta_base *= freq_scale;
// ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
for (int64_t ic = 0; ic < ne0; ic += 2) {
if (ic < n_dims) {
const int64_t ib = 0;
const int64_t i0 = ic/2;
// simplified from `(ib * n_dims + ic) * inv_ndims`
float cur_rot = inv_ndims * ic - ib;
float freq_factor = freq_factors ? freq_factors[ic/2] : 1.0f;
const float freq_factor = freq_factors ? freq_factors[i0] : 1.0f;
float cos_theta, sin_theta;
rope_yarn(
theta_base/freq_factor, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor,
theta_base/freq_factor, freq_scale, corr_dims, ic, ext_factor, attn_factor,
&cos_theta, &sin_theta
);
sin_theta *= sin_sign;
theta_base *= theta_scale;
const int64_t i0 = ib*n_dims + ic/2;
const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
@ -22742,6 +22772,16 @@ int ggml_cpu_has_neon(void) {
#endif
}
int ggml_cpu_has_sve(void) {
#if defined(__ARM_FEATURE_SVE)
// TODO: Currently, SVE 256 bit is only supported.
GGML_ASSERT(svcntb() == QK8_0);
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_arm_fma(void) {
#if defined(__ARM_FEATURE_FMA)
return 1;
@ -22830,6 +22870,14 @@ int ggml_cpu_has_sycl(void) {
#endif
}
int ggml_cpu_has_rpc(void) {
#if defined(GGML_USE_RPC)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_gpublas(void) {
return ggml_cpu_has_cuda() || ggml_cpu_has_clblast() || ggml_cpu_has_vulkan() || ggml_cpu_has_kompute() ||
ggml_cpu_has_sycl();

21
ggml.h
View file

@ -756,7 +756,6 @@ extern "C" {
GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
GGML_API GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor);
GGML_API GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor);
GGML_API GGML_CALL bool ggml_is_permuted (const struct ggml_tensor * tensor);
GGML_API GGML_CALL bool ggml_is_empty (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_scalar (const struct ggml_tensor * tensor);
@ -765,6 +764,11 @@ extern "C" {
GGML_API bool ggml_is_3d (const struct ggml_tensor * tensor);
GGML_API int ggml_n_dims (const struct ggml_tensor * tensor); // returns 1 for scalars
GGML_API GGML_CALL bool ggml_is_contiguous (const struct ggml_tensor * tensor);
GGML_API GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
GGML_API GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
GGML_API GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
@ -1007,12 +1011,13 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
// concat a and b on dim 2
// concat a and b along dim
// used in stable-diffusion
GGML_API struct ggml_tensor * ggml_concat(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * b,
int dim);
GGML_API struct ggml_tensor * ggml_abs(
struct ggml_context * ctx,
@ -1547,6 +1552,14 @@ extern "C" {
float beta_slow),
"use ggml_rope_ext_inplace instead");
struct ggml_tensor * ggml_rope_xpos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int n_dims,
float base,
bool down);
// compute correction dims for YaRN RoPE scaling
GGML_CALL void ggml_rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
@ -2404,6 +2417,7 @@ extern "C" {
GGML_API int ggml_cpu_has_avx512_bf16(void);
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_sve (void);
GGML_API int ggml_cpu_has_arm_fma (void);
GGML_API int ggml_cpu_has_metal (void);
GGML_API int ggml_cpu_has_f16c (void);
@ -2418,6 +2432,7 @@ extern "C" {
GGML_API int ggml_cpu_has_sse3 (void);
GGML_API int ggml_cpu_has_ssse3 (void);
GGML_API int ggml_cpu_has_sycl (void);
GGML_API int ggml_cpu_has_rpc (void);
GGML_API int ggml_cpu_has_vsx (void);
GGML_API int ggml_cpu_has_matmul_int8(void);

4
ggml_vk_generate_shaders.py
View file

@ -2533,14 +2533,12 @@ void main() {
const uint i = row*p.ncols + ib*p.ndims + ic/2;
const uint i2 = row/p.p_delta_rows;
const float cur_rot = p.inv_ndims * ic - ib;
const int pos = data_b[i2];
const float freq_factor = p.has_freq_facs != 0 ? data_freq_factors[ic/2] : 1.0f;
const float theta_base = pos*p.freq_scale*pow(p.theta_scale, col/2.0f) / freq_factor;
float cos_theta, sin_theta;
rope_yarn(theta_base, uint(cur_rot), cos_theta, sin_theta);
rope_yarn(theta_base, ic, cos_theta, sin_theta);
const float x0 = float(data_a[i + 0]);
const float x1 = float(data_a[i + p.ndims/2]);

