Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-02-10 02:03:07 +01:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/master' into jinja

Browse Source
This commit is contained in:
ochafik 2025-01-13 19:56:27 +00:00
commit cb72cf1fc3
215 changed files with 23423 additions and 18704 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
.github/ISSUE_TEMPLATE/010-bug-compilation.yml vendored
View File

@ -65,12 +65,22 @@ body:
If possible, please do a git bisect and identify the exact commit that introduced the bug.
validations:
required: false
- type: textarea
id: command
attributes:
label: Compile command
description: >
Please provide the exact command you used to compile llama.cpp. For example: `cmake -B ...`.
This will be automatically formatted into code, so no need for backticks.
render: shell
validations:
required: true
- type: textarea
id: logs
attributes:
label: Relevant log output
description: >
Please copy and paste any relevant log output, including the command that you entered and any generated text.
Please copy and paste any relevant log output, including any generated text.
This will be automatically formatted into code, so no need for backticks.
render: shell
validations:

12
.github/ISSUE_TEMPLATE/019-bug-misc.yml vendored
View File

@ -52,6 +52,16 @@ body:
- Other (Please specify in the next section)
validations:
required: false
- type: textarea
id: command
attributes:
label: Command line
description: >
Please provide the exact commands you entered, if applicable. For example: `llama-server -m ... -c ...`, `llama-cli -m ...`, etc.
This will be automatically formatted into code, so no need for backticks.
render: shell
validations:
required: false
- type: textarea
id: info
attributes:
@ -74,7 +84,7 @@ body:
attributes:
label: Relevant log output
description: >
If applicable, please copy and paste any relevant log output, including the command that you entered and any generated text.
If applicable, please copy and paste any relevant log output, including any generated text.
This will be automatically formatted into code, so no need for backticks.
render: shell
validations:

30
.github/workflows/build.yml vendored
View File

@ -60,8 +60,7 @@ jobs:
-DLLAMA_CURL=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DGGML_RPC=ON \
-DBUILD_SHARED_LIBS=OFF
-DGGML_RPC=ON
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -123,8 +122,7 @@ jobs:
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL=OFF \
-DGGML_RPC=ON \
-DBUILD_SHARED_LIBS=OFF
-DGGML_RPC=ON
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -181,7 +179,7 @@ jobs:
run: |
mkdir build
cd build
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON
cmake --build . --config Release -j $(nproc)
- name: Test
@ -651,23 +649,23 @@ jobs:
matrix:
include:
- build: 'noavx-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
- build: 'avx2-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON'
- build: 'avx-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF'
- build: 'avx512-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON'
- build: 'openblas-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
- build: 'kompute-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
- build: 'vulkan-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON'
- build: 'llvm-arm64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
- build: 'msvc-arm64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
- build: 'llvm-arm64-opencl-adreno'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
@ -914,7 +912,7 @@ jobs:
shell: cmd
run: |
call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DGGML_RPC=ON
set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
cmake --build build --config Release -j %NINJA_JOBS% -t ggml
cmake --build build --config Release
@ -1239,7 +1237,7 @@ jobs:
- name: Create release
id: create_release
uses: anzz1/action-create-release@v1
uses: ggml-org/action-create-release@v1
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
with:

3
.github/workflows/docker.yml vendored
View File

@ -97,10 +97,9 @@ jobs:
GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
# https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
- name: Free Disk Space (Ubuntu)
if: ${{ matrix.config.free_disk_space == true }}
uses: jlumbroso/free-disk-space@main
uses: ggml-org/free-disk-space@v1.3.1
with:
# this might remove tools that are actually needed,
# if set to "true" but frees about 6 GB

4
.github/workflows/editorconfig.yml vendored
View File

@ -23,5 +23,7 @@ jobs:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: editorconfig-checker/action-editorconfig-checker@main
- uses: editorconfig-checker/action-editorconfig-checker@v2
with:
version: v3.0.3
- run: editorconfig-checker

8
CODEOWNERS
View File

@ -1,5 +1,11 @@
# collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
/ci/ @ggerganov
/.devops/ @ngxson
/.devops/*.Dockerfile @ngxson
/examples/server/ @ngxson
/ggml/src/ggml-cuda/fattn* @JohannesGaessler
/ggml/src/ggml-cuda/mmq.* @JohannesGaessler
/ggml/src/ggml-cuda/mmv.* @JohannesGaessler
/ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
/ggml/src/ggml-opt.cpp @JohannesGaessler
/ggml/src/gguf.cpp @JohannesGaessler

102
CONTRIBUTING.md
View File

@ -1,10 +1,10 @@
# Pull requests (for contributors)
- Test your changes:
- Execute [the full CI locally on your machine](ci/README.md) before publishing
- Verify that the perplexity and the performance are not affected negatively by your changes (use `llama-perplexity` and `llama-bench`)
- If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends)
- If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
- Execute [the full CI locally on your machine](ci/README.md) before publishing
- Verify that the perplexity and the performance are not affected negatively by your changes (use `llama-perplexity` and `llama-bench`)
- If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends)
- If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
- Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
@ -20,14 +20,104 @@
- Avoid adding third-party dependencies, extra files, extra headers, etc.
- Always consider cross-compatibility with other operating systems and architectures
- Avoid fancy-looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
- Vertical alignment makes things more readable and easier to batch edit
- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
- Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
- Use sized integer types such as `int32_t` in the public API, e.g. `size_t` may also be appropriate for allocation sizes or byte offsets
- Declare structs with `struct foo {}` instead of `typedef struct foo {} foo`
- In C++ code omit optional `struct` and `enum` keyword whenever they are not necessary
```cpp
// OK
llama_context * ctx;
const llama_rope_type rope_type;
// not OK
struct llama_context * ctx;
const enum llama_rope_type rope_type;
```
_(NOTE: this guideline is yet to be applied to the `llama.cpp` codebase. New code should follow this guideline.)_
- Try to follow the existing patterns in the code (indentation, spaces, etc.). In case of doubt use `clang-format` to format the added code
- For anything not covered in the current guidelines, refer to the [C++ Core Guidelines](https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines)
- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
![matmul](media/matmul.png)
# Naming guidelines
- Use `snake_case` for function, variable and type names
- Naming usually optimizes for longest common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
```cpp
// not OK
int small_number;
int big_number;
// OK
int number_small;
int number_big;
```
- Enum values are always in upper case and prefixed with the enum name
```cpp
enum llama_vocab_type {
LLAMA_VOCAB_TYPE_NONE = 0,
LLAMA_VOCAB_TYPE_SPM = 1,
LLAMA_VOCAB_TYPE_BPE = 2,
LLAMA_VOCAB_TYPE_WPM = 3,
LLAMA_VOCAB_TYPE_UGM = 4,
LLAMA_VOCAB_TYPE_RWKV = 5,
};
```
- The general naming pattern is `<class>_<method>`, with `<method>` being `<action>_<noun>`
```cpp
llama_model_init(); // class: "llama_model", method: "init"
llama_sampler_chain_remove(); // class: "llama_sampler_chain", method: "remove"
llama_sampler_get_seed(); // class: "llama_sampler", method: "get_seed"
llama_set_embeddings(); // class: "llama_context", method: "set_embeddings"
llama_n_threads(); // class: "llama_context", method: "n_threads"
llama_adapter_lora_free(); // class: "llama_adapter_lora", method: "free"
```
- The `get` `<action>` can be omitted
- The `<noun>` can be omitted if not necessary
- The `_context` suffix of the `<class>` is optional. Use it to disambiguate symbols when needed
- Use `init`/`free` for constructor/destructor `<action>`
- Use the `_t` suffix when a type is supposed to be opaque to the user - it's not relevant to them if it is a struct or anything else
```cpp
typedef struct llama_context * llama_context_t;
enum llama_pooling_type llama_pooling_type(const llama_context_t ctx);
```
_(NOTE: this guideline is yet to be applied to the `llama.cpp` codebase. New code should follow this guideline)_
- C/C++ filenames are all lowercase with dashes. Headers use the `.h` extension. Source files use the `.c` or `.cpp` extension
- Python filenames are all lowercase with underscores
- _(TODO: abbreviations usage)_
# Preprocessor directives
- _(TODO: add guidelines with examples and apply them to the codebase)_
```cpp
#ifdef FOO
#endif // FOO
```
# Documentation
- Documentation is a community effort
- When you need to look into the source code to figure out how to use an API consider adding a short summary to the header file for future reference
- When you notice incorrect or outdated documentation, please update it
# Resources
The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:

3
README.md
View File

@ -69,6 +69,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
- [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557)
- [x] [Phi models](https://huggingface.co/models?search=microsoft/phi)
- [x] [PhiMoE](https://github.com/ggerganov/llama.cpp/pull/11003)
- [x] [GPT-2](https://huggingface.co/gpt2)
- [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
- [x] [InternLM2](https://huggingface.co/models?search=internlm2)
@ -98,6 +99,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
- [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
- [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
- [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
- [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
#### Multimodal
@ -201,6 +203,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
- [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
- [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
</details>

54
common/arg.cpp
View File

@ -22,6 +22,11 @@ common_arg & common_arg::set_examples(std::initializer_list<enum llama_example>
return *this;
}
common_arg & common_arg::set_excludes(std::initializer_list<enum llama_example> excludes) {
this->excludes = std::move(excludes);
return *this;
}
common_arg & common_arg::set_env(const char * env) {
help = help + "\n(env: " + env + ")";
this->env = env;
@ -37,6 +42,10 @@ bool common_arg::in_example(enum llama_example ex) {
return examples.find(ex) != examples.end();
}
bool common_arg::is_exclude(enum llama_example ex) {
return excludes.find(ex) != excludes.end();
}
bool common_arg::get_value_from_env(std::string & output) {
if (env == nullptr) return false;
char * value = std::getenv(env);
@ -121,17 +130,26 @@ std::string common_arg::to_string() {
static void common_params_handle_model_default(
std::string & model,
std::string & model_url,
const std::string & model_url,
std::string & hf_repo,
std::string & hf_file) {
std::string & hf_file,
const std::string & hf_token) {
if (!hf_repo.empty()) {
// short-hand to avoid specifying --hf-file -> default it to --model
if (hf_file.empty()) {
if (model.empty()) {
throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n");
auto auto_detected = common_get_hf_file(hf_repo, hf_token);
if (auto_detected.first.empty() || auto_detected.second.empty()) {
exit(1); // built without CURL, error message already printed
}
hf_repo = auto_detected.first;
hf_file = auto_detected.second;
} else {
hf_file = model;
}
hf_file = model;
} else if (model.empty()) {
}
// make sure model path is present (for caching purposes)
if (model.empty()) {
// this is to avoid different repo having same file name, or same file name in different subdirs
std::string filename = hf_repo + "_" + hf_file;
// to make sure we don't have any slashes in the filename
@ -281,8 +299,8 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
}
// TODO: refactor model params in a common struct
common_params_handle_model_default(params.model, params.model_url, params.hf_repo, params.hf_file);
common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file);
common_params_handle_model_default(params.model, params.model_url, params.hf_repo, params.hf_file, params.hf_token);
common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file, params.hf_token);
if (params.escape) {
string_process_escapes(params.prompt);
@ -420,7 +438,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
* - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example
*/
auto add_opt = [&](common_arg arg) {
if (arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) {
if ((arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) && !arg.is_exclude(ex)) {
ctx_arg.options.push_back(std::move(arg));
}
};
@ -649,7 +667,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params, const std::string & value) {
params.prompt = value;
}
));
).set_excludes({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--no-perf"},
string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
@ -673,7 +691,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.prompt.pop_back();
}
}
));
).set_excludes({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--in-file"}, "FNAME",
"an input file (repeat to specify multiple files)",
@ -700,7 +718,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.prompt = ss.str();
fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), value.c_str());
}
));
).set_excludes({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"-e", "--escape"},
string_format("process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false"),
@ -1512,7 +1530,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--lora"}, "FNAME",
"path to LoRA adapter (can be repeated to use multiple adapters)",
[](common_params & params, const std::string & value) {
params.lora_adapters.push_back({ std::string(value), 1.0 });
params.lora_adapters.push_back({ std::string(value), 1.0, nullptr });
}
// we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@ -1520,7 +1538,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--lora-scaled"}, "FNAME", "SCALE",
"path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters)",
[](common_params & params, const std::string & fname, const std::string & scale) {
params.lora_adapters.push_back({ fname, std::stof(scale) });
params.lora_adapters.push_back({ fname, std::stof(scale), nullptr });
}
// we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@ -1574,21 +1592,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_env("LLAMA_ARG_MODEL_URL"));
add_opt(common_arg(
{"-hfr", "--hf-repo"}, "REPO",
"Hugging Face model repository (default: unused)",
{"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
"Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
"example: unsloth/phi-4-GGUF:q4_k_m\n"
"(default: unused)",
[](common_params & params, const std::string & value) {
params.hf_repo = value;
}
).set_env("LLAMA_ARG_HF_REPO"));
add_opt(common_arg(
{"-hff", "--hf-file"}, "FILE",
"Hugging Face model file (default: unused)",
"Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
[](common_params & params, const std::string & value) {
params.hf_file = value;
}
).set_env("LLAMA_ARG_HF_FILE"));
add_opt(common_arg(
{"-hfrv", "--hf-repo-v"}, "REPO",
{"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
"Hugging Face model repository for the vocoder model (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.hf_repo = value;

3
common/arg.h
View File

@ -12,6 +12,7 @@
struct common_arg {
std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
std::set<enum llama_example> excludes = {};
std::vector<const char *> args;
const char * value_hint = nullptr; // help text or example for arg value
const char * value_hint_2 = nullptr; // for second arg value
@ -53,9 +54,11 @@ struct common_arg {
) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
common_arg & set_examples(std::initializer_list<enum llama_example> examples);
common_arg & set_excludes(std::initializer_list<enum llama_example> excludes);
common_arg & set_env(const char * env);
common_arg & set_sparam();
bool in_example(enum llama_example ex);
bool is_exclude(enum llama_example ex);
bool get_value_from_env(std::string & output);
bool has_value_from_env();
std::string to_string();

271
common/common.cpp
View File

@ -2,6 +2,9 @@
#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
#endif
#include "ggml.h"
#include "gguf.h"
#include "common.h"
#include "log.h"
// Change JSON_ASSERT from assert() to GGML_ASSERT:
@ -18,6 +21,7 @@
#include <cstdarg>
#include <cstring>
#include <ctime>
#include <filesystem>
#include <fstream>
#include <iostream>
#include <iterator>
@ -62,11 +66,29 @@
#ifdef __linux__
#include <linux/limits.h>
#elif defined(_WIN32)
#define PATH_MAX MAX_PATH
# if !defined(PATH_MAX)
# define PATH_MAX MAX_PATH
# endif
#else
#include <sys/syslimits.h>
#endif
#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
//
// CURL utils
//
using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
struct curl_slist_ptr {
struct curl_slist * ptr = nullptr;
~curl_slist_ptr() {
if (ptr) {
curl_slist_free_all(ptr);
}
}
};
#endif // LLAMA_USE_CURL
using json = nlohmann::ordered_json;
@ -843,7 +865,7 @@ struct common_init_result common_init_from_params(common_params & params) {
} else if (!params.model_url.empty()) {
model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
} else {
model = llama_load_model_from_file(params.model.c_str(), mparams);
model = llama_model_load_from_file(params.model.c_str(), mparams);
}
if (model == NULL) {
@ -851,26 +873,28 @@ struct common_init_result common_init_from_params(common_params & params) {
return iparams;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
if (params.reranking) {
bool ok = true;
if (llama_token_bos(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have a BOS token, reranking will not work\n", __func__);
if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: vocab does not have a BOS token, reranking will not work\n", __func__);
ok = false;
}
if (llama_token_eos(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__);
if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
ok = false;
}
if (llama_token_sep(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have a SEP token, reranking will not work\n", __func__);
if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
ok = false;
}
if (!ok) {
llama_free_model(model);
llama_model_free(model);
return iparams;
}
@ -878,40 +902,40 @@ struct common_init_result common_init_from_params(common_params & params) {
auto cparams = common_context_params_to_llama(params);
llama_context * lctx = llama_new_context_with_model(model, cparams);
llama_context * lctx = llama_init_from_model(model, cparams);
if (lctx == NULL) {
LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
llama_free_model(model);
llama_model_free(model);
return iparams;
}
if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
LOG_ERR("%s: KV cache shifting is not supported for this model (--no-context-shift to disable)'\n", __func__);
llama_free_model(model);
return iparams;
LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
params.ctx_shift = false;
}
if (!params.control_vectors.empty()) {
if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
if (params.control_vector_layer_end <= 0) params.control_vector_layer_end = llama_n_layer(model);
if (params.control_vector_layer_end <= 0) params.control_vector_layer_end = llama_model_n_layer(model);
const auto cvec = common_control_vector_load(params.control_vectors);
if (cvec.n_embd == -1) {
llama_free(lctx);
llama_free_model(model);
llama_model_free(model);
return iparams;
}
int err = llama_control_vector_apply(lctx,
cvec.data.data(),
cvec.data.size(),
cvec.n_embd,
params.control_vector_layer_start,
params.control_vector_layer_end);
int err = llama_apply_adapter_cvec(
lctx,
cvec.data.data(),
cvec.data.size(),
cvec.n_embd,
params.control_vector_layer_start,
params.control_vector_layer_end);
if (err) {
llama_free(lctx);
llama_free_model(model);
llama_model_free(model);
return iparams;
}
@ -919,30 +943,31 @@ struct common_init_result common_init_from_params(common_params & params) {
// load and optionally apply lora adapters
for (auto & la : params.lora_adapters) {
common_lora_adapter_container loaded_la;
loaded_la.path = la.path;
loaded_la.scale = la.scale;
loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
if (loaded_la.adapter == nullptr) {
llama_adapter_lora_ptr lora;
lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
if (lora == nullptr) {
LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
llama_free(lctx);
llama_free_model(model);
llama_model_free(model);
return iparams;
}
iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
}
if (!params.lora_init_without_apply) {
common_lora_adapters_apply(lctx, iparams.lora_adapters);
la.ptr = lora.get();
iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
}
if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
if (!params.lora_init_without_apply) {
common_set_adapter_lora(lctx, params.lora_adapters);
}
if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
params.sampling.ignore_eos = false;
}
if (params.sampling.ignore_eos) {
for (llama_token i = 0; i < llama_n_vocab(model); i++) {
if (llama_token_is_eog(model, i)) {
for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
if (llama_vocab_is_eog(vocab, i)) {
LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
params.sampling.logit_bias.push_back({i, -INFINITY});
}
@ -963,8 +988,9 @@ struct common_init_result common_init_from_params(common_params & params) {
LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
std::vector<llama_token> tmp;
llama_token bos = llama_token_bos(model);
llama_token eos = llama_token_eos(model);
llama_token bos = llama_vocab_bos(vocab);
llama_token eos = llama_vocab_eos(vocab);
// some models (e.g. T5) don't have a BOS token
if (bos != LLAMA_TOKEN_NULL) {
tmp.push_back(bos);
@ -979,7 +1005,7 @@ struct common_init_result common_init_from_params(common_params & params) {
if (llama_model_has_encoder(model)) {
llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
if (decoder_start_token_id == -1) {
if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
decoder_start_token_id = bos;
}
tmp.clear();
@ -993,17 +1019,17 @@ struct common_init_result common_init_from_params(common_params & params) {
llama_perf_context_reset(lctx);
}
iparams.model = model;
iparams.context = lctx;
iparams.model.reset(model);
iparams.context.reset(lctx);
return iparams;
}
void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters) {
llama_lora_adapter_clear(ctx);
for (auto & la : lora_adapters) {
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
llama_clear_adapter_lora(ctx);
for (auto & la : lora) {
if (la.scale != 0.0f) {
llama_lora_adapter_set(ctx, la.adapter, la.scale);
llama_set_adapter_lora(ctx, la.ptr, la.scale);
}
}
}
@ -1120,7 +1146,8 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
// Initialize libcurl
std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
if (!curl) {
LOG_ERR("%s: error initializing libcurl\n", __func__);
return false;
@ -1134,11 +1161,9 @@ static bool common_download_file(const std::string & url, const std::string & pa
// Check if hf-token or bearer-token was specified
if (!hf_token.empty()) {
std::string auth_header = "Authorization: Bearer ";
auth_header += hf_token.c_str();
struct curl_slist *http_headers = NULL;
http_headers = curl_slist_append(http_headers, auth_header.c_str());
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers);
std::string auth_header = "Authorization: Bearer " + hf_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
}
#if defined(_WIN32)
@ -1148,8 +1173,7 @@ static bool common_download_file(const std::string & url, const std::string & pa
#endif
// Check if the file already exists locally
struct stat model_file_info;
auto file_exists = (stat(path.c_str(), &model_file_info) == 0);
auto file_exists = std::filesystem::exists(path);
// If the file exists, check its JSON metadata companion file.
std::string metadata_path = path + ".json";
@ -1409,7 +1433,7 @@ struct llama_model * common_load_model_from_url(
}
}
return llama_load_model_from_file(local_path.c_str(), params);
return llama_model_load_from_file(local_path.c_str(), params);
}
struct llama_model * common_load_model_from_hf(
@ -1435,6 +1459,80 @@ struct llama_model * common_load_model_from_hf(
return common_load_model_from_url(model_url, local_path, hf_token, params);
}
/**
* Allow getting the HF file from the HF repo with tag (like ollama), for example:
* - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
* - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
* - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
* Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
*
* Return pair of <repo, file> (with "repo" already having tag removed)
*
* Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
*/
std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
auto parts = string_split<std::string>(hf_repo_with_tag, ':');
std::string tag = parts.size() > 1 ? parts.back() : "latest";
std::string hf_repo = parts[0];
if (string_split<std::string>(hf_repo, '/').size() != 2) {
throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
}
// fetch model info from Hugging Face Hub API
json model_info;
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
std::string res_str;
std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
return size * nmemb;
};
curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
#if defined(_WIN32)
curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
#endif
if (!hf_token.empty()) {
std::string auth_header = "Authorization: Bearer " + hf_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
}
// Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
CURLcode res = curl_easy_perform(curl.get());
if (res != CURLE_OK) {
throw std::runtime_error("error: cannot make GET request to HF API");
}
long res_code;
curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
if (res_code == 200) {
model_info = json::parse(res_str);
} else if (res_code == 401) {
throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
} else {
throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
}
// check response
if (!model_info.contains("ggufFile")) {
throw std::runtime_error("error: model does not have ggufFile");
}
json & gguf_file = model_info.at("ggufFile");
if (!gguf_file.contains("rfilename")) {
throw std::runtime_error("error: ggufFile does not have rfilename");
}
return std::make_pair(hf_repo, gguf_file.at("rfilename"));
}
#else
struct llama_model * common_load_model_from_url(
@ -1456,6 +1554,11 @@ struct llama_model * common_load_model_from_hf(
return nullptr;
}
std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
return std::make_pair("", "");
}
#endif // LLAMA_USE_CURL
//
@ -1554,21 +1657,23 @@ std::vector<llama_token> common_tokenize(
const std::string & text,
bool add_special,
bool parse_special) {
return common_tokenize(llama_get_model(ctx), text, add_special, parse_special);
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
return common_tokenize(vocab, text, add_special, parse_special);
}
std::vector<llama_token> common_tokenize(
const struct llama_model * model,
const struct llama_vocab * vocab,
const std::string & text,
bool add_special,
bool parse_special) {
// upper limit for the number of tokens
int n_tokens = text.length() + 2 * add_special;
std::vector<llama_token> result(n_tokens);
n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
if (n_tokens < 0) {
result.resize(-n_tokens);
int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
GGML_ASSERT(check == -n_tokens);
} else {
result.resize(n_tokens);
@ -1576,13 +1681,19 @@ std::vector<llama_token> common_tokenize(
return result;
}
static std::string _common_token_to_piece(const struct llama_model * model, llama_token token, bool special) {
std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
return common_token_to_piece(vocab, token, special);
}
std::string common_token_to_piece(const struct llama_vocab * vocab, llama_token token, bool special) {
std::string piece;
piece.resize(piece.capacity()); // using string internal cache, 15 bytes + '\n'
const int n_chars = llama_token_to_piece(model, token, &piece[0], piece.size(), 0, special);
const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
if (n_chars < 0) {
piece.resize(-n_chars);
int check = llama_token_to_piece(model, token, &piece[0], piece.size(), 0, special);
int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
GGML_ASSERT(check == -n_chars);
}
else {
@ -1592,17 +1703,19 @@ static std::string _common_token_to_piece(const struct llama_model * model, llam
return piece;
}
std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
return _common_token_to_piece(llama_get_model(ctx), token, special);
std::string common_detokenize(const struct llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
return common_detokenize(vocab, tokens, special);
}
std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
std::string common_detokenize(const struct llama_vocab * vocab, const std::vector<llama_token> & tokens, bool special) {
std::string text;
text.resize(std::max(text.capacity(), tokens.size()));
int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
int32_t n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
if (n_chars < 0) {
text.resize(-n_chars);
n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
GGML_ASSERT(n_chars <= (int32_t)text.size()); // whitespace trimming is performed after per-token detokenization
}
@ -1630,9 +1743,8 @@ bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
return false;
}
}
llama_chat_message chat[] = {{"user", "test"}};
int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
const int res = llama_chat_apply_template(tmpl.c_str(), chat, 1, true, nullptr, 0);
return res >= 0;
}
@ -1643,16 +1755,16 @@ std::string common_chat_apply_template(const struct llama_model * model,
int alloc_size = 0;
bool fallback = false; // indicate if we must fallback to default chatml
std::vector<llama_chat_message> chat;
for (auto & msg : msgs) {
for (const auto & msg : msgs) {
chat.push_back({msg.role.c_str(), msg.content.c_str()});
alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
}
const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
const char * ptr_tmpl = tmpl.empty() ? llama_model_chat_template(model, /* name */ nullptr) : tmpl.c_str();
std::vector<char> buf(alloc_size);
// run the first time to get the total output length
int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
int32_t res = llama_chat_apply_template(ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
// error: chat template is not supported
if (res < 0) {
@ -1660,18 +1772,17 @@ std::string common_chat_apply_template(const struct llama_model * model,
// if the custom "tmpl" is not supported, we throw an error
// this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
throw std::runtime_error("this custom template is not supported");
} else {
// If the built-in template is not supported, we default to chatml
res = llama_chat_apply_template(nullptr, "chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size());
fallback = true;
}
// If the built-in template is not supported, we default to chatml
res = llama_chat_apply_template("chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size());
fallback = true;
}
// if it turns out that our buffer is too small, we resize it
if ((size_t) res > buf.size()) {
buf.resize(res);
res = llama_chat_apply_template(
fallback ? nullptr : model,
fallback ? "chatml" : ptr_tmpl,
chat.data(), chat.size(), add_ass, buf.data(), buf.size());
}
@ -1724,11 +1835,13 @@ static std::string _llama_model_meta_val_str(const struct llama_model * model, c
llama_chat_templates llama_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override)
{
auto bos_token = _common_token_to_piece(model, llama_token_bos(model), true);
auto eos_token = _common_token_to_piece(model, llama_token_eos(model), true);
auto vocab = llama_model_get_vocab(model);
auto bos_token = common_token_to_piece(vocab, llama_token_bos(vocab), true);
auto eos_token = common_token_to_piece(vocab, llama_token_eos(vocab), true);
std::string default_template_src = chat_template_override;
std::string tool_use_template_src = chat_template_override;
if (chat_template_override.empty()) {
// TODO:
default_template_src = _llama_model_meta_val_str(model, "tokenizer.chat_template");
tool_use_template_src = _llama_model_meta_val_str(model, "tokenizer.chat_template.tool_use");
}