52
gguf-py/gguf/constants.py
View file

@ -37,11 +37,15 @@ class Keys:
CONTEXT_LENGTH = "{arch}.context_length"
EMBEDDING_LENGTH = "{arch}.embedding_length"
BLOCK_COUNT = "{arch}.block_count"
LEADING_DENSE_BLOCK_COUNT = "{arch}.leading_dense_block_count"
FEED_FORWARD_LENGTH = "{arch}.feed_forward_length"
EXPERT_FEED_FORWARD_LENGTH = "{arch}.expert_feed_forward_length"
USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout"
EXPERT_COUNT = "{arch}.expert_count"
EXPERT_USED_COUNT = "{arch}.expert_used_count"
EXPERT_SHARED_COUNT = "{arch}.expert_shared_count"
EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale"
POOLING_TYPE = "{arch}.pooling_type"
LOGIT_SCALE = "{arch}.logit_scale"
@ -55,6 +59,8 @@ class Keys:
LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon"
LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
CAUSAL = "{arch}.attention.causal"
Q_LORA_RANK = "{arch}.attention.q_lora_rank"
KV_LORA_RANK = "{arch}.attention.kv_lora_rank"
class Rope:
DIMENSION_COUNT = "{arch}.rope.dimension_count"
@ -64,6 +70,7 @@ class Keys:
SCALING_ATTN_FACTOR = "{arch}.rope.scaling.attn_factor"
SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length"
SCALING_FINETUNED = "{arch}.rope.scaling.finetuned"
SCALING_YARN_LOG_MUL = "{arch}.rope.scaling.yarn_log_multiplier"
class SSM:
CONV_KERNEL = "{arch}.ssm.conv_kernel"
@ -140,6 +147,7 @@ class MODEL_ARCH(IntEnum):
DBRX = auto()
OLMO = auto()
ARCTIC = auto()
DEEPSEEK2 = auto()
class MODEL_TENSOR(IntEnum):
@ -185,6 +193,12 @@ class MODEL_TENSOR(IntEnum):
SSM_A = auto()
SSM_D = auto()
SSM_OUT = auto()
ATTN_Q_A = auto()
ATTN_Q_B = auto()
ATTN_KV_A_MQA = auto()
ATTN_KV_B = auto()
ATTN_Q_A_NORM = auto()
ATTN_KV_A_NORM = auto()
MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@ -221,6 +235,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.DBRX: "dbrx",
MODEL_ARCH.OLMO: "olmo",
MODEL_ARCH.ARCTIC: "arctic",
MODEL_ARCH.DEEPSEEK2: "deepseek2",
}
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@ -266,6 +281,12 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a",
MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d",
MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out",
MODEL_TENSOR.ATTN_Q_A: "blk.{bid}.attn_q_a",
MODEL_TENSOR.ATTN_Q_B: "blk.{bid}.attn_q_b",
MODEL_TENSOR.ATTN_KV_A_MQA: "blk.{bid}.attn_kv_a_mqa",
MODEL_TENSOR.ATTN_KV_B: "blk.{bid}.attn_kv_b",
MODEL_TENSOR.ATTN_Q_A_NORM: "blk.{bid}.attn_q_a_norm",
MODEL_TENSOR.ATTN_KV_A_NORM: "blk.{bid}.attn_kv_a_norm",
}
MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@ -757,6 +778,33 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.DEEPSEEK2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_A,
MODEL_TENSOR.ATTN_Q_B,
MODEL_TENSOR.ATTN_KV_A_MQA,
MODEL_TENSOR.ATTN_KV_B,
MODEL_TENSOR.ATTN_Q_A_NORM,
MODEL_TENSOR.ATTN_KV_A_NORM,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
],
# TODO
}
@ -790,6 +838,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
MODEL_ARCH.DEEPSEEK2: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
}
#

21
gguf-py/gguf/gguf_writer.py
View file

@ -374,9 +374,15 @@ class GGUFWriter:
def add_block_count(self, length: int) -> None:
self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length)
def add_leading_dense_block_count(self, length: int) -> None:
self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length)
def add_feed_forward_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_expert_feed_forward_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_parallel_residual(self, use: bool) -> None:
self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
@ -407,6 +413,12 @@ class GGUFWriter:
def add_expert_used_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
def add_expert_shared_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_SHARED_COUNT.format(arch=self.arch), count)
def add_expert_weights_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
def add_layer_norm_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
@ -416,6 +428,12 @@ class GGUFWriter:
def add_causal_attention(self, value: bool) -> None:
self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
def add_q_lora_rank(self, length: int) -> None:
self.add_uint32(Keys.Attention.Q_LORA_RANK.format(arch=self.arch), length)
def add_kv_lora_rank(self, length: int) -> None:
self.add_uint32(Keys.Attention.KV_LORA_RANK.format(arch=self.arch), length)
def add_pooling_type(self, value: PoolingType) -> None:
self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
@ -440,6 +458,9 @@ class GGUFWriter:
def add_rope_scaling_finetuned(self, value: bool) -> None:
self.add_bool(Keys.Rope.SCALING_FINETUNED.format(arch=self.arch), value)
def add_rope_scaling_yarn_log_mul(self, value: float) -> None:
self.add_float32(Keys.Rope.SCALING_YARN_LOG_MUL.format(arch=self.arch), value)
def add_ssm_conv_kernel(self, value: int) -> None:
self.add_uint32(Keys.SSM.CONV_KERNEL.format(arch=self.arch), value)

29
gguf-py/gguf/tensor_mapping.py
View file

@ -256,6 +256,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_UP_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek2
),
# AWQ-activation gate
@ -285,6 +286,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_GATE_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek2
),
# Feed-forward down
@ -320,6 +322,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_DOWN_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek2
),
MODEL_TENSOR.ATTN_Q_NORM: (
@ -383,6 +386,30 @@ class TensorNameMap:
"model.layers.{bid}.out_proj",
"backbone.layers.{bid}.mixer.out_proj",
),
MODEL_TENSOR.ATTN_Q_A: (
"model.layers.{bid}.self_attn.q_a_proj", # deepseek2
),
MODEL_TENSOR.ATTN_Q_B: (
"model.layers.{bid}.self_attn.q_b_proj", # deepseek2
),
MODEL_TENSOR.ATTN_KV_A_MQA: (
"model.layers.{bid}.self_attn.kv_a_proj_with_mqa", # deepseek2
),
MODEL_TENSOR.ATTN_KV_B: (
"model.layers.{bid}.self_attn.kv_b_proj", # deepseek2
),
MODEL_TENSOR.ATTN_Q_A_NORM: (
"model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2
),
MODEL_TENSOR.ATTN_KV_A_NORM: (
"model.layers.{bid}.self_attn.kv_a_layernorm", # deepseek2
),
}
# architecture-specific block mappings
@ -415,7 +442,7 @@ class TensorNameMap:
if tensor not in MODEL_TENSORS[arch]:
continue
# TODO: make this configurable
n_experts = 128
n_experts = 160
for xid in range(n_experts):
tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
self.mapping[tensor_name] = (tensor, tensor_name)

689
llama.cpp
View file

File diff suppressed because it is too large Load diff

12
llama.h
View file

@ -85,6 +85,7 @@ extern "C" {
LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11,
LLAMA_VOCAB_PRE_TYPE_OLMO = 12,
LLAMA_VOCAB_PRE_TYPE_DBRX = 13,
LLAMA_VOCAB_PRE_TYPE_SMAUG = 14,
};
// note: these values should be synchronized with ggml_rope
@ -264,6 +265,8 @@ extern "C" {
bool check_tensors; // validate model tensor data
};
// NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
// https://github.com/ggerganov/llama.cpp/pull/7544
struct llama_context_params {
uint32_t seed; // RNG seed, -1 for random
uint32_t n_ctx; // text context, 0 = from model
@ -290,14 +293,14 @@ extern "C" {
ggml_backend_sched_eval_callback cb_eval;
void * cb_eval_user_data;
enum ggml_type type_k; // data type for K cache
enum ggml_type type_v; // data type for V cache
enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
// Keep the booleans together to avoid misalignment during copy-by-value.
bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
bool embeddings; // if true, extract embeddings (together with logits)
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
bool flash_attn; // whether to use flash attention
bool flash_attn; // whether to use flash attention [EXPERIMENTAL]
// Abort callback
// if it returns true, execution of llama_decode() will be aborted
@ -823,6 +826,9 @@ extern "C" {
// Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
// Identify if Token Id is a control token or a render-able token
LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token);
// Special tokens
LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence

4
scripts/sync-ggml-am.sh
View file

@ -106,8 +106,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
# src/ggml-kompute.h -> ggml-kompute.h
# src/ggml-metal.h -> ggml-metal.h
# src/ggml-metal.m -> ggml-metal.m
# src/ggml-mpi.h -> ggml-mpi.h
# src/ggml-mpi.c -> ggml-mpi.c
# src/ggml-opencl.cpp -> ggml-opencl.cpp
# src/ggml-opencl.h -> ggml-opencl.h
# src/ggml-quants.c -> ggml-quants.c
@ -145,8 +143,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
-e 's/src\/ggml-kompute\.h/ggml-kompute.h/g' \
-e 's/src\/ggml-metal\.h/ggml-metal.h/g' \
-e 's/src\/ggml-metal\.m/ggml-metal.m/g' \
-e 's/src\/ggml-mpi\.h/ggml-mpi.h/g' \
-e 's/src\/ggml-mpi\.c/ggml-mpi.c/g' \
-e 's/src\/ggml-opencl\.cpp/ggml-opencl.cpp/g' \
-e 's/src\/ggml-opencl\.h/ggml-opencl.h/g' \
-e 's/src\/ggml-quants\.c/ggml-quants.c/g' \

2
scripts/sync-ggml.last
View file

@ -1 +1 @@
126d34985705a5a2222723c145cb4e125ac689f3
2aae01fd9b8f9399f343cf18f46f38996ef52e2c

2
scripts/sync-ggml.sh
View file

@ -14,8 +14,6 @@ cp -rpv ../ggml/src/ggml-kompute.h ./ggml-kompute.h
cp -rpv ../ggml/src/ggml-metal.h ./ggml-metal.h
cp -rpv ../ggml/src/ggml-metal.m ./ggml-metal.m
cp -rpv ../ggml/src/ggml-metal.metal ./ggml-metal.metal
cp -rpv ../ggml/src/ggml-mpi.h ./ggml-mpi.h
cp -rpv ../ggml/src/ggml-mpi.c ./ggml-mpi.c
cp -rpv ../ggml/src/ggml-opencl.cpp ./ggml-opencl.cpp
cp -rpv ../ggml/src/ggml-opencl.h ./ggml-opencl.h
cp -rpv ../ggml/src/ggml-quants.c ./ggml-quants.c