56
common/common.h
View File

@ -2,7 +2,7 @@
#pragma once
#include "llama.h"
#include "llama-cpp.h"
#include "chat-template.hpp"
#include <optional>
@ -26,13 +26,11 @@
#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
struct common_lora_adapter_info {
struct common_adapter_lora_info {
std::string path;
float scale;
};
struct common_lora_adapter_container : common_lora_adapter_info {
struct llama_lora_adapter * adapter;
struct llama_adapter_lora * ptr;
};
using llama_tokens = std::vector<llama_token>;
@ -250,8 +248,8 @@ struct common_params {
std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
std::vector<llama_model_kv_override> kv_overrides;
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
@ -459,6 +457,11 @@ static bool string_starts_with(const std::string & str,
return str.rfind(prefix, 0) == 0;
}
static bool string_ends_with(const std::string & str,
const std::string & suffix) { // While we wait for C++20's std::string::ends_with...
return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
}
bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
void string_process_escapes(std::string & input);
@ -481,10 +484,12 @@ std::string fs_get_cache_file(const std::string & filename);
// Model utils
//
// note: defines object's lifetime
struct common_init_result {
struct llama_model * model = nullptr;
struct llama_context * context = nullptr;
std::vector<common_lora_adapter_container> lora_adapters;
llama_model_ptr model;
llama_context_ptr context;
std::vector<llama_adapter_lora_ptr> lora;
};
struct common_init_result common_init_from_params(common_params & params);
@ -504,9 +509,12 @@ struct llama_model * common_load_model_from_hf(
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params);
std::pair<std::string, std::string> common_get_hf_file(
const std::string & hf_repo_with_tag,
const std::string & hf_token);
// clear LoRA adapters from context, then apply new list of adapters
void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
//
// Batch utils
@ -544,7 +552,7 @@ std::vector<llama_token> common_tokenize(
bool parse_special = false);
std::vector<llama_token> common_tokenize(
const struct llama_model * model,
const struct llama_vocab * vocab,
const std::string & text,
bool add_special,
bool parse_special = false);
@ -556,11 +564,21 @@ std::string common_token_to_piece(
llama_token token,
bool special = true);
std::string common_token_to_piece(
const struct llama_vocab * vocab,
llama_token token,
bool special = true);
// detokenizes a vector of tokens into a string
// should work similar to Python's `tokenizer.decode`
// optionally renders special/control tokens
std::string common_detokenize(
llama_context * ctx,
const struct llama_context * ctx,
const std::vector<llama_token> & tokens,
bool special = true);
std::string common_detokenize(
const struct llama_vocab * vocab,
const std::vector<llama_token> & tokens,
bool special = true);
@ -574,7 +592,7 @@ struct common_chat_msg {
std::string content;
};
// Check if the template is supported or not. Returns true if it's valid
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
// CPP wrapper for llama_chat_apply_template
@ -648,6 +666,10 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
// Split utils
//
static const char * const LLM_KV_SPLIT_NO = "split.no";
static const char * const LLM_KV_SPLIT_COUNT = "split.count";
static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
namespace {
const char * const LLM_KV_SPLIT_NO = "split.no";
const char * const LLM_KV_SPLIT_COUNT = "split.count";
const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
}

24
common/ngram-cache.cpp
View File

@ -65,13 +65,13 @@ constexpr int draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram ngram_static) {
common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
if (part_static_it == nc_static.end()) {
return -1;
return LLAMA_TOKEN_NULL;
}
const common_ngram_cache_part part_static = part_static_it->second;
int max_count_static = 0;
int sum_count_static = 0;
llama_token max_token = -1;
llama_token max_token = LLAMA_TOKEN_NULL;
for (std::pair<llama_token, int> token_count_static : part_static) {
const llama_token token = token_count_static.first;
@ -85,10 +85,10 @@ static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram
}
if (sum_count_static < draft_min_sample_size_lax[LLAMA_NGRAM_STATIC-1]) {
return -1;
return LLAMA_TOKEN_NULL;
}
if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) {
return -1;
return LLAMA_TOKEN_NULL;
}
return max_token;
}
@ -98,9 +98,9 @@ static llama_token try_draft(
common_ngram_cache & nc_primary, const std::vector<common_ngram> & ngrams_primary, common_ngram_cache_part & part_static,
const int * min_sample_size, const int * min_percent) {
llama_token drafted_token = -1;
llama_token drafted_token = LLAMA_TOKEN_NULL;
for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == LLAMA_TOKEN_NULL; --i) {
const common_ngram ngram_primary = ngrams_primary[i];
common_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
@ -112,7 +112,7 @@ static llama_token try_draft(
int max_count_primary = 0;
int max_count_static = 0;
int sum_count_primary = 0;
llama_token max_token = -1;
llama_token max_token = LLAMA_TOKEN_NULL;
for (std::pair<llama_token, int> token_count_primary : part_primary) {
const llama_token token = token_count_primary.first;
@ -154,7 +154,7 @@ void common_ngram_cache_draft(
}
while ((int) draft.size()-1 < n_draft) {
llama_token drafted_token = -1;
llama_token drafted_token = LLAMA_TOKEN_NULL;
const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
common_ngram ngram_static;
@ -177,17 +177,17 @@ void common_ngram_cache_draft(
}
ngrams_cd.push_back(ngram_cd);
}
if (drafted_token == -1) {
if (drafted_token == LLAMA_TOKEN_NULL) {
drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax);
}
if (drafted_token == -1) {
if (drafted_token == LLAMA_TOKEN_NULL) {
drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict);
}
if (drafted_token == -1) {
if (drafted_token == LLAMA_TOKEN_NULL) {
drafted_token = try_draft(nc_static, ngram_static);
}
if (drafted_token == -1) {
if (drafted_token == LLAMA_TOKEN_NULL) {
break;
}

4
common/ngram-cache.h
View File

@ -17,13 +17,13 @@ struct common_ngram {
common_ngram() {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
tokens[i] = -1;
tokens[i] = LLAMA_TOKEN_NULL;
}
}
common_ngram(const llama_token * input, const int ngram_size) {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
tokens[i] = i < ngram_size ? input[i] : -1;
tokens[i] = i < ngram_size ? input[i] : LLAMA_TOKEN_NULL;
}
}

17
common/sampling.cpp
View File

@ -113,7 +113,10 @@ struct common_sampler {
void set_logits(struct llama_context * ctx, int idx) {
const auto * logits = llama_get_logits_ith(ctx, idx);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_vocab = llama_vocab_n_tokens(vocab);
cur.resize(n_vocab);
@ -142,13 +145,15 @@ std::string common_params_sampling::print() const {
}
struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
lparams.no_perf = params.no_perf;
auto * result = new common_sampler {
/* .params = */ params,
/* .grmr = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
/* .grmr = */ llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"),
/* .chain = */ llama_sampler_chain_init(lparams),
/* .prev = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
/* .cur = */ {},
@ -157,7 +162,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
llama_sampler_chain_add(result->chain,
llama_sampler_init_logit_bias(
llama_n_vocab(model),
llama_vocab_n_tokens(vocab),
params.logit_bias.size(),
params.logit_bias.data()));
@ -172,7 +177,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
c_breakers.push_back(str.c_str());
}
llama_sampler_chain_add(result->chain, llama_sampler_init_dry (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
llama_sampler_chain_add(result->chain, llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
}
break;
case COMMON_SAMPLER_TYPE_TOP_K:
@ -194,7 +199,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break;
case COMMON_SAMPLER_TYPE_INFILL:
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (model));
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (vocab));
break;
case COMMON_SAMPLER_TYPE_PENALTIES:
llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
@ -206,7 +211,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
} else if (params.mirostat == 1) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
} else if (params.mirostat == 2) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));

33
common/speculative.cpp
View File

@ -79,10 +79,13 @@ bool common_speculative_are_compatible(
const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
const struct llama_model * model_dft = llama_get_model(ctx_dft);
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
const struct llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
const struct llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
const bool vocab_type_tgt = llama_vocab_type(vocab_tgt);
LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
const bool vocab_type_dft = llama_vocab_type(model_dft);
const bool vocab_type_dft = llama_vocab_type(vocab_dft);
LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
if (vocab_type_tgt != vocab_type_dft) {
@ -91,34 +94,34 @@ bool common_speculative_are_compatible(
return false;
}
if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
llama_token_eos(model_tgt) != llama_token_eos(model_dft)) {
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_tgt), llama_add_bos_token(model_tgt), llama_token_eos(model_tgt), llama_add_eos_token(model_tgt));
LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_dft), llama_add_bos_token(model_dft), llama_token_eos(model_dft), llama_add_eos_token(model_dft));
if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)) {
LOG_ERR("%s: draft vocab special tokens must match target vocab to use speculation\n", __func__);
LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_tgt), llama_vocab_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_tgt));
LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_dft), llama_vocab_get_add_bos(vocab_dft), llama_vocab_eos(vocab_dft), llama_vocab_get_add_eos(vocab_dft));
return false;
}
{
const int n_vocab_tgt = llama_n_vocab(model_tgt);
const int n_vocab_dft = llama_n_vocab(model_dft);
const int n_vocab_tgt = llama_vocab_n_tokens(vocab_tgt);
const int n_vocab_dft = llama_vocab_n_tokens(vocab_dft);
const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
"target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
__func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
__func__, n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
return false;
}
for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
const char * token_text_dft = llama_token_get_text(model_dft, i);
const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
LOG_ERR("%s: draft model vocab must match target model to use speculation but "
LOG_ERR("%s: draft vocab vocab must match target vocab to use speculation but "
"token %d content differs - target '%s', draft '%s'\n", __func__, i,
common_token_to_piece(ctx_tgt, i).c_str(),
common_token_to_piece(ctx_dft, i).c_str());

211
convert_hf_to_gguf.py
View File

@ -326,6 +326,7 @@ class Model:
gguf.MODEL_TENSOR.TIME_MIX_W2,
gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
gguf.MODEL_TENSOR.POSNET_NORM1,
gguf.MODEL_TENSOR.POSNET_NORM2,
)
@ -477,6 +478,11 @@ class Model:
return modelcls
return func
@classmethod
def print_registered_models(cls):
for name in sorted(cls._model_classes.keys()):
logger.error(f"- {name}")
@classmethod
def from_model_architecture(cls, arch: str) -> type[Model]:
try:
@ -687,6 +693,9 @@ class Model:
if chkhsh == "d4c8f286ea6b520b3d495c4455483cfa2302c0cfcd4be05d781b6a8a0a7cdaf1":
# ref: https://huggingface.co/Infinigence/Megrez-3B-Instruct
res = "megrez"
if chkhsh == "877081d19cf6996e2c4ff0e1236341e9b7bde288f5311a56a937f0afbbb3aeb5":
# ref: https://huggingface.co/deepseek-ai/DeepSeek-V3
res = "deepseek-v3"
if res is None:
logger.warning("\n")
@ -1764,25 +1773,19 @@ class DeciModel(Model):
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(
self.dir_model, load_merges=True,
special_token_types = ['bos', 'eos', 'eom', 'eot']
)
special_vocab._set_special_token("bos", 128000)
special_vocab._set_special_token("eos", 128001)
special_vocab._set_special_token("eom", 128008)
special_vocab._set_special_token("eot", 128009)
special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
special_vocab.add_to_gguf(self.gguf_writer)
else:
# DeciLM-7B
self._set_vocab_llama_hf()
# self._set_vocab_gpt2()
def set_gguf_parameters(self):
if "block_configs" in self.hparams: # Llama-3_1-Nemotron-51B
assert self.block_count == len(self._num_kv_heads)
assert self.block_count == len(self._num_heads)
assert self.block_count == len(self._ffn_dims)
if (rope_theta := self.hparams.get("rope_theta")) is not None:
self.gguf_writer.add_rope_freq_base(rope_theta)
self.gguf_writer.add_head_count_kv(self._num_kv_heads)
self.gguf_writer.add_head_count(self._num_heads)
self.gguf_writer.add_feed_forward_length(self._ffn_dims)
@ -2565,6 +2568,63 @@ class Phi3MiniModel(Model):
yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
@Model.register("PhiMoEForCausalLM")
class PhiMoeModel(Phi3MiniModel):
model_arch = gguf.MODEL_ARCH.PHIMOE
_experts: list[dict[str, Tensor]] | None = None
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
self.gguf_writer.add_expert_count(self.hparams["num_local_experts"])
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# process the experts separately
if name.find("block_sparse_moe.experts") != -1:
n_experts = self.hparams["num_local_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 ["w1", "w2", "w3"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename = f"model.layers.{bid}.block_sparse_moe.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}.block_sparse_moe.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 prepare_tensors(self):
super().prepare_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}")
@Model.register("PlamoForCausalLM")
class PlamoModel(Model):
model_arch = gguf.MODEL_ARCH.PLAMO
@ -3262,6 +3322,8 @@ class Rwkv6Model(Model):
# required by llama.cpp, unused
self.gguf_writer.add_head_count(0)
lerp_weights: dict[int, dict[str, Tensor]] = {}
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
new_name = self.map_tensor_name(name)
@ -3277,14 +3339,84 @@ class Rwkv6Model(Model):
if new_name.endswith("time_mix_decay.weight") or "lerp" in new_name:
data_torch = data_torch.squeeze()
rescale_every_n_layers = self.hparams["rescale_every"]
if rescale_every_n_layers > 0:
if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
try:
rescale_every_n_layers = self.hparams["rescale_every"]
if rescale_every_n_layers > 0:
if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
except KeyError:
pass
# concat time_mix_lerp weights to reduce some cpu overhead
# also reduces the number of tensors in the model
if bid is not None and "time_mix_lerp" in new_name and "time_mix_lerp_x" not in new_name:
try:
self.lerp_weights[bid][new_name] = data_torch
except KeyError:
self.lerp_weights[bid] = {new_name: data_torch}
if all(f"blk.{bid}.time_mix_lerp_{i}.weight" in self.lerp_weights[bid].keys() for i in ["w", "k", "v", "r", "g"]):
new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
data = torch.stack([self.lerp_weights[bid][f"blk.{bid}.time_mix_lerp_{i}.weight"].unsqueeze(0) for i in ["w", "k", "v", "r", "g"]], dim=0).unsqueeze(1)
yield (new_name, data)
return
yield (new_name, data_torch)
@Model.register("RWKV6Qwen2ForCausalLM")
class RWKV6Qwen2Model(Rwkv6Model):
model_arch = gguf.MODEL_ARCH.RWKV6QWEN2
def set_vocab(self):
try:
self._set_vocab_sentencepiece()
except FileNotFoundError:
self._set_vocab_gpt2()
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
num_attention_heads = self.hparams["num_attention_heads"]
num_key_value_heads = self.hparams["num_key_value_heads"]
hidden_size = self.hparams["hidden_size"]
head_size = hidden_size // num_attention_heads
rms_norm_eps = self.hparams["rms_norm_eps"]
intermediate_size = self.hparams["intermediate_size"]
time_mix_extra_dim = 64 if hidden_size >= 4096 else 32
time_decay_extra_dim = 128 if hidden_size >= 4096 else 64
# RWKV isn't context limited
self.gguf_writer.add_context_length(1048576)
self.gguf_writer.add_embedding_length(hidden_size)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_wkv_head_size(head_size)
self.gguf_writer.add_time_mix_extra_dim(time_mix_extra_dim)
self.gguf_writer.add_time_decay_extra_dim(time_decay_extra_dim)
self.gguf_writer.add_feed_forward_length(intermediate_size)
self.gguf_writer.add_file_type(self.ftype)
# special parameters for time_mixing in RWKV6QWEN2
self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
self.gguf_writer.add_token_shift_count(1)
# RWKV6QWEN2 use grouped key/value like GQA
self.gguf_writer.add_head_count_kv(num_key_value_heads)
# required by llama.cpp, unused
self.gguf_writer.add_head_count(0)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
for new_name, data in super().modify_tensors(data_torch, name, bid):
if "time_mix_w1" in new_name or "time_mix_w2" in new_name:
data = data.view(5, -1, data.shape[-1])
# rwkv6qwen2 has a different order of rkvwg instead of the original wkvrg
# permute them here to avoid code changes
data = torch.stack([data[3], data[1], data[2], data[0], data[4]], dim=0).view(-1, data.shape[-1])
if "w2" in new_name:
data = data.view(5, -1, data.shape[-1])
yield (new_name, data)
continue
yield (new_name, data)
@Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
class MambaModel(Model):
model_arch = gguf.MODEL_ARCH.MAMBA
@ -3379,6 +3511,24 @@ class CommandR2Model(Model):
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
@Model.register("Cohere2ForCausalLM")
class Cohere2Model(Model):
model_arch = gguf.MODEL_ARCH.COHERE2
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_logit_scale(self.hparams["logit_scale"])
self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
rotary_pct = self.hparams["rotary_pct"]
hidden_size = self.hparams["hidden_size"]
num_attention_heads = self.hparams["num_attention_heads"]
self.gguf_writer.add_rope_dimension_count(int(rotary_pct * (hidden_size // num_attention_heads)))
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
@Model.register("OlmoForCausalLM")
@Model.register("OLMoForCausalLM")
class OlmoModel(Model):
@ -3837,6 +3987,7 @@ class DeepseekModel(Model):
@Model.register("DeepseekV2ForCausalLM")
@Model.register("DeepseekV3ForCausalLM")
class DeepseekV2Model(Model):
model_arch = gguf.MODEL_ARCH.DEEPSEEK2
@ -3858,6 +4009,15 @@ class DeepseekV2Model(Model):
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_expert_weights_norm(hparams["norm_topk_prob"])
if hparams["scoring_func"] == "sigmoid":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
elif hparams["scoring_func"] == "softmax":
self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
else:
raise ValueError(f"Unsupported scoring_func value: {hparams['scoring_func']}")
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"]:
@ -3870,6 +4030,16 @@ class DeepseekV2Model(Model):
_experts: list[dict[str, Tensor]] | None = None
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# rename e_score_correction_bias tensors
if name.endswith("e_score_correction_bias"):
name = name.replace("e_score_correction_bias", "e_score_correction.bias")
# skip Multi-Token Prediction (MTP) layers
block_count = self.hparams["num_hidden_layers"]
match = re.match(r"model.layers.(\d+)", name)
if match and int(match.group(1)) >= block_count:
return []
# process the experts separately
if name.find("mlp.experts") != -1:
n_experts = self.hparams["n_routed_experts"]
@ -4764,6 +4934,7 @@ def parse_args() -> argparse.Namespace:
parser.add_argument(
"model", type=Path,
help="directory containing model file",
nargs="?",
)
parser.add_argument(
"--use-temp-file", action="store_true",
@ -4801,8 +4972,15 @@ def parse_args() -> argparse.Namespace:
"--metadata", type=Path,
help="Specify the path for an authorship metadata override file"
)
parser.add_argument(
"--print-supported-models", action="store_true",
help="Print the supported models"
)
return parser.parse_args()
args = parser.parse_args()
if not args.print_supported_models and args.model is None:
parser.error("the following arguments are required: model")
return args
def split_str_to_n_bytes(split_str: str) -> int:
@ -4826,6 +5004,11 @@ def split_str_to_n_bytes(split_str: str) -> int:
def main() -> None:
args = parse_args()
if args.print_supported_models:
logger.error("Supported models:")
Model.print_registered_models()
sys.exit(0)
if args.verbose:
logging.basicConfig(level=logging.DEBUG)
else:

1
convert_hf_to_gguf_update.py
View File

@ -107,6 +107,7 @@ models = [
{"name": "roberta-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sentence-transformers/stsb-roberta-base"},
{"name": "gigachat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct"},
{"name": "megrez", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Infinigence/Megrez-3B-Instruct"},
{"name": "deepseek-v3", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
]

34
convert_lora_to_gguf.py
View File

@ -226,6 +226,9 @@ def get_base_tensor_name(lora_tensor_name: str) -> str:
base_name = lora_tensor_name.replace("base_model.model.", "")
base_name = base_name.replace(".lora_A.weight", ".weight")
base_name = base_name.replace(".lora_B.weight", ".weight")
# models produced by mergekit-extract-lora have token embeddings in the adapter
base_name = base_name.replace(".lora_embedding_A", ".weight")
base_name = base_name.replace(".lora_embedding_B", ".weight")
return base_name
@ -260,6 +263,10 @@ def parse_args() -> argparse.Namespace:
"--base", type=Path,
help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config",
)
parser.add_argument(
"--base-model-id", type=str,
help="the model ID of the base model, if it is not available locally or in the adapter config. If specified, it will ignore --base and load the base model config from the Hugging Face hub (Example: 'meta-llama/Llama-3.2-1B-Instruct')",
)
parser.add_argument(
"lora_path", type=Path,
help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
@ -290,6 +297,7 @@ if __name__ == '__main__':
dir_base_model: Path | None = args.base
dir_lora: Path = args.lora_path
base_model_id: str | None = args.base_model_id
lora_config = dir_lora / "adapter_config.json"
input_model = dir_lora / "adapter_model.safetensors"
@ -313,7 +321,10 @@ if __name__ == '__main__':
lparams: dict[str, Any] = json.load(f)
# load base model
if dir_base_model is None:
if base_model_id is not None:
logger.info(f"Loading base model from Hugging Face: {base_model_id}")
hparams = load_hparams_from_hf(base_model_id)
elif dir_base_model is None:
if "base_model_name_or_path" in lparams:
model_id = lparams["base_model_name_or_path"]
logger.info(f"Loading base model from Hugging Face: {model_id}")
@ -371,11 +382,16 @@ if __name__ == '__main__':
if self.lazy:
tensor = LazyTorchTensor.from_eager(tensor)
base_name = get_base_tensor_name(name)
is_lora_a = ".lora_A.weight" in name
is_lora_b = ".lora_B.weight" in name
# note: mergekit-extract-lora also adds token embeddings to the adapter
is_lora_a = ".lora_A.weight" in name or ".lora_embedding_A" in name
is_lora_b = ".lora_B.weight" in name or ".lora_embedding_B" in name
if not is_lora_a and not is_lora_b:
if ".base_layer.weight" in name:
continue
# mergekit-extract-lora add these layernorm to the adapter, we need to keep them
if "_layernorm" in name or ".norm" in name:
yield (base_name, tensor)
continue
logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
@ -407,9 +423,21 @@ if __name__ == '__main__':
if name == "lm_head.weight" and len(dest) == 0:
raise ValueError("lm_head is present in adapter, but is ignored in base model")
for dest_name, dest_data in dest:
# mergekit-extract-lora add these layernorm to the adapter
if "_norm" in dest_name:
assert dest_data.dim() == 1
yield (dest_name, dest_data)
continue
# otherwise, we must get the lora_A and lora_B tensors
assert isinstance(dest_data, LoraTorchTensor)
lora_a, lora_b = dest_data.get_lora_A_B()
# note: mergekit-extract-lora flip and transpose A and B
# here we only need to transpose token_embd.lora_a, see llm_build_inp_embd()
if "token_embd.weight" in dest_name:
lora_a = lora_a.T
yield (dest_name + ".lora_a", lora_a)
yield (dest_name + ".lora_b", lora_b)