116
tests/test-backend-ops.cpp
View file

@ -1138,26 +1138,37 @@ struct test_soft_max : public test_case {
// GGML_OP_ROPE
struct test_rope : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
const std::array<int64_t, 4> ne_a;
int n_dims;
int mode;
int n_ctx;
float fs; // freq_scale
float ef; // ext_factor
float af; // attn_factor
bool ff;
int v; // view (1 : non-contiguous a)
std::string vars() override {
return VARS_TO_STR6(type, ne, n_dims, mode, n_ctx, ff);
return VARS_TO_STR10(type, ne_a, n_dims, mode, n_ctx, fs, ef, af, ff, v);
}
test_rope(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 1},
int n_dims = 10, int mode = 0, int n_ctx = 512, bool ff = false)
: type(type), ne(ne), n_dims(n_dims), mode(mode), n_ctx(n_ctx), ff(ff) {}
std::array<int64_t, 4> ne_a = {10, 10, 10, 1},
int n_dims = 10, int mode = 0, int n_ctx = 512, float fs = 1.0f, float ef = 0.0f, float af = 0.0f, bool ff = false, int v = 0)
: type(type), ne_a(ne_a), n_dims(n_dims), mode(mode), n_ctx(n_ctx), fs(fs), ef(ef), af(af), ff(ff), v(v) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2]);
ggml_tensor * a;
if (v & 1) {
auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3;
a = ggml_new_tensor(ctx, type, 4, ne.data());
a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0);
} else {
a = ggml_new_tensor(ctx, type, 4, ne_a.data());
}
ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne_a[2]);
ggml_tensor * freq = ff ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_dims/2) : nullptr;
ggml_tensor * out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
ggml_tensor * out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, n_ctx, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
return out;
}
@ -1165,11 +1176,11 @@ struct test_rope : public test_case {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (t->type == GGML_TYPE_I32) {
// pos
std::vector<int> data(ne[2]);
for (int i = 0; i < ne[2]; i++) {
std::vector<int> data(ne_a[2]);
for (int i = 0; i < ne_a[2]; i++) {
data[i] = rand() % n_ctx;
}
ggml_backend_tensor_set(t, data.data(), 0, ne[2] * sizeof(int));
ggml_backend_tensor_set(t, data.data(), 0, ne_a[2] * sizeof(int));
} else {
if (t->ne[0] == n_dims/2) {
// frequency factors in the range [0.9f, 1.1f]
@ -1259,22 +1270,41 @@ struct test_im2col : public test_case {
// GGML_OP_CONCAT
struct test_concat : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
const int64_t b_ne2;
const std::array<int64_t, 4> ne_a;
const int64_t ne_b_d;
const int dim;
const int v; // view (1 << 0: non-cont a, 1 << 1: non-cont b)
std::string vars() override {
return VARS_TO_STR3(type, ne, b_ne2);
return VARS_TO_STR5(type, ne_a, ne_b_d, dim, v);
}
test_concat(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 10},
int64_t b_ne2 = 10)
: type(type), ne(ne), b_ne2(b_ne2) {}
std::array<int64_t, 4> ne_a = {10, 10, 10, 10},
int64_t ne_b_d = 10,
int dim = 2, int v = 0)
: type(type), ne_a(ne_a), ne_b_d(ne_b_d), dim(dim), v(v) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_tensor * b = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], b_ne2, ne[3]);
ggml_tensor * out = ggml_concat(ctx, a, b);
auto ne_b = ne_a;
ne_b[dim] = ne_b_d;
ggml_tensor * a;
if (v & 1) {
auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3;
a = ggml_new_tensor(ctx, type, 4, ne.data());
a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0);
} else {
a = ggml_new_tensor(ctx, type, 4, ne_a.data());
}
ggml_tensor * b;
if (v & 2) {
auto ne = ne_b; ne[0] *= 3; ne[1] *= 2; ne[2] *= 4;
b = ggml_new_tensor(ctx, type, 4, ne.data());
b = ggml_view_4d(ctx, b, ne_b[0], ne_b[1], ne_b[2], ne_b[3], b->nb[1], b->nb[2], b->nb[3], 0);
} else {
b = ggml_new_tensor(ctx, type, 4, ne_b.data());
}
ggml_tensor * out = ggml_concat(ctx, a, b, dim);
return out;
}
};
@ -2194,25 +2224,47 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 8.0f));
{
bool all = true;
for (float v : { 0, 1 }) {
for (float fs : { 1.0f, 1.4245f }) {
for (float ef : { 0.0f, 0.7465f }) {
for (float af : { 1.0f, 1.4245f }) {
for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
// TODO: ff not supported yet for !neox
test_cases.emplace_back(new test_rope(type, {128, 32, 10, 1}, 128, 0, 512, false)); // llama 7B
test_cases.emplace_back(new test_rope(type, {128, 40, 10, 1}, 128, 0, 512, false)); // llama 13B
test_cases.emplace_back(new test_rope(type, {128, 52, 10, 1}, 128, 0, 512, false)); // llama 30B
test_cases.emplace_back(new test_rope(type, {128, 64, 10, 1}, 128, 0, 512, false)); // llama 65B
test_cases.emplace_back(new test_rope(type, {128, 32, 10, 1}, 128, 0, 512, fs, ef, af, false, v)); // llama 7B
if (all) {
test_cases.emplace_back(new test_rope(type, {128, 40, 10, 1}, 128, 0, 512, fs, ef, af, false, v)); // llama 13B
test_cases.emplace_back(new test_rope(type, {128, 52, 10, 1}, 128, 0, 512, fs, ef, af, false, v)); // llama 30B
test_cases.emplace_back(new test_rope(type, {128, 64, 10, 1}, 128, 0, 512, fs, ef, af, false, v)); // llama 65B
}
for (bool ff : {false, true}) { // freq_factors
test_cases.emplace_back(new test_rope(type, { 64, 1, 10, 1}, 64, 2, 512, ff)); // neox (falcon 7B)
test_cases.emplace_back(new test_rope(type, { 64, 71, 10, 1}, 64, 2, 512, ff)); // neox (falcon 7B)
test_cases.emplace_back(new test_rope(type, { 64, 8, 10, 1}, 64, 2, 512, ff)); // neox (falcon 40B)
test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1}, 64, 2, 512, ff)); // neox (falcon 40B)
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 20, 2, 512, ff)); // neox (stablelm)
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 32, 2, 512, ff)); // neox (phi-2)
if (all) {
test_cases.emplace_back(new test_rope(type, { 64, 1, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 7B)
test_cases.emplace_back(new test_rope(type, { 64, 71, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 7B)
test_cases.emplace_back(new test_rope(type, { 64, 8, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 40B)
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 20, 2, 512, fs, ef, af, ff, v)); // neox (stablelm)
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 32, 2, 512, fs, ef, af, ff, v)); // neox (phi-2)
}
test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 40B)
}
}
all = false;
}
}
}
}
}
test_cases.emplace_back(new test_concat(GGML_TYPE_F32));
test_cases.emplace_back(new test_concat(GGML_TYPE_I32));
for (int v : { 0, 1, 2, 3 }) {
for (int dim : { 0, 1, 2, 3, }) {
test_cases.emplace_back(new test_concat(GGML_TYPE_F32, {11, 12, 13, 14}, 7, dim, v));
test_cases.emplace_back(new test_concat(GGML_TYPE_I32, {11, 12, 13, 14}, 7, dim, v));
}
}
for (ggml_sort_order order : {GGML_SORT_ORDER_ASC, GGML_SORT_ORDER_DESC}) {
test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {8, 1, 1, 1}, order));