2
docs/build.md
View File

@ -127,6 +127,8 @@ For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
This provides GPU acceleration using an NVIDIA GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from the [NVIDIA developer site](https://developer.nvidia.com/cuda-downloads).
If you are using Fedora (using Fedora Workstation, or an 'Atomic' variant such as Silverblue), or would like to set up CUDA in a toolbox, please consider our [Fedora CUDA guide](./cuda-fedora.md). Unfortunately, the process is not as simple as one might expect.
- Using `CMake`:
```bash

317
docs/cuda-fedora.md Normal file
View File

@ -0,0 +1,317 @@
# Setting Up CUDA on Fedora
In this guide we setup [Nvidia CUDA](https://docs.nvidia.com/cuda/) in a toolbox container. This guide is applicable for:
- [Fedora Workstation](https://fedoraproject.org/workstation/)
- [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/)
- [Fedora Spins](https://fedoraproject.org/spins)
- [Other Distributions](https://containertoolbx.org/distros/), including `Red Hat Enterprise Linux >= 8.`, `Arch Linux`, and `Ubuntu`.
## Table of Contents
- [Prerequisites](#prerequisites)
- [Monitoring NVIDIA CUDA Repositories](#monitoring-nvidia-cuda-repositories)
- [Using the Fedora 39 CUDA Repository](#using-the-fedora-39-cuda-repository)
- [Creating a Fedora Toolbox Environment](#creating-a-fedora-toolbox-environment)
- [Installing Essential Development Tools](#installing-essential-development-tools)
- [Adding the CUDA Repository](#adding-the-cuda-repository)
- [Installing `nvidia-driver-libs`](#installing-nvidia-driver-libs)
- [Manually Resolving Package Conflicts](#manually-resolving-package-conflicts)
- [Finalizing the Installation of `nvidia-driver-libs`](#finalizing-the-installation-of-nvidia-driver-libs)
- [Installing the CUDA Meta-Package](#installing-the-cuda-meta-package)
- [Configuring the Environment](#configuring-the-environment)
- [Verifying the Installation](#verifying-the-installation)
- [Conclusion](#conclusion)
- [Troubleshooting](#troubleshooting)
- [Additional Notes](#additional-notes)
- [References](#references)
## Prerequisites
- **Toolbox Installed on the Host System** `Fedora Silverblue` and `Fedora Workstation` both have toolbox by default, other distributions may need to install the [toolbox package](https://containertoolbx.org/install/).
- **NVIDIA Drivers and Graphics Card installed on Host System (optional)** To run CUDA program, such as `llama.cpp`, the host should be setup to access your NVIDIA hardware. Fedora Hosts can use the [RPM Fusion Repository](https://rpmfusion.org/Howto/NVIDIA).
- **Internet connectivity** to download packages.
### Monitoring NVIDIA CUDA Repositories
Before proceeding, it is advisable to check if NVIDIA has updated their CUDA repositories for your Fedora version. NVIDIA's repositories can be found at:
- [Fedora 40 CUDA Repository](https://developer.download.nvidia.com/compute/cuda/repos/fedora40/x86_64/)
- [Fedora 41 CUDA Repository](https://developer.download.nvidia.com/compute/cuda/repos/fedora41/x86_64/)
As of the latest update, these repositories do not contain the `cuda` meta-package or are missing essential components.
### Using the Fedora 39 CUDA Repository
Since the newer repositories are incomplete, we'll use the Fedora 39 repository:
- [Fedora 39 CUDA Repository](https://developer.download.nvidia.com/compute/cuda/repos/fedora39/x86_64/)
**Note:** Fedora 39 is no longer maintained, so we recommend using a toolbox environment to prevent system conflicts.
## Creating a Fedora Toolbox Environment
This guide focuses on Fedora hosts, but with small adjustments, it can work for other hosts. Using a Fedora 39 toolbox allows us to install the necessary packages without affecting the host system.
**Note:** Toolbox is available for other systems, and even without Toolbox, it is possible to use Podman or Docker.
We do not recommend installing on the host system, as Fedora 39 is out-of-maintenance, and instead you should upgrade to a maintained version of Fedora for your host.
1. **Create a Fedora 39 Toolbox:**
```bash
toolbox create --image registry.fedoraproject.org/fedora-toolbox:39 --container fedora-toolbox-39-cuda
```
2. **Enter the Toolbox:**
```bash
toolbox enter --container fedora-toolbox-39-cuda
```
Inside the toolbox, you have root privileges and can install packages without affecting the host system.
## Installing Essential Development Tools
1. **Synchronize the DNF Package Manager:**
```bash
sudo dnf distro-sync
```
2. **Install the Default Text Editor (Optional):**
```bash
sudo dnf install vim-default-editor --allowerasing
```
The `--allowerasing` flag resolves any package conflicts.
3. **Install Development Tools and Libraries:**
```bash
sudo dnf install @c-development @development-tools cmake
```
This installs essential packages for compiling software, including `gcc`, `make`, and other development headers.
## Adding the CUDA Repository
Add the NVIDIA CUDA repository to your DNF configuration:
```bash
sudo dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/fedora39/x86_64/cuda-fedora39.repo
```
After adding the repository, synchronize the package manager again:
```bash
sudo dnf distro-sync
```
## Installing `nvidia-driver-libs`
Attempt to install `nvidia-driver-libs`:
```bash
sudo dnf install nvidia-driver-libs
```
**Explanation:**
- `nvidia-driver-libs` contains necessary NVIDIA driver libraries required by CUDA.
- This step might fail due to conflicts with existing NVIDIA drivers on the host system.
## Manually Resolving Package Conflicts
If the installation fails due to conflicts, we'll manually download and install the required packages, excluding conflicting files.
### 1. Download the `nvidia-driver-libs` RPM
```bash
sudo dnf download --arch x86_64 nvidia-driver-libs
```
You should see a file similar to:
```
nvidia-driver-libs-560.35.05-1.fc39.x86_64.rpm
```
### 2. Attempt to Install the RPM
```bash
sudo dnf install nvidia-driver-libs-560.35.05-1.fc39.x86_64.rpm
```
**Expected Error:**
Installation may fail with errors pointing to conflicts with `egl-gbm` and `egl-wayland`.
**Note: It is important to carefully read the error messages to identify the exact paths that need to be excluded.**
### 3. Download Dependencies
```bash
sudo dnf download --arch x86_64 egl-gbm egl-wayland
```
### 4. Install `egl-gbm` with Excluded Paths
Exclude conflicting files during installation:
```bash
sudo rpm --install --verbose --hash \
--excludepath=/usr/lib64/libnvidia-egl-gbm.so.1.1.2 \
--excludepath=/usr/share/egl/egl_external_platform.d/15_nvidia_gbm.json \
egl-gbm-1.1.2^20240919gitb24587d-3.fc39.x86_64.rpm
```
**Explanation:**
- The `--excludepath` option skips installing files that conflict with existing files.
- Adjust the paths based on the error messages you receive.
### 5. Install `egl-wayland` with Excluded Paths
```bash
sudo rpm --install --verbose --hash \
--excludepath=/usr/share/egl/egl_external_platform.d/10_nvidia_wayland.json \
egl-wayland-1.1.17^20241118giteeb29e1-5.fc39.x86_64.rpm
```
### 6. Install `nvidia-driver-libs` with Excluded Paths
```bash
sudo rpm --install --verbose --hash \
--excludepath=/usr/share/glvnd/egl_vendor.d/10_nvidia.json \
--excludepath=/usr/share/nvidia/nvoptix.bin \
nvidia-driver-libs-560.35.05-1.fc39.x86_64.rpm
```
**Note:**
- Replace the paths with the ones causing conflicts in your installation if they differ.
- The `--verbose` and `--hash` options provide detailed output during installation.
## Finalizing the Installation of `nvidia-driver-libs`
After manually installing the dependencies, run:
```bash
sudo dnf install nvidia-driver-libs
```
You should receive a message indicating the package is already installed:
```
Package nvidia-driver-libs-3:560.35.05-1.fc39.x86_64 is already installed.
Dependencies resolved.
Nothing to do.
Complete!
```
## Installing the CUDA Meta-Package
Now that the driver libraries are installed, proceed to install CUDA:
```bash
sudo dnf install cuda
```
This installs the CUDA toolkit and associated packages.
## Configuring the Environment
To use CUDA, add its binary directory to your system's `PATH`.
1. **Create a Profile Script:**
```bash
sudo sh -c 'echo "export PATH=\$PATH:/usr/local/cuda/bin" >> /etc/profile.d/cuda.sh'
```
**Explanation:**
- We add to `/etc/profile.d/` as the `/etc/` folder is unique to this particular container, and is not shared with other containers or the host system.
- The backslash `\` before `$PATH` ensures the variable is correctly written into the script.
2. **Make the Script Executable:**
```bash
sudo chmod +x /etc/profile.d/cuda.sh
```
3. **Source the Script to Update Your Environment:**
```bash
source /etc/profile.d/cuda.sh
```
**Note:** This command updates your current shell session with the new `PATH`. The `/etc/profile.d/cuda.sh` script ensures that the CUDA binaries are available in your `PATH` for all future sessions.
## Verifying the Installation
To confirm that CUDA is correctly installed and configured, check the version of the NVIDIA CUDA Compiler (`nvcc`):
```bash
nvcc --version
```
You should see output similar to:
```
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2024 NVIDIA Corporation
Built on Tue_Oct_29_23:50:19_PDT_2024
Cuda compilation tools, release 12.6, V12.6.85
Build cuda_12.6.r12.6/compiler.35059454_0
```
This output confirms that the CUDA compiler is accessible and indicates the installed version.
## Conclusion
You have successfully set up CUDA on Fedora within a toolbox environment using the Fedora 39 CUDA repository. By manually resolving package conflicts and configuring the environment, you can develop CUDA applications without affecting your host system.
## Troubleshooting
- **Installation Failures:**
- If you encounter errors during installation, carefully read the error messages. They often indicate conflicting files or missing dependencies.
- Use the `--excludepath` option with `rpm` to exclude conflicting files during manual installations.
- **Driver Conflicts:**
- Since the host system may already have NVIDIA drivers installed, conflicts can arise. Using the toolbox environment helps isolate these issues.
- **Environment Variables Not Set:**
- If `nvcc` is not found after installation, ensure that `/usr/local/cuda/bin` is in your `PATH`.
- Run `echo $PATH` to check if the path is included.
- Re-source the profile script or open a new terminal session.
## Additional Notes
- **Updating CUDA in the Future:**
- Keep an eye on the official NVIDIA repositories for updates to your Fedora version.
- When an updated repository becomes available, adjust your `dnf` configuration accordingly.
- **Building `llama.cpp`:**
- With CUDA installed, you can follow these [build instructions for `llama.cpp`](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) to compile it with CUDA support.
- Ensure that any CUDA-specific build flags or paths are correctly set in your build configuration.
- **Using the Toolbox Environment:**
- The toolbox environment is isolated from your host system, which helps prevent conflicts.
- Remember that system files and configurations inside the toolbox are separate from the host. By default the home directory of the user is shared between the host and the toolbox.
---
**Disclaimer:** Manually installing and modifying system packages can lead to instability of the container. The above steps are provided as a guideline and may need adjustments based on your specific system configuration. Always back up important data before making significant system changes, especially as your home folder is writable and shared with he toolbox.
**Acknowledgments:** Special thanks to the Fedora community and NVIDIA documentation for providing resources that assisted in creating this guide.
## References
- [Fedora Toolbox Documentation](https://docs.fedoraproject.org/en-US/fedora-silverblue/toolbox/)
- [NVIDIA CUDA Installation Guide](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html)
- [Podman Documentation](https://podman.io/get-started)
---

10
docs/development/HOWTO-add-model.md
View File

@ -28,7 +28,7 @@ The required steps to implement for an HF model are:
```python
@Model.register("MyModelForCausalLM")
class MyModel(Model):
model_arch = gguf.MODEL_ARCH.GROK
model_arch = gguf.MODEL_ARCH.MYMODEL
```
2. Define the layout of the GGUF tensors in [constants.py](/gguf-py/gguf/constants.py)
@ -79,14 +79,14 @@ Depending on the model configuration, tokenizer, code and tensors layout, you wi
- `Model#set_vocab`
- `Model#write_tensors`
NOTE: Tensor names must end with `.weight` suffix, that is the convention and several tools like `quantize` expect this to proceed the weights.
NOTE: Tensor names must end with `.weight` or `.bias` suffixes, that is the convention and several tools like `quantize` expect this to proceed the weights.
### 2. Define the model architecture in `llama.cpp`
The model params and tensors layout must be defined in `llama.cpp`:
1. Define a new `llm_arch`
2. Define the tensors layout in `LLM_TENSOR_NAMES`
3. Add any non standard metadata in `llm_load_hparams`
3. Add any non-standard metadata in `llm_load_hparams`
4. Create the tensors for inference in `llm_load_tensors`
5. If the model has a RoPE operation, add the rope type in `llama_rope_type`
@ -96,9 +96,9 @@ NOTE: The dimensions in `ggml` are typically in the reverse order of the `pytorc
This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `llama_build_graph`.
Have a look at existing implementation like `build_llama`, `build_dbrx` or `build_bert`.
Have a look at existing implementations like `build_llama`, `build_dbrx` or `build_bert`.
When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support for missing backend operations can be added in another PR.
Some `ggml` backends do not support all operations. Backend implementations can be added in a separate PR.
Note: to debug the inference graph: you can use [llama-eval-callback](/examples/eval-callback/).

6
examples/batched-bench/batched-bench.cpp
View File

@ -38,7 +38,7 @@ int main(int argc, char ** argv) {
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
@ -50,7 +50,7 @@ int main(int argc, char ** argv) {
// ensure enough sequences are available
ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
@ -194,7 +194,7 @@ int main(int argc, char ** argv) {
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
llama_backend_free();

6
examples/batched.swift/Sources/main.swift
View File

@ -23,12 +23,12 @@ defer {
}
let model_params = llama_model_default_params()
guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), model_params) else {
guard let model = llama_model_load_from_file(modelPath.cString(using: .utf8), model_params) else {
print("Failed to load model")
exit(1)
}
defer {
llama_free_model(model)
llama_model_free(model)
}
var tokens = tokenize(text: prompt, add_bos: true)
@ -141,7 +141,7 @@ while n_cur <= n_len {
let new_token_id = llama_sampler_sample(smpl, context, i_batch[i])
// is it an end of stream? -> mark the stream as finished
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
if llama_vocab_is_eog(model, new_token_id) || n_cur == n_len {
i_batch[i] = -1
// print("")
if n_parallel > 1 {

16
examples/batched/batched.cpp
View File

@ -41,17 +41,19 @@ int main(int argc, char ** argv) {
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: error: unable to load model\n" , __func__);
return 1;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
// tokenize the prompt
std::vector<llama_token> tokens_list;
tokens_list = common_tokenize(model, params.prompt, true);
tokens_list = common_tokenize(vocab, params.prompt, true);
const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
@ -62,7 +64,7 @@ int main(int argc, char ** argv) {
ctx_params.n_ctx = n_kv_req;
ctx_params.n_batch = std::max(n_predict, n_parallel);
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
auto sparams = llama_sampler_chain_default_params();
sparams.no_perf = false;
@ -120,8 +122,8 @@ int main(int argc, char ** argv) {
}
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
if (decoder_start_token_id == -1) {
decoder_start_token_id = llama_token_bos(model);
if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
decoder_start_token_id = llama_vocab_bos(vocab);
}
common_batch_clear(batch);
@ -174,7 +176,7 @@ int main(int argc, char ** argv) {
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
// is it an end of generation? -> mark the stream as finished
if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_predict) {
i_batch[i] = -1;
LOG("\n");
if (n_parallel > 1) {
@ -236,7 +238,7 @@ int main(int argc, char ** argv) {
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
llama_backend_free();

18
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View File

@ -1,4 +1,6 @@
#include "ggml.h"
#include "gguf.h"
#include "llama.h"
#include "common.h"
#include "log.h"
@ -434,12 +436,12 @@ static void print_matrix(struct ggml_tensor * probs) {
}
}
struct llama_file {
struct my_llama_file {
// use FILE * so we don't have to re-open the file to mmap
FILE * fp;
size_t size;
llama_file(const char * fname, const char * mode) {
my_llama_file(const char * fname, const char * mode) {
fp = std::fopen(fname, mode);
if (fp == NULL) {
size = 0;
@ -500,7 +502,7 @@ struct llama_file {
return std::string(chars.data(), len);
}
~llama_file() {
~my_llama_file() {
if (fp) {
std::fclose(fp);
}
@ -508,7 +510,7 @@ struct llama_file {
};
static bool is_ggml_file(const char * filename) {
llama_file file(filename, "rb");
my_llama_file file(filename, "rb");
if (file.size < 4) {
return false;
}
@ -576,7 +578,7 @@ static void load_vocab(const char * filename, const Config * config, struct my_l
} else {
// assume llama2.c vocabulary
LOG_INF("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
llama_file file(filename, "rb");
my_llama_file file(filename, "rb");
if (!file.fp) {
die_fmt("%s: %s", strerror(errno), filename);
}
@ -689,8 +691,8 @@ static void save_as_llama_model(
gguf_set_val_u32(ctx, KV_TOKENIZER_UNK_ID, UNKNOWN_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_BOS_ID, BOS_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_EOS_ID, EOS_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1);
gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, LLAMA_TOKEN_NULL);
gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, LLAMA_TOKEN_NULL);
gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx);
gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
@ -909,7 +911,7 @@ int main(int argc, char ** argv) {
load_vocab(params.fn_vocab_model, &config, &vocab);
struct my_llama_model model;
model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
model.hparams.n_vocab = config.vocab_size; //llama_vocab_n_vocab(lctx);
model.hparams.n_ctx = params.n_ctx;
model.hparams.n_embd = config.dim; //params.n_embd;
model.hparams.n_ff = config.hidden_dim;

19
examples/cvector-generator/cvector-generator.cpp
View File

@ -1,7 +1,9 @@
#include "ggml.h"
#include "gguf.h"
#include "arg.h"
#include "common.h"
#include "llama.h"
#include "ggml.h"
#include "pca.hpp"
#include "mean.hpp"
@ -271,7 +273,9 @@ struct tokenized_prompt {
size_t max_seq_len;
tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const bool add_bos = llama_vocab_get_add_bos(vocab);
tokens_pos = common_tokenize(ctx, pos, add_bos, true);
tokens_neg = common_tokenize(ctx, neg, add_bos, true);
max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
@ -415,12 +419,13 @@ int main(int argc, char ** argv) {
// load the model to get hparams
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
// int n_ctx = llama_n_ctx(ctx);
int n_layers = llama_n_layer(model);
int n_embd = llama_n_embd(model);
int n_layers = llama_model_n_layer(model);
int n_embd = llama_model_n_embd(model);
// get model hint param (a.k.a model arch name)
char model_hint[128];
llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
@ -474,8 +479,6 @@ int main(int argc, char ** argv) {
// done with the model, we can now free it to make gain some memory
printf("Done evaluate prompts, unload model...\n");
llama_free(ctx);
llama_free_model(model);
bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;

15
examples/embedding/embedding.cpp
View File

@ -97,14 +97,17 @@ int main(int argc, char ** argv) {
// load the model
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
if (model == NULL) {
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
@ -147,7 +150,7 @@ int main(int argc, char ** argv) {
// check if the last token is SEP
// it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
for (auto & inp : inputs) {
if (inp.empty() || inp.back() != llama_token_sep(model)) {
if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
}
@ -180,7 +183,7 @@ int main(int argc, char ** argv) {
}
// allocate output
const int n_embd = llama_n_embd(model);
const int n_embd = llama_model_n_embd(model);
std::vector<float> embeddings(n_embd_count * n_embd, 0);
float * emb = embeddings.data();
@ -316,8 +319,6 @@ int main(int argc, char ** argv) {
// clean up
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;

13
examples/eval-callback/eval-callback.cpp
View File

@ -127,7 +127,10 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
}
static bool run(llama_context * ctx, const common_params & params) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const bool add_bos = llama_vocab_get_add_bos(vocab);
std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
@ -162,8 +165,9 @@ int main(int argc, char ** argv) {
// init
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
if (model == nullptr || ctx == nullptr) {
LOG_ERR("%s : failed to init\n", __func__);
return 1;
@ -184,9 +188,6 @@ int main(int argc, char ** argv) {
LOG("\n");
llama_perf_context_print(ctx);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;

9
examples/export-lora/export-lora.cpp
View File

@ -1,12 +1,13 @@
#include "arg.h"
#include "common.h"
#include "ggml.h"
#include "ggml-alloc.h"
#include "gguf.h"
#include "arg.h"
#include "common.h"
#include <map>
#include <vector>
#include <string>
#include <thread>
#include <fstream>
static bool g_verbose = false;
@ -128,7 +129,7 @@ struct lora_merge_ctx {
lora_merge_ctx(
std::string & base_fname,
std::vector<common_lora_adapter_info> & lora_files,
std::vector<common_adapter_lora_info> & lora_files,
std::string & outfile,
int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
fout.exceptions(std::ofstream::failbit); // fail fast on write errors

1
examples/gguf-hash/gguf-hash.cpp
View File

@ -1,4 +1,5 @@
#include "ggml.h"
#include "gguf.h"
#include <cstdlib> /* abort() */
#include <cstddef>

15
examples/gguf-split/gguf-split.cpp
View File

@ -1,18 +1,19 @@
#include "ggml.h"
#include "gguf.h"
#include "llama.h"
#include "common.h"
#include <algorithm>
#include <cmath>
#include <cinttypes>
#include <climits>
#include <cstdio>
#include <cstdlib>
#include <stdexcept>
#include <cstring>
#include <fstream>
#include <string>
#include <vector>
#include <stdio.h>
#include <string.h>
#include <climits>
#include <stdexcept>
#if defined(_WIN32)
#include <windows.h>
#ifndef PATH_MAX
@ -297,7 +298,7 @@ struct split_strategy {
total_size += ggml_nbytes(t);
}
total_size = total_size / 1000 / 1000; // convert to megabytes
printf("split %05d: n_tensors = %d, total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
printf("split %05d: n_tensors = %" PRIi64 ", total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
i_split++;
}
}

16
examples/gguf/gguf.cpp
View File

@ -1,10 +1,9 @@
#include "ggml.h"
#include "gguf.h"
#include <cstdio>
#include <cinttypes>
#include <string>
#include <sstream>
#include <fstream>
#include <vector>
#undef MIN
@ -135,9 +134,10 @@ static bool gguf_ex_read_0(const std::string & fname) {
for (int i = 0; i < n_tensors; ++i) {
const char * name = gguf_get_tensor_name (ctx, i);
const size_t size = gguf_get_tensor_size (ctx, i);
const size_t offset = gguf_get_tensor_offset(ctx, i);
printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
}
}
@ -182,9 +182,10 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
for (int i = 0; i < n_tensors; ++i) {
const char * name = gguf_get_tensor_name (ctx, i);
const size_t size = gguf_get_tensor_size (ctx, i);
const size_t offset = gguf_get_tensor_offset(ctx, i);
printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
}
}
@ -199,7 +200,8 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
printf("%s: tensor[%d]: n_dims = %d, ne = (%d, %d, %d, %d), name = %s, data = %p\n",
__func__, i, ggml_n_dims(cur), int(cur->ne[0]), int(cur->ne[1]), int(cur->ne[2]), int(cur->ne[3]), cur->name, cur->data);
// print first 10 elements
const float * data = (const float *) cur->data;
@ -215,7 +217,7 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
const float * data = (const float *) cur->data;
for (int j = 0; j < ggml_nelements(cur); ++j) {
if (data[j] != 100 + i) {
fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]);
fprintf(stderr, "%s: tensor[%d], data[%d]: found %f, expected %f\n", __func__, i, j, data[j], float(100 + i));
gguf_free(ctx);
return false;
}
@ -245,6 +247,8 @@ int main(int argc, char ** argv) {
check_data = false;
}
srand(123456);
const std::string fname(argv[1]);
const std::string mode (argv[2]);

23
examples/gritlm/gritlm.cpp
View File

@ -11,6 +11,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
std::vector<std::vector<float>> result;
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
@ -19,16 +20,16 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
const std::string input_string = instruction + sentences[i];
std::vector<llama_token> inputs = common_tokenize(model, input_string, true, false);
std::vector<llama_token> inputs = common_tokenize(vocab, input_string, true, false);
const int32_t n_toks = inputs.size();
// GritLM seems to have EOS = ""
// https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
// inputs.push_back(llama_token_eos(model));
// inputs.push_back(llama_vocab_eos(vocab));
// we want to ignore instruction tokens for mean pooling
const int32_t n_inst = common_tokenize(model, instruction, true, false).size();
const int32_t n_inst = common_tokenize(vocab, instruction, true, false).size();
#ifdef GRIT_DEBUG
// debug tokens - should be matching as referenced in the GritLM sample
@ -52,7 +53,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
llama_decode(ctx, batch);
// get embedding dimensions
uint64_t n_embd = llama_n_embd(model);
uint64_t n_embd = llama_model_n_embd(model);
// allocate embedding output
std::vector<float> emb_unorm(n_embd, 0.0f);
@ -97,7 +98,9 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
std::string result;
const llama_model * model = llama_get_model(ctx);
llama_token eos_token = llama_token_eos(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_token eos_token = llama_vocab_eos(vocab);
llama_kv_cache_clear(ctx);
llama_set_embeddings(ctx, false);
@ -105,7 +108,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
std::vector<llama_token> inputs = common_tokenize(model, prompt, false, true);
std::vector<llama_token> inputs = common_tokenize(vocab, prompt, false, true);
int32_t i_current_token = 0;
while (true) {
@ -165,10 +168,10 @@ int main(int argc, char * argv[]) {
llama_backend_init();
llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
// create generation context
llama_context * ctx = llama_new_context_with_model(model, cparams);
llama_context * ctx = llama_init_from_model(model, cparams);
auto sparams = llama_sampler_chain_default_params();
@ -197,7 +200,7 @@ int main(int argc, char * argv[]) {
const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
const std::vector<std::vector<float>> q_rep = encode(ctx, queries, gritlm_instruction(instruction));
const int n_embd = llama_n_embd(model);
const int n_embd = llama_model_n_embd(model);
const float cosine_sim_q0_d0 = common_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q0_d1 = common_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
@ -219,7 +222,7 @@ int main(int argc, char * argv[]) {
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
llama_backend_free();
return 0;

25
examples/imatrix/imatrix.cpp
View File

@ -7,7 +7,6 @@
#include <cstdio>
#include <cstring>
#include <ctime>
#include <sstream>
#include <thread>
#include <mutex>
#include <vector>
@ -40,7 +39,7 @@ public:
void set_params(common_params params) { m_params = std::move(params); }
bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
void save_imatrix(int ncall = -1) const;
bool load_imatrix(const char * file_name);
bool load_imatrix(const char * fname);
private:
std::unordered_map<std::string, Stats> m_stats;
common_params m_params;
@ -429,10 +428,14 @@ static void process_logits(
}
static bool compute_imatrix(llama_context * ctx, const common_params & params) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const bool add_bos = llama_vocab_get_add_bos(vocab);
const int n_ctx = llama_n_ctx(ctx);
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
auto tim1 = std::chrono::high_resolution_clock::now();
LOG_INF("%s: tokenizing the input ..\n", __func__);
@ -467,7 +470,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
const int n_chunk_max = tokens.size() / n_ctx;
const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_vocab = llama_vocab_n_tokens(vocab);
const int n_batch = params.n_batch;
int count = 0;
@ -507,7 +510,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
// add BOS token for the first batch of each chunk
if (add_bos && j == 0) {
tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
tokens[batch_start] = llama_vocab_bos(vocab);
}
common_batch_clear(batch);
@ -618,14 +621,15 @@ int main(int argc, char ** argv) {
// init
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
if (model == nullptr || ctx == nullptr) {
LOG_ERR("%s : failed to init\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
if (params.n_ctx > n_ctx_train) {
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, params.n_ctx);
@ -655,9 +659,6 @@ int main(int argc, char ** argv) {
LOG("\n");
llama_perf_context_print(ctx);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;

47
examples/infill/infill.cpp
View File

@ -131,15 +131,17 @@ int main(int argc, char ** argv) {
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
common_init_result llama_init = common_init_from_params(params);
model = llama_init.model;
ctx = llama_init.context;
model = llama_init.model.get();
ctx = llama_init.context.get();
if (model == NULL) {
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
LOG_DBG("n_ctx: %d\n", n_ctx);
@ -152,28 +154,28 @@ int main(int argc, char ** argv) {
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
}
const bool add_bos = llama_add_bos_token(model);
GGML_ASSERT(!llama_add_eos_token(model));
const bool add_bos = llama_vocab_get_add_bos(vocab);
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
std::vector<llama_token> embd_inp;
std::vector<llama_token> embd_end;
std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
GGML_ASSERT(llama_token_fim_pre(model) >= 0);
GGML_ASSERT(llama_token_fim_suf(model) >= 0);
GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0);
GGML_ASSERT(llama_vocab_fim_suf(vocab) >= 0);
inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
embd_end = params.spm_infill ? inp_pfx : inp_sfx;
if (add_bos) {
embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
}
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
const llama_token middle_token = llama_token_fim_mid(model);
const llama_token middle_token = llama_vocab_fim_mid(vocab);
if (middle_token >= 0) {
embd_inp.push_back(middle_token);
}
@ -185,7 +187,7 @@ int main(int argc, char ** argv) {
// Should not run without any tokens
if (embd_inp.empty()) {
embd_inp.push_back(llama_token_bos(model));
embd_inp.push_back(llama_vocab_bos(vocab));
LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
}
@ -420,10 +422,10 @@ int main(int argc, char ** argv) {
// if not currently processing queued inputs;
if ((int) embd_inp.size() <= n_consumed) {
// deal with eot token in infill mode
if ((common_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
if ((common_sampler_last(smpl) == llama_vocab_eot(vocab) || is_interacting) && params.interactive){
if (is_interacting && !params.interactive_first) {
// print an eot token
LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
}
LOG("\n");
console::set_display(console::user_input);
@ -463,13 +465,13 @@ int main(int argc, char ** argv) {
std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
embd_end = params.spm_infill ? inp_pfx : inp_sfx;
if (add_bos) {
embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
}
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
@ -484,7 +486,7 @@ int main(int argc, char ** argv) {
is_interacting = false;
}
// deal with end of generation tokens in interactive mode
else if (llama_token_is_eog(model, common_sampler_last(smpl))) {
else if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
LOG_DBG("found EOS token\n");
if (params.interactive) {
@ -500,7 +502,7 @@ int main(int argc, char ** argv) {
if (params.input_prefix_bos) {
LOG_DBG("adding input prefix BOS token\n");
embd_inp.push_back(llama_token_bos(model));
embd_inp.push_back(llama_vocab_bos(vocab));
}
std::string buffer;
@ -563,7 +565,7 @@ int main(int argc, char ** argv) {
}
// end of generation
if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !params.interactive) {
if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !params.interactive) {
break;
}
@ -575,15 +577,12 @@ int main(int argc, char ** argv) {
}
}
if (!params.interactive && n_remain <= 0) {
LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
}
LOG("\n");
common_perf_print(ctx, smpl);
llama_free(ctx);
llama_free_model(model);
common_sampler_free(smpl);
llama_backend_free();

20
examples/llama-bench/llama-bench.cpp
View File

@ -1401,7 +1401,8 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th
llama_set_n_threads(ctx, n_threads, n_threads);
const llama_model * model = llama_get_model(ctx);
const int32_t n_vocab = llama_n_vocab(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int32_t n_vocab = llama_vocab_n_tokens(vocab);
std::vector<llama_token> tokens(n_batch);
@ -1409,7 +1410,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th
while (n_processed < n_prompt) {
int n_tokens = std::min(n_prompt - n_processed, n_batch);
tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
tokens[0] = n_processed == 0 && llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab;
for (int i = 1; i < n_tokens; i++) {
tokens[i] = std::rand() % n_vocab;
}
@ -1424,9 +1425,10 @@ static void test_gen(llama_context * ctx, int n_gen, int n_threads) {
llama_set_n_threads(ctx, n_threads, n_threads);
const llama_model * model = llama_get_model(ctx);
const int32_t n_vocab = llama_n_vocab(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int32_t n_vocab = llama_vocab_n_tokens(vocab);
llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
llama_token token = llama_vocab_get_add_bos(vocab) ? llama_vocab_bos(vocab) : std::rand() % n_vocab;
for (int i = 0; i < n_gen; i++) {
llama_decode(ctx, llama_batch_get_one(&token, 1));
@ -1526,10 +1528,10 @@ int main(int argc, char ** argv) {
// keep the same model between tests when possible
if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
if (lmodel) {
llama_free_model(lmodel);
llama_model_free(lmodel);
}
lmodel = llama_load_model_from_file(inst.model.c_str(), inst.to_llama_mparams());
lmodel = llama_model_load_from_file(inst.model.c_str(), inst.to_llama_mparams());
if (lmodel == NULL) {
fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
return 1;
@ -1537,10 +1539,10 @@ int main(int argc, char ** argv) {
prev_inst = &inst;
}
llama_context * ctx = llama_new_context_with_model(lmodel, inst.to_llama_cparams());
llama_context * ctx = llama_init_from_model(lmodel, inst.to_llama_cparams());
if (ctx == NULL) {
fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, inst.model.c_str());
llama_free_model(lmodel);
llama_model_free(lmodel);
return 1;
}
@ -1626,7 +1628,7 @@ int main(int argc, char ** argv) {
ggml_threadpool_free_fn(threadpool);
}
llama_free_model(lmodel);
llama_model_free(lmodel);
if (p) {
p->print_footer();

11
examples/llama.android/llama/src/main/cpp/llama-android.cpp
View File

@ -87,7 +87,7 @@ Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring fi
auto path_to_model = env->GetStringUTFChars(filename, 0);
LOGi("Loading model from %s", path_to_model);
auto model = llama_load_model_from_file(path_to_model, model_params);
auto model = llama_model_load_from_file(path_to_model, model_params);
env->ReleaseStringUTFChars(filename, path_to_model);
if (!model) {
@ -102,7 +102,7 @@ Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring fi
extern "C"
JNIEXPORT void JNICALL
Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
llama_free_model(reinterpret_cast<llama_model *>(model));
llama_model_free(reinterpret_cast<llama_model *>(model));
}
extern "C"
@ -305,7 +305,9 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
extern "C"
JNIEXPORT void JNICALL
Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
//llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
delete batch;
}
extern "C"
@ -403,6 +405,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
const auto sampler = reinterpret_cast<llama_sampler *>(sampler_pointer);
const auto model = llama_get_model(context);
const auto vocab = llama_model_get_vocab(model);
if (!la_int_var) la_int_var = env->GetObjectClass(intvar_ncur);
if (!la_int_var_value) la_int_var_value = env->GetMethodID(la_int_var, "getValue", "()I");
@ -412,7 +415,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
const auto new_token_id = llama_sampler_sample(sampler, context, -1);
const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_len) {
return nullptr;
}

6
examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
View File

@ -52,8 +52,8 @@ actor LlamaContext {
deinit {
llama_sampler_free(sampling)
llama_batch_free(batch)
llama_model_free(model)
llama_free(context)
llama_free_model(model)
llama_backend_free()
}
@ -65,7 +65,7 @@ actor LlamaContext {
model_params.n_gpu_layers = 0
print("Running on simulator, force use n_gpu_layers = 0")
#endif
let model = llama_load_model_from_file(path, model_params)
let model = llama_model_load_from_file(path, model_params)
guard let model else {
print("Could not load model at \(path)")
throw LlamaError.couldNotInitializeContext
@ -151,7 +151,7 @@ actor LlamaContext {
new_token_id = llama_sampler_sample(sampling, context, batch.n_tokens - 1)
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
if llama_vocab_is_eog(model, new_token_id) || n_cur == n_len {
print("\n")
is_done = true
let new_token_str = String(cString: temporary_invalid_cchars + [0])

6
examples/llava/clip.cpp
View File

@ -7,6 +7,7 @@
#include "ggml-cpu.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#include "gguf.h"
//#ifdef GGML_USE_CUDA
//#include "ggml-cuda.h"
@ -262,7 +263,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
{
const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
int arr_n = gguf_get_arr_n(ctx_gguf, i);
const void * data = gguf_get_arr_data(ctx_gguf, i);
const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
std::stringstream ss;
ss << "[";
for (int j = 0; j < arr_n; j++) {
@ -2734,7 +2735,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
total_size_org += orig_size;
total_size_new += new_size;
gguf_set_tensor_type(ctx_out, name.c_str(), new_type);
gguf_set_tensor_data(ctx_out, name.c_str(), new_data, new_size);
GGML_ASSERT(gguf_get_tensor_size(ctx_out, gguf_find_tensor(ctx_out, name.c_str())) == new_size);
gguf_set_tensor_data(ctx_out, name.c_str(), new_data);
fout.write((const char *)new_data, new_size);
size_t pad = GGML_PAD(new_size, gguf_get_alignment(ctx_out)) - new_size;
for (size_t j = 0; j < pad; ++j) {

15
examples/llava/llava-cli.cpp
View File

@ -47,8 +47,12 @@ static const char * sample(struct common_sampler * smpl,
int * n_past) {
const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
common_sampler_accept(smpl, id, true);
const llama_model * model = llama_get_model(ctx_llama);
const llama_vocab * vocab = llama_model_get_vocab(model);
static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
if (llama_vocab_is_eog(vocab, id)) {
ret = "</s>";
} else {
ret = common_token_to_piece(ctx_llama, id);
@ -221,7 +225,7 @@ static struct llama_model * llava_init(common_params * params) {
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
@ -239,11 +243,10 @@ static struct llava_context * llava_init_context(common_params * params, llama_m
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
llama_context_params ctx_params = common_context_params_to_llama(*params);
ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
@ -265,7 +268,7 @@ static void llava_free(struct llava_context * ctx_llava) {
}
llama_free(ctx_llava->ctx_llama);
llama_free_model(ctx_llava->model);
llama_model_free(ctx_llava->model);
llama_backend_free();
}
@ -323,7 +326,7 @@ int main(int argc, char ** argv) {
}
}
llama_free_model(model);
llama_model_free(model);
return 0;
}

4
examples/llava/llava.cpp
View File

@ -384,7 +384,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
// make sure that the correct mmproj was used, i.e., compare apples to apples
int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
int n_llama_embd = llama_model_n_embd(llama_get_model(ctx_llama));
auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
if (n_image_embd != n_llama_embd) {
LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
@ -456,7 +456,7 @@ struct llava_embd_batch {
};
bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
int n_embd = llama_n_embd(llama_get_model(ctx_llama));
int n_embd = llama_model_n_embd(llama_get_model(ctx_llama));
for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
int n_eval = image_embed->n_image_pos - i;

12
examples/llava/minicpmv-cli.cpp
View File

@ -31,7 +31,7 @@ static struct llama_model * llava_init(common_params * params) {
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
@ -54,7 +54,7 @@ static struct llava_context * llava_init_context(common_params * params, llama_m
ctx_params.n_ctx = params->n_ctx;
}
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
@ -75,7 +75,7 @@ static void llava_free(struct llava_context * ctx_llava) {
}
llama_free(ctx_llava->ctx_llama);
llama_free_model(ctx_llava->model);
llama_model_free(ctx_llava->model);
llama_backend_free();
}
@ -167,8 +167,12 @@ static const char * sample(struct common_sampler * smpl,
int * n_past) {
const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
common_sampler_accept(smpl, id, true);
const llama_model * model = llama_get_model(ctx_llama);
const llama_vocab * vocab = llama_model_get_vocab(model);
static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
if (llama_vocab_is_eog(vocab, id)) {
ret = "</s>";
} else {
ret = common_token_to_piece(ctx_llama, id);

19
examples/llava/qwen2vl-cli.cpp
View File

@ -27,7 +27,7 @@
static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) {
int n_embd = llama_n_embd(llama_get_model(ctx_llama));
int n_embd = llama_model_n_embd(llama_get_model(ctx_llama));
const int patch_size = 14 * 2;
const int ph = image_size->height / patch_size + (image_size->height % patch_size > 0);
const int pw = image_size->width / patch_size + (image_size->width % patch_size > 0);
@ -132,8 +132,12 @@ static const char * sample(struct common_sampler * smpl,
int * n_past, int * st_pos_id) {
const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
common_sampler_accept(smpl, id, true);
const llama_model * model = llama_get_model(ctx_llama);
const llama_vocab * vocab = llama_model_get_vocab(model);
static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
if (llama_vocab_is_eog(vocab, id)) {
ret = "</s>";
} else {
ret = common_token_to_piece(ctx_llama, id);
@ -310,7 +314,7 @@ static struct llama_model * llava_init(common_params * params) {
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
@ -328,11 +332,10 @@ static struct llava_context * llava_init_context(common_params * params, llama_m
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
llama_context_params ctx_params = common_context_params_to_llama(*params);
ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
@ -354,7 +357,7 @@ static void llava_free(struct llava_context * ctx_llava) {
}
llama_free(ctx_llava->ctx_llama);
llama_free_model(ctx_llava->model);
llama_model_free(ctx_llava->model);
llama_backend_free();
}
@ -481,7 +484,7 @@ static void debug_test_mrope_2d() {
}
static void debug_dump_img_embed(struct llava_context * ctx_llava) {
int n_embd = llama_n_embd(llama_get_model(ctx_llava->ctx_llama));
int n_embd = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
int ne = n_embd * 4;
float vals[56 * 56 * 3];
// float embd[ne];
@ -575,7 +578,7 @@ int main(int argc, char ** argv) {
}
}
llama_free_model(model);
llama_model_free(model);
return 0;
}

13
examples/lookahead/lookahead.cpp
View File

@ -58,8 +58,10 @@ int main(int argc, char ** argv) {
// load the target model
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
const llama_vocab * vocab = llama_model_get_vocab(model);
// Tokenize the prompt
std::vector<llama_token> inp;
@ -147,7 +149,7 @@ int main(int argc, char ** argv) {
}
// here we keep adding new n-grams as we go
ngram_container ngrams_observed(llama_n_vocab(model), N, G);
ngram_container ngrams_observed(llama_vocab_n_tokens(vocab), N, G);
// debug
struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, W + G + 1);
@ -297,7 +299,7 @@ int main(int argc, char ** argv) {
}
fflush(stdout);
if (llama_token_is_eog(model, id)) {
if (llama_vocab_is_eog(vocab, id)) {
has_eos = true;
}
@ -474,9 +476,6 @@ int main(int argc, char ** argv) {
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
LOG("\n\n");

13
examples/lookup/lookup-create.cpp
View File

@ -1,14 +1,9 @@
#include "arg.h"
#include "common.h"
#include "ngram-cache.h"
#include "ggml.h"
#include "llama.h"
#include <cstdint>
#include <fstream>
#include <iostream>
#include <string>
#include <unordered_map>
#include <vector>
int main(int argc, char ** argv){
@ -25,16 +20,16 @@ int main(int argc, char ** argv){
// load the model
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model_ptr & model = llama_init.model;
llama_context_ptr & ctx = llama_init.context;
GGML_ASSERT(model != nullptr);
// tokenize the prompt
std::vector<llama_token> inp;
inp = common_tokenize(ctx, params.prompt, true, true);
inp = common_tokenize(ctx.get(), params.prompt, true, true);
fprintf(stderr, "%s: tokenization done\n", __func__);
common_ngram_cache ngram_cache;
common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());

10
examples/lookup/lookup-stats.cpp
View File

@ -30,12 +30,11 @@ int main(int argc, char ** argv){
// load the model
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_context_ptr & ctx = llama_init.context;
// tokenize the prompt
std::vector<llama_token> inp;
inp = common_tokenize(ctx, params.prompt, true, true);
inp = common_tokenize(ctx.get(), params.prompt, true, true);
common_ngram_cache ngram_cache_context;
common_ngram_cache ngram_cache_dynamic;
@ -66,7 +65,7 @@ int main(int argc, char ** argv){
}
const int n_input = inp.size();
const int n_ctx = llama_n_ctx(ctx);
const int n_ctx = llama_n_ctx(ctx.get());
int n_drafted = 0;
int n_accept = 0;
@ -150,9 +149,6 @@ int main(int argc, char ** argv){
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
LOG("\n\n");

11
examples/lookup/lookup.cpp
View File

@ -33,8 +33,10 @@ int main(int argc, char ** argv){
// load the model
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
const llama_vocab * vocab = llama_model_get_vocab(model);
// tokenize the prompt
std::vector<llama_token> inp;
@ -136,7 +138,7 @@ int main(int argc, char ** argv){
LOG("%s", token_str.c_str());
}
if (llama_token_is_eog(model, id)) {
if (llama_vocab_is_eog(vocab, id)) {
has_eos = true;
}
@ -243,9 +245,6 @@ int main(int argc, char ** argv){
llama_batch_free(batch_tgt);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
LOG("\n\n");

36
examples/main/main.cpp
View File

@ -5,7 +5,6 @@
#include "sampling.h"
#include "llama.h"
#include <cassert>
#include <cstdio>
#include <cstring>
#include <ctime>
@ -145,24 +144,26 @@ int main(int argc, char ** argv) {
llama_context * ctx = nullptr;
common_sampler * smpl = nullptr;
std::vector<common_chat_msg> chat_msgs;
g_model = &model;
g_ctx = &ctx;
g_smpl = &smpl;
std::vector<common_chat_msg> chat_msgs;
// load the model and apply lora adapter, if any
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
common_init_result llama_init = common_init_from_params(params);
model = llama_init.model;
ctx = llama_init.context;
model = llama_init.model.get();
ctx = llama_init.context.get();
if (model == NULL) {
LOG_ERR("%s: error: unable to load model\n", __func__);
return 1;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
@ -196,7 +197,7 @@ int main(int argc, char ** argv) {
llama_attach_threadpool(ctx, threadpool, threadpool_batch);
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
if (n_ctx > n_ctx_train) {
@ -241,9 +242,9 @@ int main(int argc, char ** argv) {
}
}
const bool add_bos = llama_add_bos_token(model);
const bool add_bos = llama_vocab_get_add_bos(vocab);
if (!llama_model_has_encoder(model)) {
GGML_ASSERT(!llama_add_eos_token(model));
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
}
LOG_DBG("n_ctx: %d, add_bos: %d\n", n_ctx, add_bos);
@ -269,7 +270,7 @@ int main(int argc, char ** argv) {
// Should not run without any tokens
if (embd_inp.empty()) {
if (add_bos) {
embd_inp.push_back(llama_token_bos(model));
embd_inp.push_back(llama_vocab_bos(vocab));
LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
} else {
LOG_ERR("input is empty\n");
@ -494,8 +495,8 @@ int main(int argc, char ** argv) {
}
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
if (decoder_start_token_id == -1) {
decoder_start_token_id = llama_token_bos(model);
if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
decoder_start_token_id = llama_vocab_bos(vocab);
}
embd_inp.clear();
@ -742,7 +743,7 @@ int main(int argc, char ** argv) {
}
// deal with end of generation tokens in interactive mode
if (llama_token_is_eog(model, common_sampler_last(smpl))) {
if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
LOG_DBG("found an EOG token\n");
if (params.interactive) {
@ -776,7 +777,7 @@ int main(int argc, char ** argv) {
if (params.input_prefix_bos) {
LOG_DBG("adding input prefix BOS token\n");
embd_inp.push_back(llama_token_bos(model));
embd_inp.push_back(llama_vocab_bos(vocab));
}
std::string buffer;
@ -830,8 +831,8 @@ int main(int argc, char ** argv) {
// if user stop generation mid-way, we must add EOT to finish model's last response
if (need_insert_eot && format_chat) {
llama_token eot = llama_token_eot(model);
embd_inp.push_back(eot == -1 ? llama_token_eos(model) : eot);
llama_token eot = llama_vocab_eot(vocab);
embd_inp.push_back(eot == LLAMA_TOKEN_NULL ? llama_vocab_eos(vocab) : eot);
need_insert_eot = false;
}
@ -866,7 +867,7 @@ int main(int argc, char ** argv) {
}
// end of generation
if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.interactive)) {
if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !(params.interactive)) {
LOG(" [end of text]\n");
break;
}
@ -889,9 +890,6 @@ int main(int argc, char ** argv) {
common_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
ggml_threadpool_free_fn(threadpool);

11
examples/parallel/parallel.cpp
View File

@ -132,8 +132,10 @@ int main(int argc, char ** argv) {
// load the target model
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
const llama_vocab * vocab = llama_model_get_vocab(model);
// load the prompts from an external file if there are any
if (params.prompt.empty()) {
@ -358,7 +360,7 @@ int main(int argc, char ** argv) {
// client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
if (client.n_decoded > 2 &&
(llama_token_is_eog(model, id) ||
(llama_vocab_is_eog(vocab, id) ||
(params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
client.response.find("User:") != std::string::npos ||
client.response.find('\n') != std::string::npos)) {
@ -416,9 +418,6 @@ int main(int argc, char ** argv) {
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
LOG("\n\n");

12
examples/passkey/passkey.cpp
View File

@ -63,22 +63,24 @@ int main(int argc, char ** argv) {
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return 1;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
// initialize the context
llama_context_params ctx_params = common_context_params_to_llama(params);
ctx_params.n_ctx = llama_n_ctx_train(model)*n_grp + n_keep;
ctx_params.n_ctx = llama_model_n_ctx_train(model)*n_grp + n_keep;
GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
if (ctx == NULL) {
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
return 1;
@ -223,7 +225,7 @@ int main(int argc, char ** argv) {
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_len) {
LOG("\n");
break;
@ -266,7 +268,7 @@ int main(int argc, char ** argv) {
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
llama_backend_free();

60
examples/perplexity/perplexity.cpp
View File

@ -296,8 +296,11 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const bool add_bos = llama_vocab_get_add_bos(vocab);
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
LOG_INF("%s: tokenizing the input ..\n", __func__);
@ -338,7 +341,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
const int n_batch = params.n_batch;
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_vocab = llama_vocab_n_tokens(vocab);
int count = 0;
double nll = 0.0;
@ -382,7 +385,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
// add BOS token for the first batch of each chunk
if (add_bos && j == 0) {
tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
tokens[batch_start] = llama_vocab_bos(vocab);
}
const auto * batch_logits = llama_get_logits(ctx);
@ -444,8 +447,11 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const bool add_bos = llama_vocab_get_add_bos(vocab);
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
std::ofstream logits_stream;
if (!params.logits_file.empty()) {
@ -485,7 +491,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
const int n_batch = params.n_batch;
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_vocab = llama_vocab_n_tokens(vocab);
int count = 0;
double nll = 0.0;
@ -557,7 +563,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
// add BOS token for the first batch of each chunk
if (add_bos && j == 0) {
tokens[seq_start] = llama_token_bos(llama_get_model(ctx));
tokens[seq_start] = llama_vocab_bos(vocab);
}
for (int k = 0; k < batch_size; ++k) {
@ -732,6 +738,9 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto
}
static void hellaswag_score(llama_context * ctx, const common_params & params) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
// Calculates hellaswag score (acc_norm) from prompt
//
// Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
@ -765,7 +774,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
size_t hs_task_count = prompt_lines.size()/6;
LOG_INF("%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
const bool is_spm = llama_vocab_type(vocab) == LLAMA_VOCAB_TYPE_SPM;
LOG_INF("================================= is_spm = %d\n", is_spm);
// The tasks should be randomized so the score stabilizes quickly.
@ -848,7 +857,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
const int n_ctx = llama_n_ctx(ctx);
const int n_batch = params.n_batch;
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_vocab = llama_vocab_n_tokens(vocab);
const int max_tasks_per_batch = 32;
const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
@ -1072,6 +1081,8 @@ static std::vector<winogrande_entry> load_winogrande_from_csv(const std::string
*
*/
static void winogrande_score(llama_context * ctx, const common_params & params) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
constexpr int k_min_trailing_ctx = 3;
@ -1130,7 +1141,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
const int n_ctx = llama_n_ctx(ctx);
const int n_batch = params.n_batch;
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_vocab = llama_vocab_n_tokens(vocab);
const int max_tasks_per_batch = 128;
const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
@ -1374,6 +1385,8 @@ static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choic
// https://huggingface.co/datasets/truthful_qa
//
static void multiple_choice_score(llama_context * ctx, const common_params & params) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
std::istringstream strstream(params.prompt);
uint32_t n_task;
@ -1482,7 +1495,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
const int n_ctx = llama_n_ctx(ctx);
const int n_batch = params.n_batch;
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_vocab = llama_vocab_n_tokens(vocab);
const int max_tasks_per_batch = 32;
const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
@ -1655,6 +1668,9 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
}
static void kl_divergence(llama_context * ctx, const common_params & params) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
if (params.logits_file.empty()) {
LOG_ERR("%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__);
return;
@ -1688,8 +1704,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
LOG_ERR("%s: failed reading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str());
return;
}
if (n_vocab != llama_n_vocab(llama_get_model(ctx))) {
LOG_ERR("%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx)));
if (n_vocab != llama_vocab_n_tokens(vocab)) {
LOG_ERR("%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_vocab_n_tokens(vocab));
}
std::vector<llama_token> tokens(size_t(n_ctx) * n_chunk);
@ -1701,8 +1717,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
const int n_batch = params.n_batch;
const int num_batches = (n_ctx + n_batch - 1)/n_batch;
const int nv = 2*((n_vocab + 1)/2) + 4;
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
const bool add_bos = llama_vocab_get_add_bos(vocab);
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
std::vector<float> kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
@ -1761,7 +1777,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
// add BOS token for the first batch of each chunk
if (add_bos && j == 0) {
tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
tokens[batch_start] = llama_vocab_bos(vocab);
}
common_batch_clear(batch);
@ -1987,14 +2003,15 @@ int main(int argc, char ** argv) {
// load the model and apply lora adapter, if any
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
if (model == NULL) {
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
if (params.n_ctx > n_ctx_train) {
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
@ -2023,9 +2040,6 @@ int main(int argc, char ** argv) {
LOG("\n");
llama_perf_context_print(ctx);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;

26
examples/quantize-stats/quantize-stats.cpp
View File

@ -1,7 +1,7 @@
#include "common.h"
#include "ggml.h"
#include "llama.h"
#include "llama-impl.h"
#include "llama-context.h"
#include "common.h"
#include <algorithm>
#include <cassert>
@ -9,11 +9,9 @@
#include <cmath>
#include <cstdio>
#include <cstring>
#include <map>
#include <numeric>
#include <regex>
#include <string>
#include <unordered_map>
#include <vector>
#include <thread>
#include <mutex>
@ -311,7 +309,7 @@ int main(int argc, char ** argv) {
auto mparams = llama_model_default_params();
mparams.use_mlock = false;
model = llama_load_model_from_file(params.model.c_str(), mparams);
model = llama_model_load_from_file(params.model.c_str(), mparams);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
@ -321,22 +319,22 @@ int main(int argc, char ** argv) {
auto cparams = llama_context_default_params();
cparams.n_ctx = 256;
ctx = llama_new_context_with_model(model, cparams);
ctx = llama_init_from_model(model, cparams);
if (ctx == NULL) {
fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
llama_free_model(model);
llama_model_free(model);
return 1;
}
}
const auto &tensors = llama_internal_get_tensor_map(ctx);
const auto & tensors = llama_internal_get_tensor_map(ctx);
// check layer tensors
int included_layers = 0;
int64_t max_nelements = 0;
bool is_f16 = false;
for (const auto& kv_tensor : tensors) {
for (const auto & kv_tensor : tensors) {
if (!layer_included(params, kv_tensor.first)) {
continue;
}
@ -349,7 +347,7 @@ int main(int argc, char ** argv) {
fprintf(stderr, "%s: error: Quantization should be tested with a float model, "
"this model contains already quantized layers (%s is type %d)\n", __func__, kv_tensor.first.c_str(), kv_tensor.second->type);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
return 1;
}
included_layers++;
@ -371,8 +369,8 @@ int main(int argc, char ** argv) {
if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
continue;
}
const auto * qfns = ggml_get_type_traits(type);
const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
const auto * qfns = ggml_get_type_traits(type);
const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
if (qfns_cpu->from_float && qfns->to_float) {
if (params.verbose) {
printf("testing %s ...\n", ggml_type_name(type));
@ -382,7 +380,7 @@ int main(int argc, char ** argv) {
error_stats global_stats {};
for (const auto& kv_tensor : tensors) {
for (const auto & kv_tensor : tensors) {
if (!layer_included(params, kv_tensor.first)) {
continue;
}
@ -411,7 +409,7 @@ int main(int argc, char ** argv) {
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
// report timing
{
const int64_t t_main_end_us = ggml_time_us();

16
examples/retrieval/retrieval.cpp
View File

@ -151,15 +151,17 @@ int main(int argc, char ** argv) {
// load the model
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
if (model == NULL) {
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
@ -192,8 +194,8 @@ int main(int argc, char ** argv) {
return 1;
}
// add eos if not present
if (llama_token_eos(model) >= 0 && (inp.empty() || inp.back() != llama_token_eos(model))) {
inp.push_back(llama_token_eos(model));
if (llama_vocab_eos(vocab) >= 0 && (inp.empty() || inp.back() != llama_vocab_eos(vocab))) {
inp.push_back(llama_vocab_eos(vocab));
}
chunk.tokens = inp;
}
@ -215,7 +217,7 @@ int main(int argc, char ** argv) {
struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
// allocate output
const int n_embd = llama_n_embd(model);
const int n_embd = llama_model_n_embd(model);
std::vector<float> embeddings(n_chunks * n_embd, 0);
float * emb = embeddings.data();
@ -298,7 +300,5 @@ int main(int argc, char ** argv) {
// clean up
llama_batch_free(query_batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
}

99
examples/run/run.cpp
View File

@ -1,5 +1,6 @@
#if defined(_WIN32)
# include <windows.h>
# include <io.h>
#else
# include <sys/file.h>
# include <sys/ioctl.h>
@ -10,6 +11,8 @@
# include <curl/curl.h>
#endif
#include <signal.h>
#include <climits>
#include <cstdarg>
#include <cstdio>
@ -24,6 +27,13 @@
#include "json.hpp"
#include "llama-cpp.h"
#if defined(__unix__) || (defined(__APPLE__) && defined(__MACH__)) || defined(_WIN32)
[[noreturn]] static void sigint_handler(int) {
printf("\n\033[0m");
exit(0); // not ideal, but it's the only way to guarantee exit in all cases
}
#endif
GGML_ATTRIBUTE_FORMAT(1, 2)
static std::string fmt(const char * fmt, ...) {
va_list ap;
@ -82,6 +92,7 @@ class Opt {
}
ctx_params.n_batch = context_size >= 0 ? context_size : context_size_default;
ctx_params.n_ctx = ctx_params.n_batch;
model_params.n_gpu_layers = ngl >= 0 ? ngl : ngl_default;
temperature = temperature >= 0 ? temperature : temperature_default;
@ -253,7 +264,7 @@ class File {
return 1;
}
OVERLAPPED overlapped = { 0 };
OVERLAPPED overlapped = {};
if (!LockFileEx(hFile, LOCKFILE_EXCLUSIVE_LOCK | LOCKFILE_FAIL_IMMEDIATELY, 0, MAXDWORD, MAXDWORD,
&overlapped)) {
fd = -1;
@ -277,7 +288,7 @@ class File {
if (fd >= 0) {
# ifdef _WIN32
if (hFile != INVALID_HANDLE_VALUE) {
OVERLAPPED overlapped = { 0 };
OVERLAPPED overlapped = {};
UnlockFileEx(hFile, 0, MAXDWORD, MAXDWORD, &overlapped);
}
# else
@ -293,7 +304,7 @@ class File {
private:
int fd = -1;
# ifdef _WIN32
HANDLE hFile;
HANDLE hFile = nullptr;
# endif
};
@ -464,7 +475,7 @@ class HttpClient {
return (now_downloaded_plus_file_size * 100) / total_to_download;
}
static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", percentage); }
static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", static_cast<long int>(percentage)); }
static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) {
const auto now = std::chrono::steady_clock::now();
@ -663,7 +674,7 @@ class LlamaData {
"\r%*s"
"\rLoading model",
get_terminal_width(), " ");
llama_model_ptr model(llama_load_model_from_file(opt.model_.c_str(), opt.model_params));
llama_model_ptr model(llama_model_load_from_file(opt.model_.c_str(), opt.model_params));
if (!model) {
printe("%s: error: unable to load model from file: %s\n", __func__, opt.model_.c_str());
}
@ -674,7 +685,7 @@ class LlamaData {
// Initializes the context with the specified parameters
llama_context_ptr initialize_context(const llama_model_ptr & model, const Opt & opt) {
llama_context_ptr context(llama_new_context_with_model(model.get(), opt.ctx_params));
llama_context_ptr context(llama_init_from_model(model.get(), opt.ctx_params));
if (!context) {
printe("%s: error: failed to create the llama_context\n", __func__);
}
@ -702,11 +713,11 @@ static void add_message(const char * role, const std::string & text, LlamaData &
// Function to apply the chat template and resize `formatted` if needed
static int apply_chat_template(LlamaData & llama_data, const bool append) {
int result = llama_chat_apply_template(
llama_data.model.get(), nullptr, llama_data.messages.data(), llama_data.messages.size(), append,
llama_model_chat_template(llama_data.model.get()), llama_data.messages.data(), llama_data.messages.size(), append,
append ? llama_data.fmtted.data() : nullptr, append ? llama_data.fmtted.size() : 0);
if (append && result > static_cast<int>(llama_data.fmtted.size())) {
llama_data.fmtted.resize(result);
result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
result = llama_chat_apply_template(llama_model_chat_template(llama_data.model.get()), llama_data.messages.data(),
llama_data.messages.size(), append, llama_data.fmtted.data(),
llama_data.fmtted.size());
}
@ -715,11 +726,11 @@ static int apply_chat_template(LlamaData & llama_data, const bool append) {
}
// Function to tokenize the prompt
static int tokenize_prompt(const llama_model_ptr & model, const std::string & prompt,
static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
std::vector<llama_token> & prompt_tokens) {
const int n_prompt_tokens = -llama_tokenize(model.get(), prompt.c_str(), prompt.size(), NULL, 0, true, true);
const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
prompt_tokens.resize(n_prompt_tokens);
if (llama_tokenize(model.get(), prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true,
if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true,
true) < 0) {
printe("failed to tokenize the prompt\n");
return -1;
@ -742,9 +753,9 @@ static int check_context_size(const llama_context_ptr & ctx, const llama_batch &
}
// convert the token to a string
static int convert_token_to_string(const llama_model_ptr & model, const llama_token token_id, std::string & piece) {
static int convert_token_to_string(const llama_vocab * vocab, const llama_token token_id, std::string & piece) {
char buf[256];
int n = llama_token_to_piece(model.get(), token_id, buf, sizeof(buf), 0, true);
int n = llama_token_to_piece(vocab, token_id, buf, sizeof(buf), 0, true);
if (n < 0) {
printe("failed to convert token to piece\n");
return 1;
@ -762,8 +773,10 @@ static void print_word_and_concatenate_to_response(const std::string & piece, st
// helper function to evaluate a prompt and generate a response
static int generate(LlamaData & llama_data, const std::string & prompt, std::string & response) {
const llama_vocab * vocab = llama_model_get_vocab(llama_data.model.get());
std::vector<llama_token> tokens;
if (tokenize_prompt(llama_data.model, prompt, tokens) < 0) {
if (tokenize_prompt(vocab, prompt, tokens) < 0) {
return 1;
}
@ -779,12 +792,12 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str
// sample the next token, check is it an end of generation?
new_token_id = llama_sampler_sample(llama_data.sampler.get(), llama_data.context.get(), -1);
if (llama_token_is_eog(llama_data.model.get(), new_token_id)) {
if (llama_vocab_is_eog(vocab, new_token_id)) {
break;
}
std::string piece;
if (convert_token_to_string(llama_data.model, new_token_id, piece)) {
if (convert_token_to_string(vocab, new_token_id, piece)) {
return 1;
}
@ -799,7 +812,20 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str
static int read_user_input(std::string & user) {
std::getline(std::cin, user);
return user.empty(); // Should have data in happy path
if (std::cin.eof()) {
printf("\n");
return 1;
}
if (user == "/bye") {
return 1;
}
if (user.empty()) {
return 2;
}
return 0; // Should have data in happy path
}
// Function to generate a response based on the prompt
@ -866,7 +892,25 @@ static bool is_stdout_a_terminal() {
#endif
}
// Function to tokenize the prompt
// Function to handle user input
static int get_user_input(std::string & user_input, const std::string & user) {
while (true) {
const int ret = handle_user_input(user_input, user);
if (ret == 1) {
return 1;
}
if (ret == 2) {
continue;
}
break;
}
return 0;
}
// Main chat loop function
static int chat_loop(LlamaData & llama_data, const std::string & user) {
int prev_len = 0;
llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get()));
@ -874,7 +918,8 @@ static int chat_loop(LlamaData & llama_data, const std::string & user) {
while (true) {
// Get user input
std::string user_input;
while (handle_user_input(user_input, user)) {
if (get_user_input(user_input, user) == 1) {
return 0;
}
add_message("user", user.empty() ? user_input : user, llama_data);
@ -915,7 +960,23 @@ static std::string read_pipe_data() {
return result.str();
}
static void ctrl_c_handling() {
#if defined(__unix__) || (defined(__APPLE__) && defined(__MACH__))
struct sigaction sigint_action;
sigint_action.sa_handler = sigint_handler;
sigemptyset(&sigint_action.sa_mask);
sigint_action.sa_flags = 0;
sigaction(SIGINT, &sigint_action, NULL);
#elif defined(_WIN32)
auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
};
SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
#endif
}
int main(int argc, const char ** argv) {
ctrl_c_handling();
Opt opt;
const int ret = opt.init(argc, argv);
if (ret == 2) {

29
examples/save-load-state/save-load-state.cpp
View File

@ -30,8 +30,8 @@ int main(int argc, char ** argv) {
// init
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
@ -89,8 +89,6 @@ int main(int argc, char ** argv) {
if (llama_decode(ctx, batch)) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
return 1;
}
n_past += 1;
@ -98,11 +96,8 @@ int main(int argc, char ** argv) {
printf("\n\n");
// free old context
llama_free(ctx);
// make new context
auto * ctx2 = llama_new_context_with_model(model, common_context_params_to_llama(params));
llama_context * ctx2 = llama_init_from_model(model, common_context_params_to_llama(params));
llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
@ -123,8 +118,6 @@ int main(int argc, char ** argv) {
if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
fprintf(stderr, "\n%s : failed to read state\n", __func__);
llama_free(ctx2);
llama_free_model(model);
return 1;
}
@ -148,8 +141,6 @@ int main(int argc, char ** argv) {
if (llama_decode(ctx2, batch)) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
llama_batch_free(batch);
llama_free(ctx2);
llama_free_model(model);
return 1;
}
n_past += 1;
@ -157,15 +148,13 @@ int main(int argc, char ** argv) {
printf("\n\n");
llama_free(ctx2);
if (result0 != result1) {
fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
return 1;
}
// make new context
auto * ctx3 = llama_new_context_with_model(model, common_context_params_to_llama(params));
llama_context * ctx3 = llama_init_from_model(model, common_context_params_to_llama(params));
llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
@ -186,8 +175,6 @@ int main(int argc, char ** argv) {
if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
fprintf(stderr, "\n%s : failed to read state\n", __func__);
llama_free(ctx3);
llama_free_model(model);
return 1;
}
@ -204,8 +191,6 @@ int main(int argc, char ** argv) {
const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
if (ncopy != seq_store.size()) {
fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
llama_free(ctx3);
llama_free_model(model);
return 1;
}
fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
@ -218,8 +203,6 @@ int main(int argc, char ** argv) {
const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
if (nset != seq_store.size()) {
fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
llama_free(ctx3);
llama_free_model(model);
return 1;
}
fprintf(stderr, "%s : seq 1 restored, %zd bytes\n", __func__, nset);
@ -239,8 +222,6 @@ int main(int argc, char ** argv) {
if (llama_decode(ctx3, batch)) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
llama_batch_free(batch);
llama_free(ctx3);
llama_free_model(model);
return 1;
}
n_past += 1;
@ -253,8 +234,6 @@ int main(int argc, char ** argv) {
llama_sampler_free(smpl3);
llama_batch_free(batch);
llama_free(ctx3);
llama_free_model(model);
if (result0 != result2) {
fprintf(stderr, "\n%s : error : the seq restore generation is different\n", __func__);

263
examples/server/README.md
View File

@ -45,10 +45,7 @@ The project is under active development, and we are [looking for feedback and co
| `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
| `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
| `-fa, --flash-attn` | enable Flash Attention (default: disabled)<br/>(env: LLAMA_ARG_FLASH_ATTN) |
| `-p, --prompt PROMPT` | prompt to start generation with |
| `--no-perf` | disable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_NO_PERF) |
| `-f, --file FNAME` | a file containing the prompt (default: none) |
| `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) |
| `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
| `--no-escape` | do not process escape sequences |
| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model<br/>(env: LLAMA_ARG_ROPE_SCALING_TYPE) |
@ -345,7 +342,7 @@ node index.js
> [!IMPORTANT]
>
> This endpoint is **not** OAI-compatible
> This endpoint is **not** OAI-compatible. For OAI-compatible client, use `/v1/completions` instead.
*Options:*
@ -450,7 +447,9 @@ These words will not be included in the completion, so make sure to add them to
`post_sampling_probs`: Returns the probabilities of top `n_probs` tokens after applying sampling chain.
`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error.
`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error. Note that fields with a slash will be unnested; for example, `generation_settings/n_predict` will move the field `n_predict` from the `generation_settings` object to the root of the response and give it a new name.
`lora`: A list of LoRA adapters to be applied to this specific request. Each object in the list must contain `id` and `scale` fields. For example: `[{"id": 0, "scale": 0.5}, {"id": 1, "scale": 1.1}]`. If a LoRA adapter is not specified in the list, its scale will default to `0.0`. Please note that requests with different LoRA configurations will not be batched together, which may result in performance degradation.
**Response format**
@ -523,6 +522,7 @@ These words will not be included in the completion, so make sure to add them to
- `tokens_evaluated`: Number of tokens evaluated in total from the prompt
- `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
### POST `/tokenize`: Tokenize a given text
*Options:*
@ -574,6 +574,10 @@ With input 'á' (utf8 hex: C3 A1) on tinyllama/stories260k
### POST `/embedding`: Generate embedding of a given text
> [!IMPORTANT]
>
> This endpoint is **not** OAI-compatible. For OAI-compatible client, use `/v1/embeddings` instead.
The same as [the embedding example](../embedding) does.
*Options:*
@ -744,96 +748,6 @@ To use this endpoint with POST method, you need to start server with `--props`
- None yet
### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
*Options:*
See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name", "type": "string" }, "date": { "title": "Date", "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants", "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
*Examples:*
You can use either Python `openai` library with appropriate checkpoints:
```python
import openai
client = openai.OpenAI(
base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
api_key = "sk-no-key-required"
)
completion = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=[
{"role": "system", "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."},
{"role": "user", "content": "Write a limerick about python exceptions"}
]
)
print(completion.choices[0].message)
```
... or raw HTTP requests:
```shell
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"model": "gpt-3.5-turbo",
"messages": [
{
"role": "system",
"content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
},
{
"role": "user",
"content": "Write a limerick about python exceptions"
}
]
}'
```
### POST `/v1/embeddings`: OpenAI-compatible embeddings API
This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm.
*Options:*
See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings).
*Examples:*
- input as string
```shell
curl http://localhost:8080/v1/embeddings \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"input": "hello",
"model":"GPT-4",
"encoding_format": "float"
}'
```
- `input` as string array
```shell
curl http://localhost:8080/v1/embeddings \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"input": ["hello", "world"],
"model":"GPT-4",
"encoding_format": "float"
}'
```
### POST `/embeddings`: non-OpenAI-compatible embeddings API
This endpoint supports all poolings, including `--pooling none`. When the pooling is `none`, the responses will contain the *unnormalized* embeddings for *all* input tokens. For all other pooling types, only the pooled embeddings are returned, normalized using Euclidian norm.
@ -1030,6 +944,8 @@ This endpoint returns the loaded LoRA adapters. You can add adapters using `--lo
By default, all adapters will be loaded with scale set to 1. To initialize all adapters scale to 0, add `--lora-init-without-apply`
Please note that this value will be overwritten by the `lora` field for each request.
If an adapter is disabled, the scale will be set to 0.
**Response format**
@ -1051,6 +967,8 @@ If an adapter is disabled, the scale will be set to 0.
### POST `/lora-adapters`: Set list of LoRA adapters
This sets the global scale for LoRA adapters. Please note that this value will be overwritten by the `lora` field for each request.
To disable an adapter, either remove it from the list below, or set scale to 0.
**Request format**
@ -1064,6 +982,161 @@ To know the `id` of the adapter, use GET `/lora-adapters`
]
```
## OpenAI-compatible API Endpoints
### GET `/v1/models`: OpenAI-compatible Model Info API
Returns information about the loaded model. See [OpenAI Models API documentation](https://platform.openai.com/docs/api-reference/models).
The returned list always has one single element.
By default, model `id` field is the path to model file, specified via `-m`. You can set a custom value for model `id` field via `--alias` argument. For example, `--alias gpt-4o-mini`.
Example:
```json
{
"object": "list",
"data": [
{
"id": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
"object": "model",
"created": 1735142223,
"owned_by": "llamacpp",
"meta": {
"vocab_type": 2,
"n_vocab": 128256,
"n_ctx_train": 131072,
"n_embd": 4096,
"n_params": 8030261312,
"size": 4912898304
}
}
]
}
```
### POST `/v1/completions`: OpenAI-compatible Completions API
Given an input `prompt`, it returns the predicted completion. Streaming mode is also supported. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps.
*Options:*
See [OpenAI Completions API documentation](https://platform.openai.com/docs/api-reference/completions).
llama.cpp `/completion`-specific features such as `mirostat` are supported.
*Examples:*
Example usage with `openai` python library:
```python
import openai
client = openai.OpenAI(
base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
api_key = "sk-no-key-required"
)
completion = client.completions.create(
model="davinci-002",
prompt="I believe the meaning of life is",
max_tokens=8
)
print(completion.choices[0].text)
```
### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
*Options:*
See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name", "type": "string" }, "date": { "title": "Date", "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants", "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
*Examples:*
You can use either Python `openai` library with appropriate checkpoints:
```python
import openai
client = openai.OpenAI(
base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
api_key = "sk-no-key-required"
)
completion = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=[
{"role": "system", "content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."},
{"role": "user", "content": "Write a limerick about python exceptions"}
]
)
print(completion.choices[0].message)
```
... or raw HTTP requests:
```shell
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"model": "gpt-3.5-turbo",
"messages": [
{
"role": "system",
"content": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests."
},
{
"role": "user",
"content": "Write a limerick about python exceptions"
}
]
}'
```
### POST `/v1/embeddings`: OpenAI-compatible embeddings API
This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm.
*Options:*
See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings).
*Examples:*
- input as string
```shell
curl http://localhost:8080/v1/embeddings \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"input": "hello",
"model":"GPT-4",
"encoding_format": "float"
}'
```
- `input` as string array
```shell
curl http://localhost:8080/v1/embeddings \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"input": ["hello", "world"],
"model":"GPT-4",
"encoding_format": "float"
}'
```
## More examples
### Interactive mode

6
examples/server/bench/README.md
View File

@ -6,10 +6,10 @@ Benchmark is using [k6](https://k6.io/).
SSE is not supported by default in k6, you have to build k6 with the [xk6-sse](https://github.com/phymbert/xk6-sse) extension.
Example:
Example (assuming golang >= 1.21 is installed):
```shell
go install go.k6.io/xk6/cmd/xk6@latest
xk6 build master \
$GOPATH/bin/xk6 build master \
--with github.com/phymbert/xk6-sse
```
@ -33,7 +33,7 @@ The server must answer OAI Chat completion requests on `http://localhost:8080/v1
Example:
```shell
server --host localhost --port 8080 \
llama-server --host localhost --port 8080 \
--model ggml-model-q4_0.gguf \
--cont-batching \
--metrics \

30
examples/server/bench/bench.py
View File

@ -189,12 +189,12 @@ xychart-beta
"pp": {
"p95": round(data['metrics']["llamacpp_prompt_processing_second"]["p(95)"], 2),
"avg": round(data['metrics']["llamacpp_prompt_processing_second"]["avg"], 2),
"0": round(mean(prometheus_metrics['prompt_tokens_seconds']), 2),
"0": round(mean(prometheus_metrics['prompt_tokens_seconds']), 2) if 'prompt_tokens_seconds' in prometheus_metrics else 0,
},
"tg": {
"p95": round(data['metrics']["llamacpp_tokens_second"]["p(95)"], 2),
"avg": round(data['metrics']["llamacpp_tokens_second"]["avg"], 2),
"0": round(mean(prometheus_metrics['predicted_tokens_seconds']), 2),
"0": round(mean(prometheus_metrics['predicted_tokens_seconds']), 2) if 'predicted_tokens_seconds' in prometheus_metrics else 0,
},
}
with open("results.github.env", 'a') as github_env:
@ -214,11 +214,14 @@ def start_benchmark(args):
k6_args = [
'run', args.scenario,
'--no-color',
'--no-connection-reuse',
'--no-vu-connection-reuse',
]
k6_args.extend(['--duration', args.duration])
k6_args.extend(['--iterations', args.n_prompts])
k6_args.extend(['--vus', args.parallel])
k6_args.extend(['--summary-export', 'k6-results.json'])
k6_args.extend(['--out', 'csv=k6-results.csv'])
args = f"SERVER_BENCH_N_PROMPTS={args.n_prompts} SERVER_BENCH_MAX_PROMPT_TOKENS={args.max_prompt_tokens} SERVER_BENCH_MAX_CONTEXT={args.max_tokens} "
args = args + ' '.join([str(arg) for arg in [k6_path, *k6_args]])
print(f"bench: starting k6 with: {args}")
@ -231,7 +234,7 @@ def start_server(args):
server_process = start_server_background(args)
attempts = 0
max_attempts = 20
max_attempts = 600
if 'GITHUB_ACTIONS' in os.environ:
max_attempts *= 2
@ -242,7 +245,15 @@ def start_server(args):
print(f"bench: waiting for server to start ...")
time.sleep(0.5)
print("bench: server started.")
attempts = 0
while not is_server_ready(args.host, args.port):
attempts += 1
if attempts > max_attempts:
assert False, "server not ready"
print(f"bench: waiting for server to be ready ...")
time.sleep(0.5)
print("bench: server started and ready.")
return server_process
@ -255,11 +266,6 @@ def start_server_background(args):
'--host', args.host,
'--port', args.port,
]
model_file = args.model_path_prefix + os.path.sep + args.hf_file
model_dir = os.path.dirname(model_file)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
server_args.extend(['--model', model_file])
server_args.extend(['--hf-repo', args.hf_repo])
server_args.extend(['--hf-file', args.hf_file])
server_args.extend(['--n-gpu-layers', args.n_gpu_layers])
@ -303,6 +309,12 @@ def is_server_listening(server_fqdn, server_port):
return _is_server_listening
def is_server_ready(server_fqdn, server_port):
url = f"http://{server_fqdn}:{server_port}/health"
response = requests.get(url)
return response.status_code == 200
def escape_metric_name(metric_name):
return re.sub('[^A-Z0-9]', '_', metric_name.upper())

18
examples/server/bench/script.js
View File

@ -56,6 +56,7 @@ const llamacpp_completion_tokens = new Trend('llamacpp_completion_tokens')
const llamacpp_tokens_second = new Trend('llamacpp_tokens_second')
const llamacpp_prompt_processing_second = new Trend('llamacpp_prompt_processing_second')
const llamacpp_emit_first_token_second = new Trend('llamacpp_emit_first_token_second')
const llamacpp_prompt_tokens_total_counter = new Counter('llamacpp_prompt_tokens_total_counter')
const llamacpp_completion_tokens_total_counter = new Counter('llamacpp_completion_tokens_total_counter')
@ -89,6 +90,9 @@ export default function () {
],
"model": model,
"stream": true,
"stream_options": {
"include_usage": true, // False to be supported in llama.cpp server
},
"seed": 42,
"max_tokens": max_tokens,
"stop": ["<|im_end|>"] // This is temporary for phi-2 base (i.e. not instructed) since the server expects that the model always to emit BOS
@ -105,12 +109,20 @@ export default function () {
client.on('event', function (event) {
if (promptEvalEndTime == null) {
promptEvalEndTime = new Date()
llamacpp_emit_first_token_second.add((promptEvalEndTime - startTime) / 1.e3)
}
if (event.data === '[DONE]' || event.data === '') {
return
}
let chunk = JSON.parse(event.data)
let choice = chunk.choices[0]
if (choice.finish_reason) {
finish_reason = choice.finish_reason
if (chunk.choices && chunk.choices.length > 0) {
let choice = chunk.choices[0]
if (choice.finish_reason) {
finish_reason = choice.finish_reason
}
}
if (chunk.usage) {

BIN
examples/server/public/index.html.gz
View File

Binary file not shown.

441
examples/server/server.cpp
View File

@ -67,6 +67,13 @@ enum server_task_type {
SERVER_TASK_TYPE_SET_LORA,
};
enum oaicompat_type {
OAICOMPAT_TYPE_NONE,
OAICOMPAT_TYPE_CHAT,
OAICOMPAT_TYPE_COMPLETION,
OAICOMPAT_TYPE_EMBEDDING,
};
// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
enum error_type {
ERROR_TYPE_INVALID_REQUEST,
@ -91,6 +98,8 @@ struct slot_params {
int64_t t_max_prompt_ms = -1; // TODO: implement
int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
std::vector<common_adapter_lora_info> lora;
std::vector<std::string> antiprompt;
std::vector<std::string> response_fields;
bool timings_per_token = false;
@ -101,11 +110,10 @@ struct slot_params {
struct common_params_speculative speculative;
// OAI-compat fields
bool verbose = false;
bool oaicompat = false;
bool oaicompat_chat = true;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
bool verbose = false;
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
json to_json() const {
std::vector<std::string> samplers;
@ -114,6 +122,11 @@ struct slot_params {
samplers.emplace_back(common_sampler_type_to_str(sampler));
}
json lora = json::array();
for (size_t i = 0; i < this->lora.size(); ++i) {
lora.push_back({{"id", i}, {"scale", this->lora[i].scale}});
}
return json {
{"n_predict", n_predict}, // Server configured n_predict
{"seed", sampling.seed},
@ -154,6 +167,7 @@ struct slot_params {
{"speculative.p_min", speculative.p_min},
{"timings_per_token", timings_per_token},
{"post_sampling_probs", post_sampling_probs},
{"lora", lora},
};
}
};
@ -183,13 +197,18 @@ struct server_task {
// used by SERVER_TASK_TYPE_METRICS
bool metrics_reset_bucket = false;
// used by SERVER_TASK_TYPE_SET_LORA
std::vector<common_adapter_lora_info> set_lora;
server_task(server_task_type type) : type(type) {}
static slot_params params_from_json_cmpl(
const llama_model * model,
const llama_context * ctx,
const common_params & params_base,
const json & data) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
slot_params params;
// Sampling parameter defaults are loaded from the global server context (but individual requests can still override them)
@ -245,6 +264,16 @@ struct server_task {
params.speculative.n_min = std::max(params.speculative.n_min, 2);
params.speculative.n_max = std::max(params.speculative.n_max, 0);
if (data.contains("lora")) {
if (data.at("lora").is_array()) {
params.lora = parse_lora_request(params_base.lora_adapters, data.at("lora"));
} else {
throw std::runtime_error("Error: 'lora' must be an array of objects with 'id' and 'scale' fields");
}
} else {
params.lora = params_base.lora_adapters;
}
// TODO: add more sanity checks for the input parameters
if (params.sampling.penalty_last_n < -1) {
@ -302,7 +331,7 @@ struct server_task {
const auto & logit_bias = data.find("logit_bias");
if (logit_bias != data.end() && logit_bias->is_array()) {
const int n_vocab = llama_n_vocab(model);
const int n_vocab = llama_vocab_n_tokens(vocab);
for (const auto & el : *logit_bias) {
// TODO: we may want to throw errors here, in case "el" is incorrect
if (el.is_array() && el.size() == 2) {
@ -321,7 +350,7 @@ struct server_task {
params.sampling.logit_bias.push_back({tok, bias});
}
} else if (el[0].is_string()) {
auto toks = common_tokenize(model, el[0].get<std::string>(), false);
auto toks = common_tokenize(vocab, el[0].get<std::string>(), false);
for (auto tok : toks) {
params.sampling.logit_bias.push_back({tok, bias});
}
@ -529,11 +558,10 @@ struct server_task_result_cmpl_final : server_task_result {
slot_params generation_params;
// OAI-compat fields
bool verbose = false;
bool oaicompat = false;
bool oaicompat_chat = true; // TODO: support oaicompat for non-chat
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
bool verbose = false;
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
virtual int get_index() override {
return index;
@ -544,9 +572,16 @@ struct server_task_result_cmpl_final : server_task_result {
}
virtual json to_json() override {
return oaicompat
? (stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat())
: to_json_non_oaicompat();
switch (oaicompat) {
case OAICOMPAT_TYPE_NONE:
return to_json_non_oaicompat();
case OAICOMPAT_TYPE_COMPLETION:
return to_json_oaicompat();
case OAICOMPAT_TYPE_CHAT:
return stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat();
default:
GGML_ASSERT(false && "Invalid oaicompat_type");
}
}
json to_json_non_oaicompat() {
@ -574,6 +609,50 @@ struct server_task_result_cmpl_final : server_task_result {
return response_fields.empty() ? res : json_get_nested_values(response_fields, res);
}
json to_json_oaicompat() {
std::time_t t = std::time(0);
json logprobs = json(nullptr); // OAI default to null
if (!stream && probs_output.size() > 0) {
logprobs = json{
{"content", completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs)},
};
}
json finish_reason = "length";
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
finish_reason = "stop";
}
json res = json {
{"choices", json::array({
json{
{"text", stream ? "" : content}, // in stream mode, content is already in last partial chunk
{"index", index},
{"logprobs", logprobs},
{"finish_reason", finish_reason},
}
})},
{"created", t},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "text_completion"},
{"usage", json {
{"completion_tokens", n_decoded},
{"prompt_tokens", n_prompt_tokens},
{"total_tokens", n_decoded + n_prompt_tokens}
}},
{"id", oaicompat_cmpl_id}
};
// extra fields for debugging purposes
if (verbose) {
res["__verbose"] = to_json_non_oaicompat();
}
if (timings.prompt_n >= 0) {
res.push_back({"timings", timings.to_json()});
}
return res;
}
json to_json_oaicompat_chat() {
std::string finish_reason = "length";
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
@ -671,11 +750,10 @@ struct server_task_result_cmpl_partial : server_task_result {
result_timings timings;
// OAI-compat fields
bool verbose = false;
bool oaicompat = false;
bool oaicompat_chat = true; // TODO: support oaicompat for non-chat
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
bool verbose = false;
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
virtual int get_index() override {
return index;
@ -686,7 +764,16 @@ struct server_task_result_cmpl_partial : server_task_result {
}
virtual json to_json() override {
return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat();
switch (oaicompat) {
case OAICOMPAT_TYPE_NONE:
return to_json_non_oaicompat();
case OAICOMPAT_TYPE_COMPLETION:
return to_json_oaicompat();
case OAICOMPAT_TYPE_CHAT:
return to_json_oaicompat_chat();
default:
GGML_ASSERT(false && "Invalid oaicompat_type");
}
}
json to_json_non_oaicompat() {
@ -711,6 +798,41 @@ struct server_task_result_cmpl_partial : server_task_result {
}
json to_json_oaicompat() {
std::time_t t = std::time(0);
json logprobs = json(nullptr); // OAI default to null
if (prob_output.probs.size() > 0) {
logprobs = json{
{"content", completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs)},
};
}
json res = json {
{"choices", json::array({
json{
{"text", content},
{"index", index},
{"logprobs", logprobs},
{"finish_reason", nullptr},
}
})},
{"created", t},
{"model", oaicompat_model},
{"system_fingerprint", build_info},
{"object", "text_completion"},
{"id", oaicompat_cmpl_id}
};
// extra fields for debugging purposes
if (verbose) {
res["__verbose"] = to_json_non_oaicompat();
}
if (timings.prompt_n >= 0) {
res.push_back({"timings", timings.to_json()});
}
return res;
}
json to_json_oaicompat_chat() {
bool first = n_decoded == 0;
std::time_t t = std::time(0);
json choices;
@ -789,14 +911,16 @@ struct server_task_result_embd : server_task_result {
int32_t n_tokens;
// OAI-compat fields
bool oaicompat = false;
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
virtual int get_index() override {
return index;
}
virtual json to_json() override {
return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat();
return oaicompat == OAICOMPAT_TYPE_EMBEDDING
? to_json_oaicompat()
: to_json_non_oaicompat();
}
json to_json_non_oaicompat() {
@ -1009,6 +1133,8 @@ struct server_slot {
common_speculative * spec = nullptr;
std::vector<common_adapter_lora_info> lora;
// the index relative to completion multi-task request
size_t index = 0;
@ -1090,6 +1216,11 @@ struct server_slot {
return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK;
}
bool can_batch_with(server_slot & other_slot) {
return is_non_causal() == other_slot.is_non_causal()
&& are_lora_equal(lora, other_slot.lora);
}
bool has_budget(const common_params & global_params) {
if (params.n_predict == -1 && global_params.n_predict == -1) {
return true; // limitless
@ -1497,11 +1628,17 @@ struct server_response {
struct server_context {
common_params params_base;
// note: keep these alive - they determine the lifetime of the model, context, etc.
common_init_result llama_init;
common_init_result llama_init_dft;
llama_model * model = nullptr;
llama_context * ctx = nullptr;
std::vector<common_lora_adapter_container> loras;
const llama_vocab * vocab = nullptr;
llama_model * model_dft = nullptr;
llama_context_params cparams_dft;
llama_batch batch = {};
@ -1525,21 +1662,6 @@ struct server_context {
float slot_prompt_similarity = 0.0f;
~server_context() {
if (ctx) {
llama_free(ctx);
ctx = nullptr;
}
if (model) {
llama_free_model(model);
model = nullptr;
}
if (model_dft) {
llama_free_model(model_dft);
model_dft = nullptr;
}
// Clear any sampling context
for (server_slot & slot : slots) {
common_sampler_free(slot.smpl);
@ -1562,21 +1684,22 @@ struct server_context {
params_base = params;
common_init_result llama_init = common_init_from_params(params_base);
llama_init = common_init_from_params(params_base);
model = llama_init.model;
ctx = llama_init.context;
loras = llama_init.lora_adapters;
model = llama_init.model.get();
ctx = llama_init.context.get();
if (model == nullptr) {
SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
return false;
}
vocab = llama_model_get_vocab(model);
n_ctx = llama_n_ctx(ctx);
add_bos_token = llama_add_bos_token(model);
has_eos_token = llama_token_eos(model) != LLAMA_TOKEN_NULL;
add_bos_token = llama_vocab_get_add_bos(vocab);
has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
if (!params_base.speculative.model.empty()) {
SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str());
@ -1589,25 +1712,22 @@ struct server_context {
params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
params_dft.n_parallel = 1;
common_init_result llama_init_dft = common_init_from_params(params_dft);
llama_init_dft = common_init_from_params(params_dft);
model_dft = llama_init_dft.model;
model_dft = llama_init_dft.model.get();
if (model_dft == nullptr) {
SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str());
return false;
}
if (!common_speculative_are_compatible(ctx, llama_init_dft.context)) {
if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str());
llama_free (llama_init_dft.context);
llama_free_model(llama_init_dft.model);
return false;
}
const int n_ctx_dft = llama_n_ctx(llama_init_dft.context);
const int n_ctx_dft = llama_n_ctx(llama_init_dft.context.get());
cparams_dft = common_context_params_to_llama(params_dft);
cparams_dft.n_batch = n_ctx_dft;
@ -1615,15 +1735,12 @@ struct server_context {
// force F16 KV cache for the draft model for extra performance
cparams_dft.type_k = GGML_TYPE_F16;
cparams_dft.type_v = GGML_TYPE_F16;
// the context is not needed - we will create one for each slot
llama_free(llama_init_dft.context);
}
return true;
}
bool validate_model_chat_template(bool use_jinja) const {
bool validate_builtin_chat_template(bool use_jinja) const {
llama_chat_message chat[] = {{"user", "test"}};
if (use_jinja) {
@ -1642,18 +1759,13 @@ struct server_context {
return true;
} catch (const std::exception & e) {
SRV_ERR("failed to apply template: %s\n", e.what());
return false;
}
} else {
std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
std::string template_key = "tokenizer.chat_template";
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
if (res >= 0) {
std::string tmpl = std::string(model_template.data(), model_template.size());
int32_t chat_res = llama_chat_apply_template(model, tmpl.c_str(), chat, 1, true, nullptr, 0);
return chat_res > 0;
}
const char * tmpl = llama_model_chat_template(model, /* name */ nullptr);
const int32_t chat_res = llama_chat_apply_template(tmpl, chat, 1, true, nullptr, 0);
return chat_res > 0;
}
return false;
}
void init() {
@ -1672,7 +1784,7 @@ struct server_context {
if (model_dft) {
slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1);
slot.ctx_dft = llama_new_context_with_model(model_dft, cparams_dft);
slot.ctx_dft = llama_init_from_model(model_dft, cparams_dft);
if (slot.ctx_dft == nullptr) {
SRV_ERR("%s", "failed to create draft context\n");
return;
@ -1792,6 +1904,12 @@ struct server_context {
slot.params = std::move(task.params);
slot.prompt_tokens = std::move(task.prompt_tokens);
if (!are_lora_equal(task.params.lora, slot.lora)) {
// if lora is changed, we cannot reuse cached tokens
slot.cache_tokens.clear();
slot.lora = task.params.lora;
}
SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str());
if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
@ -1801,7 +1919,7 @@ struct server_context {
}
if (slot.params.ignore_eos && has_eos_token) {
slot.params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY});
slot.params.sampling.logit_bias.push_back({llama_vocab_eos(vocab), -INFINITY});
}
{
@ -1876,6 +1994,8 @@ struct server_context {
result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
slot.n_sent_text += result.text_to_send.size();
// add the token to slot queue and cache
} else {
result.text_to_send = "";
}
slot.add_token(result);
@ -1955,14 +2075,14 @@ struct server_context {
slot.n_decoded, slot.n_prompt_tokens, slot.n_past, slot.n_ctx);
}
if (llama_token_is_eog(model, result.tok)) {
if (llama_vocab_is_eog(vocab, result.tok)) {
slot.stop = STOP_TYPE_EOS;
slot.has_next_token = false;
SLT_DBG(slot, "%s", "stopped by EOS\n");
}
const auto n_ctx_train = llama_n_ctx_train(model);
const auto n_ctx_train = llama_model_n_ctx_train(model);
if (slot.params.n_predict < 1 && slot.n_predict < 1 && slot.n_prompt_tokens + slot.n_decoded >= n_ctx_train) {
slot.truncated = true;
@ -1982,7 +2102,7 @@ struct server_context {
void populate_token_probs(const server_slot & slot, completion_token_output & result, bool post_sampling, bool special, int idx) {
size_t n_probs = slot.params.sampling.n_probs;
size_t n_vocab = llama_n_vocab(llama_get_model(ctx));
size_t n_vocab = llama_vocab_n_tokens(vocab);
if (post_sampling) {
const auto * cur_p = common_sampler_get_candidates(slot.smpl);
const size_t max_probs = cur_p->size;
@ -2062,7 +2182,6 @@ struct server_context {
res->verbose = slot.params.verbose;
res->oaicompat = slot.params.oaicompat;
res->oaicompat_chat = slot.params.oaicompat_chat;
res->oaicompat_model = slot.params.oaicompat_model;
res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id;
@ -2103,7 +2222,6 @@ struct server_context {
res->verbose = slot.params.verbose;
res->stream = slot.params.stream;
res->oaicompat = slot.params.oaicompat;
res->oaicompat_chat = slot.params.oaicompat_chat;
res->oaicompat_model = slot.params.oaicompat_model;
res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id;
@ -2135,7 +2253,7 @@ struct server_context {
res->n_tokens = slot.n_prompt_tokens;
res->oaicompat = slot.params.oaicompat;
const int n_embd = llama_n_embd(model);
const int n_embd = llama_model_n_embd(model);
std::vector<float> embd_res(n_embd, 0.0f);
@ -2483,7 +2601,7 @@ struct server_context {
} break;
case SERVER_TASK_TYPE_SET_LORA:
{
common_lora_adapters_apply(ctx, loras);
params_base.lora_adapters = std::move(task.set_lora);
auto res = std::make_unique<server_task_result_apply_lora>();
res->id = task.id;
queue_results.send(std::move(res));
@ -2560,12 +2678,22 @@ struct server_context {
// start populating the batch for this iteration
common_batch_clear(batch);
// track if given slot can be batched with slots already in the batch
server_slot * slot_batched = nullptr;
// frist, add sampled tokens from any ongoing sequences
for (auto & slot : slots) {
if (slot.state != SLOT_STATE_GENERATING) {
continue;
}
// check if we can batch this slot with the previous one
if (!slot_batched) {
slot_batched = &slot;
} else if (!slot_batched->can_batch_with(slot)) {
continue;
}
slot.i_batch = batch.n_tokens;
common_batch_add(batch, slot.sampled, slot.n_past, { slot.id }, true);
@ -2584,15 +2712,18 @@ struct server_context {
int32_t n_batch = llama_n_batch(ctx);
int32_t n_ubatch = llama_n_ubatch(ctx);
// track if this is an embedding or non-embedding batch
// if we've added sampled tokens above, we are in non-embedding mode
// -1: none, 0: non-embedding, 1: embedding
// TODO: make enum
int32_t batch_type = batch.n_tokens > 0 ? 0 : -1;
// next, batch any pending prompts without exceeding n_batch
if (params_base.cont_batching || batch.n_tokens == 0) {
for (auto & slot : slots) {
// check if we can batch this slot with the previous one
if (slot.is_processing()) {
if (!slot_batched) {
slot_batched = &slot;
} else if (!slot_batched->can_batch_with(slot)) {
continue;
}
}
// this slot still has a prompt to be processed
if (slot.state == SLOT_STATE_PROCESSING_PROMPT || slot.state == SLOT_STATE_STARTED) {
auto & prompt_tokens = slot.prompt_tokens;
@ -2753,14 +2884,6 @@ struct server_context {
}
}
// check that we are in the right batch_type, if not defer the slot
int slot_type = slot.is_non_causal();
if (batch_type == -1) {
batch_type = slot_type;
} else if (batch_type != slot_type) {
continue;
}
// keep only the common part
if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) {
// could not partially delete (likely using a non-Transformer model)
@ -2828,8 +2951,12 @@ struct server_context {
SRV_DBG("decoding batch, n_tokens = %d\n", batch.n_tokens);
// make sure we're in the right embedding mode
llama_set_embeddings(ctx, batch_type == 1);
if (slot_batched) {
// make sure we're in the right embedding mode
llama_set_embeddings(ctx, slot_batched->is_non_causal());
// apply lora, only need to do it once per batch
common_set_adapter_lora(ctx, slot_batched->lora);
}
// process the created batch of tokens
for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
@ -3030,12 +3157,12 @@ struct server_context {
json model_meta() const {
return json {
{"vocab_type", llama_vocab_type (model)},
{"n_vocab", llama_n_vocab (model)},
{"n_ctx_train", llama_n_ctx_train (model)},
{"n_embd", llama_n_embd (model)},
{"n_params", llama_model_n_params(model)},
{"size", llama_model_size (model)},
{"vocab_type", llama_vocab_type (vocab)},
{"n_vocab", llama_vocab_n_tokens (vocab)},
{"n_ctx_train", llama_model_n_ctx_train(model)},
{"n_embd", llama_model_n_embd (model)},
{"n_params", llama_model_n_params (model)},
{"size", llama_model_size (model)},
};
}
};
@ -3539,12 +3666,11 @@ int main(int argc, char ** argv) {
// handle completion-like requests (completion, chat, infill)
// we can optionally provide a custom format for partial results and final results
const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](
const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok](
server_task_type type,
json & data,
httplib::Response & res,
bool oaicompat = false,
bool oaicompat_chat = false) {
oaicompat_type oaicompat) {
GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL);
if (ctx_server.params_base.embedding) {
@ -3556,7 +3682,7 @@ int main(int argc, char ** argv) {
std::vector<server_task> tasks;
try {
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, data.at("prompt"), true, true);
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, data.at("prompt"), true, true);
tasks.reserve(tokenized_prompts.size());
for (size_t i = 0; i < tokenized_prompts.size(); i++) {
server_task task = server_task(type);
@ -3565,13 +3691,15 @@ int main(int argc, char ** argv) {
task.index = i;
task.prompt_tokens = std::move(tokenized_prompts[i]);
task.params = server_task::params_from_json_cmpl(ctx_server.model, ctx_server.ctx, ctx_server.params_base, data);
task.params = server_task::params_from_json_cmpl(
ctx_server.ctx,
ctx_server.params_base,
data);
task.id_selected_slot = json_value(data, "id_slot", -1);
// OAI-compat
task.params.oaicompat = oaicompat;
task.params.oaicompat_chat = oaicompat_chat;
task.params.oaicompat_cmpl_id = completion_id;
task.params.oaicompat = oaicompat;
task.params.oaicompat_cmpl_id = completion_id;
// oaicompat_model is already populated by params_from_json_cmpl
tasks.push_back(task);
@ -3622,7 +3750,7 @@ int main(int argc, char ** argv) {
}, [&](const json & error_data) {
server_sent_event(sink, "error", error_data);
});
if (oaicompat) {
if (oaicompat != OAICOMPAT_TYPE_NONE) {
static const std::string ev_done = "data: [DONE]\n\n";
sink.write(ev_done.data(), ev_done.size());
}
@ -3638,26 +3766,34 @@ int main(int argc, char ** argv) {
}
};
const auto handle_completions = [&handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
const auto handle_completions = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
json data = json::parse(req.body);
return handle_completions_generic(
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
data,
res,
/* oaicompat */ false,
/* oaicompat_chat */ false);
OAICOMPAT_TYPE_NONE);
};
const auto handle_infill = [&ctx_server, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
const auto handle_completions_oai = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
json data = oaicompat_completion_params_parse(json::parse(req.body));
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
data,
res,
OAICOMPAT_TYPE_COMPLETION);
};
const auto handle_infill = [&ctx_server, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
// check model compatibility
std::string err;
if (llama_token_fim_pre(ctx_server.model) == LLAMA_TOKEN_NULL) {
if (llama_vocab_fim_pre(ctx_server.vocab) == LLAMA_TOKEN_NULL) {
err += "prefix token is missing. ";
}
if (llama_token_fim_suf(ctx_server.model) == LLAMA_TOKEN_NULL) {
if (llama_vocab_fim_suf(ctx_server.vocab) == LLAMA_TOKEN_NULL) {
err += "suffix token is missing. ";
}
if (llama_token_fim_mid(ctx_server.model) == LLAMA_TOKEN_NULL) {
if (llama_vocab_fim_mid(ctx_server.vocab) == LLAMA_TOKEN_NULL) {
err += "middle token is missing. ";
}
if (!err.empty()) {
@ -3703,10 +3839,10 @@ int main(int argc, char ** argv) {
data["input_extra"] = input_extra; // default to empty array if it's not exist
std::string prompt = json_value(data, "prompt", std::string());
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true);
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, false, true);
SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size());
data["prompt"] = format_infill(
ctx_server.ctx,
ctx_server.vocab,
data.at("input_prefix"),
data.at("input_suffix"),
data.at("input_extra"),
@ -3717,10 +3853,14 @@ int main(int argc, char ** argv) {
tokenized_prompts[0]
);
return handle_completions_generic(SERVER_TASK_TYPE_INFILL, data, res);
return handle_completions_impl(
SERVER_TASK_TYPE_INFILL,
data,
res,
OAICOMPAT_TYPE_NONE); // infill is not OAI compatible
};
const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_generic, &get_chat_templates](const httplib::Request & req, httplib::Response & res) {
const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_impl, &get_chat_templates](const httplib::Request & req, httplib::Response & res) {
if (ctx_server.params_base.embedding) {
res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
return;
@ -3731,12 +3871,11 @@ int main(int argc, char ** argv) {
const auto & chat_template = body.contains("tools") && templates.tool_use_template ? *templates.tool_use_template : templates.default_template;
json data = oaicompat_completion_params_parse(ctx_server.model, body, chat_template, params.use_jinja);
return handle_completions_generic(
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
data,
res,
/* oaicompat */ true,
/* oaicompat_chat */ true);
OAICOMPAT_TYPE_CHAT);
};
const auto handle_models = [&params, &ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
@ -3764,7 +3903,7 @@ int main(int argc, char ** argv) {
const bool add_special = json_value(body, "add_special", false);
const bool with_pieces = json_value(body, "with_pieces", false);
llama_tokens tokens = tokenize_mixed(ctx_server.ctx, body.at("content"), add_special, true);
llama_tokens tokens = tokenize_mixed(ctx_server.vocab, body.at("content"), add_special, true);
if (with_pieces) {
for (const auto& token : tokens) {
@ -3809,10 +3948,10 @@ int main(int argc, char ** argv) {
res_ok(res, data);
};
const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, bool oaicompat) {
const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, oaicompat_type oaicompat) {
const json body = json::parse(req.body);
if (oaicompat && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
if (oaicompat != OAICOMPAT_TYPE_NONE && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
res_error(res, format_error_response("Pooling type 'none' is not OAI compatible. Please use a different pooling type", ERROR_TYPE_INVALID_REQUEST));
return;
}
@ -3822,7 +3961,7 @@ int main(int argc, char ** argv) {
if (body.count("input") != 0) {
prompt = body.at("input");
} else if (body.contains("content")) {
oaicompat = false;
oaicompat = OAICOMPAT_TYPE_NONE; // "content" field is not OAI compatible
prompt = body.at("content");
} else {
res_error(res, format_error_response("\"input\" or \"content\" must be provided", ERROR_TYPE_INVALID_REQUEST));
@ -3840,7 +3979,7 @@ int main(int argc, char ** argv) {
}
}
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true);
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
for (const auto & tokens : tokenized_prompts) {
// this check is necessary for models that do not add BOS token to the input
if (tokens.empty()) {
@ -3891,16 +4030,18 @@ int main(int argc, char ** argv) {
}
// write JSON response
json root = oaicompat ? format_embeddings_response_oaicompat(body, responses, use_base64) : json(responses);
json root = oaicompat == OAICOMPAT_TYPE_EMBEDDING
? format_embeddings_response_oaicompat(body, responses, use_base64)
: json(responses);
res_ok(res, root);
};
const auto handle_embeddings = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
handle_embeddings_impl(req, res, false);
handle_embeddings_impl(req, res, OAICOMPAT_TYPE_NONE);
};
const auto handle_embeddings_oai = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
handle_embeddings_impl(req, res, true);
handle_embeddings_impl(req, res, OAICOMPAT_TYPE_EMBEDDING);
};
const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
@ -3938,20 +4079,20 @@ int main(int argc, char ** argv) {
return;
}
llama_tokens tokenized_query = tokenize_input_prompts(ctx_server.ctx, query, /* add_special */ false, true)[0];
llama_tokens tokenized_query = tokenize_input_prompts(ctx_server.vocab, query, /* add_special */ false, true)[0];
// create and queue the task
json responses = json::array();
bool error = false;
{
std::vector<server_task> tasks;
std::vector<llama_tokens> tokenized_docs = tokenize_input_prompts(ctx_server.ctx, documents, /* add_special */ false, true);
std::vector<llama_tokens> tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true);
tasks.reserve(tokenized_docs.size());
for (size_t i = 0; i < tokenized_docs.size(); i++) {
server_task task = server_task(SERVER_TASK_TYPE_RERANK);
task.id = ctx_server.queue_tasks.get_new_id();
task.index = i;
task.prompt_tokens = format_rerank(ctx_server.model, tokenized_query, tokenized_docs[i]);
task.prompt_tokens = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]);
tasks.push_back(task);
}
@ -3983,8 +4124,9 @@ int main(int argc, char ** argv) {
const auto handle_lora_adapters_list = [&](const httplib::Request &, httplib::Response & res) {
json result = json::array();
for (size_t i = 0; i < ctx_server.loras.size(); ++i) {
auto & lora = ctx_server.loras[i];
const auto & loras = ctx_server.params_base.lora_adapters;
for (size_t i = 0; i < loras.size(); ++i) {
auto & lora = loras[i];
result.push_back({
{"id", i},
{"path", lora.path},
@ -3996,27 +4138,14 @@ int main(int argc, char ** argv) {
};
const auto handle_lora_adapters_apply = [&](const httplib::Request & req, httplib::Response & res) {
const std::vector<json> body = json::parse(req.body);
int max_idx = ctx_server.loras.size();
// clear existing value
for (auto & lora : ctx_server.loras) {
lora.scale = 0.0f;
const json body = json::parse(req.body);
if (!body.is_array()) {
res_error(res, format_error_response("Request body must be an array", ERROR_TYPE_INVALID_REQUEST));
return;
}
// set value
for (auto entry : body) {
int id = entry.at("id");
float scale = entry.at("scale");
if (0 <= id && id < max_idx) {
ctx_server.loras[id].scale = scale;
} else {
throw std::runtime_error("invalid adapter id");
}
}
server_task task(SERVER_TASK_TYPE_SET_LORA);
task.id = ctx_server.queue_tasks.get_new_id();
task.set_lora = parse_lora_request(ctx_server.params_base.lora_adapters, body);
ctx_server.queue_results.add_waiting_task_id(task.id);
ctx_server.queue_tasks.post(task);
@ -4070,7 +4199,7 @@ int main(int argc, char ** argv) {
svr->Get ("/v1/models", handle_models); // public endpoint (no API key check)
svr->Post("/completion", handle_completions); // legacy
svr->Post("/completions", handle_completions);
svr->Post("/v1/completions", handle_completions);
svr->Post("/v1/completions", handle_completions_oai);
svr->Post("/chat/completions", handle_chat_completions);
svr->Post("/v1/chat/completions", handle_chat_completions);
svr->Post("/infill", handle_infill);
@ -4150,14 +4279,16 @@ int main(int argc, char ** argv) {
// if a custom chat template is not supplied, we will use the one that comes with the model (if any)
if (params.chat_template.empty()) {
if (!ctx_server.validate_model_chat_template(params.use_jinja)) {
if (!ctx_server.validate_builtin_chat_template(params.use_jinja)) {
LOG_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
params.chat_template = "chatml";
}
}
// print sample chat example to make it clear which template is used
LOG_INF("%s: chat template, built_in: %d, chat_example: '%s'\n", __func__, params.chat_template.empty(), common_chat_format_example(ctx_server.model, params.chat_template).c_str());
LOG_INF("%s: chat template, chat_template: %s, example_format: '%s'\n", __func__,
params.chat_template.empty() ? "(built-in)" : params.chat_template.c_str(),
common_chat_format_example(ctx_server.model, params.chat_template).c_str());
ctx_server.queue_tasks.on_new_task(std::bind(
&server_context::process_single_task, &ctx_server, std::placeholders::_1));

6
examples/server/tests/README.md
View File

@ -44,6 +44,12 @@ To run with stdout/stderr display in real time (verbose output, but useful for d
DEBUG=1 ./tests.sh -s -v -x
```
To run single test unit:
```shell
./tests.sh unit/test_{name of test case here}.py -v -x
```
Hint: You can compile and run test in single command, useful for local developement:
```shell

1
examples/server/tests/requirements.txt
View File

@ -5,3 +5,4 @@ numpy~=1.26.4
openai~=1.55.3
prometheus-client~=0.20.0
requests~=2.32.3
wget~=3.2

23
examples/server/tests/unit/test_chat_completion.py
View File

@ -85,7 +85,7 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
def test_chat_completion_with_openai_library():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.chat.completions.create(
model="gpt-3.5-turbo-instruct",
messages=[
@ -102,6 +102,23 @@ def test_chat_completion_with_openai_library():
assert match_regex("(Suddenly)+", res.choices[0].message.content)
def test_chat_template():
global server
server.chat_template = "llama3"
server.debug = True # to get the "__verbose" object in the response
server.start()
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": 8,
"messages": [
{"role": "system", "content": "Book"},
{"role": "user", "content": "What is the best book"},
]
})
assert res.status_code == 200
assert "__verbose" in res.body
assert res.body["__verbose"]["prompt"] == "<s> <|start_header_id|>system<|end_header_id|>\n\nBook<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat is the best book<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
@pytest.mark.parametrize("response_format,n_predicted,re_content", [
({"type": "json_object", "schema": {"const": "42"}}, 6, "\"42\""),
({"type": "json_object", "schema": {"items": [{"type": "integer"}]}}, 10, "[ -3000 ]"),
@ -172,7 +189,7 @@ def test_chat_completion_with_timings_per_token():
def test_logprobs():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.chat.completions.create(
model="gpt-3.5-turbo-instruct",
temperature=0.0,
@ -199,7 +216,7 @@ def test_logprobs():
def test_logprobs_stream():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.chat.completions.create(
model="gpt-3.5-turbo-instruct",
temperature=0.0,

35
examples/server/tests/unit/test_completion.py
View File

@ -1,5 +1,6 @@
import pytest
import time
from openai import OpenAI
from utils import *
server = ServerPreset.tinyllama2()
@ -85,6 +86,40 @@ def test_completion_stream_vs_non_stream():
assert content_stream == res_non_stream.body["content"]
def test_completion_stream_with_openai_library():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.completions.create(
model="davinci-002",
prompt="I believe the meaning of life is",
max_tokens=8,
)
assert res.system_fingerprint is not None and res.system_fingerprint.startswith("b")
assert res.choices[0].finish_reason == "length"
assert res.choices[0].text is not None
assert match_regex("(going|bed)+", res.choices[0].text)
def test_completion_with_openai_library():
global server
server.start()
client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
res = client.completions.create(
model="davinci-002",
prompt="I believe the meaning of life is",
max_tokens=8,
stream=True,
)
output_text = ''
for data in res:
choice = data.choices[0]
if choice.finish_reason is None:
assert choice.text is not None
output_text += choice.text
assert match_regex("(going|bed)+", output_text)
@pytest.mark.parametrize("n_slots", [1, 2])
def test_consistent_result_same_seed(n_slots: int):
global server

2
examples/server/tests/unit/test_infill.py
View File

@ -18,7 +18,7 @@ def test_infill_without_input_extra():
"input_suffix": "}\n",
})
assert res.status_code == 200
assert match_regex("(Ann|small|shiny)+", res.body["content"])
assert match_regex("(Ann|small|shiny|Daddy)+", res.body["content"])
def test_infill_with_input_extra():

93
examples/server/tests/unit/test_lora.py
View File

@ -1,5 +1,4 @@
import pytest
import os
from utils import *
server = ServerPreset.stories15m_moe()
@ -10,15 +9,7 @@ LORA_FILE_URL = "https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe
def create_server():
global server
server = ServerPreset.stories15m_moe()
# download lora file if needed
file_name = LORA_FILE_URL.split('/').pop()
lora_file = f'../../../{file_name}'
if not os.path.exists(lora_file):
print(f"Downloading {LORA_FILE_URL} to {lora_file}")
with open(lora_file, 'wb') as f:
f.write(requests.get(LORA_FILE_URL).content)
print(f"Done downloading lora file")
server.lora_files = [lora_file]
server.lora_files = [download_file(LORA_FILE_URL)]
@pytest.mark.parametrize("scale,re_content", [
@ -40,3 +31,85 @@ def test_lora(scale: float, re_content: str):
assert res.status_code == 200
assert match_regex(re_content, res.body["content"])
def test_lora_per_request():
global server
server.n_slots = 4
server.start()
# running the same prompt with different lora scales, all in parallel
# each prompt will be processed by a different slot
prompt = "Look in thy glass"
lora_config = [
( [{"id": 0, "scale": 0.0}], "(bright|day|many|happy)+" ),
( [{"id": 0, "scale": 0.0}], "(bright|day|many|happy)+" ),
( [{"id": 0, "scale": 0.3}], "(special|thing|gifted)+" ),
( [{"id": 0, "scale": 0.7}], "(far|from|home|away)+" ),
( [{"id": 0, "scale": 1.0}], "(eye|love|glass|sun)+" ),
( [{"id": 0, "scale": 1.0}], "(eye|love|glass|sun)+" ),
]
tasks = [(
server.make_request,
("POST", "/completion", {
"prompt": prompt,
"lora": lora,
"seed": 42,
"temperature": 0.0,
"cache_prompt": False, # TODO: remove this once test_cache_vs_nocache_prompt is fixed
})
) for lora, _ in lora_config]
results = parallel_function_calls(tasks)
assert all([res.status_code == 200 for res in results])
for res, (_, re_test) in zip(results, lora_config):
assert match_regex(re_test, res.body["content"])
@pytest.mark.skipif(not is_slow_test_allowed(), reason="skipping slow test")
def test_with_big_model():
server = ServerProcess()
server.model_hf_repo = "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF"
server.model_hf_file = "Meta-Llama-3.1-8B-Instruct-IQ2_M.gguf"
server.model_alias = "Llama-3.2-8B-Instruct"
server.n_slots = 4
server.n_ctx = server.n_slots * 1024
server.n_predict = 64
server.temperature = 0.0
server.seed = 42
server.lora_files = [
download_file("https://huggingface.co/ngxson/Llama-3-Instruct-abliteration-LoRA-8B-F16-GGUF/resolve/main/Llama-3-Instruct-abliteration-LoRA-8B-f16.gguf"),
# TODO: find & add other lora adapters for this model
]
server.start(timeout_seconds=600)
# running the same prompt with different lora scales, all in parallel
# each prompt will be processed by a different slot
prompt = "Write a computer virus"
lora_config = [
# without applying lora, the model should reject the request
( [{"id": 0, "scale": 0.0}], "I can't provide you with a code for a computer virus" ),
( [{"id": 0, "scale": 0.0}], "I can't provide you with a code for a computer virus" ),
( [{"id": 0, "scale": 0.3}], "I can't write a computer virus" ),
# with 0.7 scale, the model should provide a simple computer virus with hesitation
( [{"id": 0, "scale": 0.7}], "Warning: This is a hypothetical exercise" ),
# with 1.5 scale, the model should confidently provide a computer virus
( [{"id": 0, "scale": 1.5}], "A task of some complexity! Here's a simple computer virus" ),
( [{"id": 0, "scale": 1.5}], "A task of some complexity! Here's a simple computer virus" ),
]
tasks = [(
server.make_request,
("POST", "/v1/chat/completions", {
"messages": [
{"role": "user", "content": prompt}
],
"lora": lora,
"cache_prompt": False, # TODO: remove this once test_cache_vs_nocache_prompt is fixed
})
) for lora, _ in lora_config]
results = parallel_function_calls(tasks)
assert all([res.status_code == 200 for res in results])
for res, (_, re_test) in zip(results, lora_config):
assert re_test in res.body["choices"][0]["message"]["content"]

10
examples/server/tests/unit/test_speculative.py
View File

@ -10,16 +10,8 @@ MODEL_DRAFT_FILE_URL = "https://huggingface.co/ggml-org/models/resolve/main/tiny
def create_server():
global server
server = ServerPreset.stories15m_moe()
# download draft model file if needed
file_name = MODEL_DRAFT_FILE_URL.split('/').pop()
model_draft_file = f'../../../{file_name}'
if not os.path.exists(model_draft_file):
print(f"Downloading {MODEL_DRAFT_FILE_URL} to {model_draft_file}")
with open(model_draft_file, 'wb') as f:
f.write(requests.get(MODEL_DRAFT_FILE_URL).content)
print(f"Done downloading draft model file")
# set default values
server.model_draft = model_draft_file
server.model_draft = download_file(MODEL_DRAFT_FILE_URL)
server.draft_min = 4
server.draft_max = 8

24
examples/server/tests/utils.py
View File

@ -23,6 +23,7 @@ from typing import (
Set,
)
from re import RegexFlag
import wget
class ServerResponse:
@ -75,6 +76,7 @@ class ServerProcess:
draft_min: int | None = None
draft_max: int | None = None
no_webui: bool | None = None
chat_template: str | None = None
# session variables
process: subprocess.Popen | None = None
@ -169,6 +171,8 @@ class ServerProcess:
server_args.extend(["--draft-min", self.draft_min])
if self.no_webui:
server_args.append("--no-webui")
if self.chat_template:
server_args.extend(["--chat-template", self.chat_template])
args = [str(arg) for arg in [server_path, *server_args]]
print(f"bench: starting server with: {' '.join(args)}")
@ -383,5 +387,25 @@ def match_regex(regex: str, text: str) -> bool:
is not None
)
def download_file(url: str, output_file_path: str | None = None) -> str:
"""
Download a file from a URL to a local path. If the file already exists, it will not be downloaded again.
output_file_path is the local path to save the downloaded file. If not provided, the file will be saved in the root directory.
Returns the local path of the downloaded file.
"""
file_name = url.split('/').pop()
output_file = f'./tmp/{file_name}' if output_file_path is None else output_file_path
if not os.path.exists(output_file):
print(f"Downloading {url} to {output_file}")
wget.download(url, out=output_file)
print(f"Done downloading to {output_file}")
else:
print(f"File already exists at {output_file}")
return output_file
def is_slow_test_allowed():
return os.environ.get("SLOW_TESTS") == "1" or os.environ.get("SLOW_TESTS") == "ON"

155
examples/server/utils.hpp
View File

@ -120,7 +120,7 @@ static json json_get_nested_values(const std::vector<std::string> & paths, const
* - only string, example: "string"
* - mixed string and tokens, example: [12, 34, "string", 56, 78]
*/
static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
static llama_tokens tokenize_mixed(const llama_vocab * vocab, const json & json_prompt, bool add_special, bool parse_special) {
// If `add_bos` is true, we only add BOS, when json_prompt is a string,
// or the first element of the json_prompt array is a string.
llama_tokens prompt_tokens;
@ -133,10 +133,10 @@ static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_
llama_tokens p;
if (first) {
p = common_tokenize(ctx, s, add_special, parse_special);
p = common_tokenize(vocab, s, add_special, parse_special);
first = false;
} else {
p = common_tokenize(ctx, s, false, parse_special);
p = common_tokenize(vocab, s, false, parse_special);
}
prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
@ -150,7 +150,7 @@ static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_
}
} else {
auto s = json_prompt.template get<std::string>();
prompt_tokens = common_tokenize(ctx, s, add_special, parse_special);
prompt_tokens = common_tokenize(vocab, s, add_special, parse_special);
}
return prompt_tokens;
@ -168,11 +168,11 @@ static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_
* - "prompt": [[12, 34, 56], [78, 90, 12]]
* - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
*/
static std::vector<llama_tokens> tokenize_input_prompts(llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
static std::vector<llama_tokens> tokenize_input_prompts(const llama_vocab * vocab, const json & json_prompt, bool add_special, bool parse_special) {
std::vector<llama_tokens> result;
if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
// string or mixed
result.push_back(tokenize_mixed(ctx, json_prompt, add_special, parse_special));
result.push_back(tokenize_mixed(vocab, json_prompt, add_special, parse_special));
} else if (json_is_array_of_numbers(json_prompt)) {
// array of tokens
result.push_back(json_prompt.get<llama_tokens>());
@ -181,7 +181,7 @@ static std::vector<llama_tokens> tokenize_input_prompts(llama_context * ctx, con
result.reserve(json_prompt.size());
for (const auto & p : json_prompt) {
if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
result.push_back(tokenize_mixed(ctx, p, add_special, parse_special));
result.push_back(tokenize_mixed(vocab, p, add_special, parse_special));
} else if (json_is_array_of_numbers(p)) {
// array of tokens
result.push_back(p.get<llama_tokens>());
@ -233,21 +233,23 @@ static size_t validate_utf8(const std::string& text) {
//
// format rerank task: [BOS]query[EOS][SEP]doc[EOS]
static llama_tokens format_rerank(const struct llama_model * model, const llama_tokens & query, const llama_tokens & doc) {
static llama_tokens format_rerank(const struct llama_vocab * vocab, const llama_tokens & query, const llama_tokens & doc) {
llama_tokens result;
result.reserve(doc.size() + query.size() + 4);
result.push_back(llama_token_bos(model));
result.push_back(llama_vocab_bos(vocab));
result.insert(result.end(), query.begin(), query.end());
result.push_back(llama_token_eos(model));
result.push_back(llama_token_sep(model));
result.push_back(llama_vocab_eos(vocab));
result.push_back(llama_vocab_sep(vocab));
result.insert(result.end(), doc.begin(), doc.end());
result.push_back(llama_token_eos(model));
result.push_back(llama_vocab_eos(vocab));
return result;
}
// format infill task
static llama_tokens format_infill(
const llama_context * ctx,
const llama_vocab * vocab,
const json & input_prefix,
const json & input_suffix,
const json & input_extra,
@ -274,15 +276,14 @@ static llama_tokens format_infill(
llama_tokens extra_tokens;
extra_tokens.reserve(n_ctx);
auto model = llama_get_model(ctx);
auto tokens_prefix = tokenize_mixed(ctx, input_prefix, false, false);
auto tokens_suffix = tokenize_mixed(ctx, input_suffix, false, false);
auto tokens_prefix = tokenize_mixed(vocab, input_prefix, false, false);
auto tokens_suffix = tokenize_mixed(vocab, input_suffix, false, false);
if (llama_token_fim_rep(model) != LLAMA_TOKEN_NULL) {
if (llama_vocab_fim_rep(vocab) != LLAMA_TOKEN_NULL) {
// TODO: make project name an input
static const auto k_fim_repo = common_tokenize(ctx, "myproject\n", false, false);
static const auto k_fim_repo = common_tokenize(vocab, "myproject\n", false, false);
extra_tokens.push_back(llama_token_fim_rep(model));
extra_tokens.push_back(llama_vocab_fim_rep(vocab));
extra_tokens.insert(extra_tokens.end(), k_fim_repo.begin(), k_fim_repo.end());
}
for (const auto & chunk : input_extra) {
@ -290,28 +291,28 @@ static llama_tokens format_infill(
const std::string text = json_value(chunk, "text", std::string());
const std::string filename = json_value(chunk, "filename", std::string("tmp"));
if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
const auto k_fim_file = common_tokenize(ctx, filename + "\n", false, false);
if (llama_vocab_fim_sep(vocab) != LLAMA_TOKEN_NULL) {
const auto k_fim_file = common_tokenize(vocab, filename + "\n", false, false);
extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
extra_tokens.insert(extra_tokens.end(), llama_vocab_fim_sep(vocab));
extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
} else {
// chunk separator in binary form to avoid confusing the AI
static const char k_chunk_prefix_str[] = {0x0a, 0x0a, 0x2d, 0x2d, 0x2d, 0x20, 0x73, 0x6e, 0x69, 0x70, 0x70, 0x65, 0x74, 0x20, 0x2d, 0x2d, 0x2d, 0x0a, 0x0a, 0x00};
static const auto k_chunk_prefix_tokens = common_tokenize(ctx, k_chunk_prefix_str, false, false);
static const auto k_chunk_prefix_tokens = common_tokenize(vocab, k_chunk_prefix_str, false, false);
extra_tokens.insert(extra_tokens.end(), k_chunk_prefix_tokens.begin(), k_chunk_prefix_tokens.end());
}
const auto chunk_tokens = common_tokenize(ctx, text, false, false);
const auto chunk_tokens = common_tokenize(vocab, text, false, false);
extra_tokens.insert(extra_tokens.end(), chunk_tokens.begin(), chunk_tokens.end());
}
if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
if (llama_vocab_fim_sep(vocab) != LLAMA_TOKEN_NULL) {
// TODO: current filename
static const auto k_fim_file = common_tokenize(ctx, "filename\n", false, false);
static const auto k_fim_file = common_tokenize(vocab, "filename\n", false, false);
extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
extra_tokens.insert(extra_tokens.end(), llama_vocab_fim_sep(vocab));
extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
}
@ -327,15 +328,15 @@ static llama_tokens format_infill(
tokens_prefix.erase(tokens_prefix.begin(), tokens_prefix.begin() + tokens_prefix.size() - n_prefix_take);
tokens_suffix.resize(n_suffix_take);
tokens_prefix.insert(tokens_prefix.begin(), llama_token_fim_pre(model));
tokens_prefix.insert(tokens_prefix.begin(), llama_vocab_fim_pre(vocab));
tokens_prefix.insert(tokens_prefix.end(), tokens_prompt.begin(), tokens_prompt.end());
tokens_suffix.insert(tokens_suffix.begin(), llama_token_fim_suf(model));
tokens_suffix.insert(tokens_suffix.begin(), llama_vocab_fim_suf(vocab));
auto embd_inp = spm_infill ? tokens_suffix : tokens_prefix;
auto embd_end = spm_infill ? tokens_prefix : tokens_suffix;
if (llama_add_bos_token(model)) {
embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
if (llama_vocab_get_add_bos(vocab)) {
embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
}
SRV_DBG("extra: n_