2
tests/test-tokenizer-0.sh
View file

@ -28,6 +28,8 @@ printf "Tokenizing using (cpp) llama.cpp ...\n"
cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in"
cat /tmp/test-tokenizer-0-$name-cpp.log | grep "tokenized in"
set +e
diff $input.tok $input.tokcpp > /dev/null 2>&1
if [ $? -eq 0 ]; then

20
tests/test-tokenizer-random.py
View file

@ -167,8 +167,10 @@ def generator_random_special_tokens(tokenizer, iterations=100) -> Iterator[str]:
for m in range(iterations):
rand.seed(m)
words = rand.choices(special_tokens, k=500)
if tokenizer.add_bos_token: # skip spam warning of double BOS
while words and words[0] == tokenizer.bos_token:
if words[0] == tokenizer.bos_token: # skip spam warning of double BOS
while len(words) > 1 and words[1] == tokenizer.bos_token: # leave one starting BOS
words.pop(0)
if tokenizer.add_bos_token: # drop all starting BOS
words.pop(0)
yield "".join(words)
@ -293,15 +295,17 @@ def main(argv: list[str] = None):
model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=4096))
tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer)
tokenizer.add_bos_token = getattr(tokenizer, "add_bos_token", True)
tokenizer.add_eos_token = getattr(tokenizer, "add_eos_token", False)
def func_tokenize1(text: str):
return model.tokenize(text, add_special=True, parse_special=True)
def func_tokenize2(text: str):
return tokenizer.encode(text, add_special_tokens=True)
ids = func_tokenize2("a")
assert 1 <= len(ids) <= 3
add_bos_token = len(ids) > 1 and tokenizer.bos_token_id == ids[0]
tokenizer.add_bos_token = getattr(tokenizer, "add_bos_token", add_bos_token)
vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True)))
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text())
test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases())
@ -324,8 +328,10 @@ if __name__ == "__main__":
# import os
# tokenizers = os.listdir(path_tokenizers)
tokenizers = [
"llama-spm", # SPM
"phi-3", # SPM
# "llama-spm", # SPM
# "phi-3", # SPM
"jina-v2-en", # WPM
"bert-bge", # WPM
]
for tokenizer in tokenizers: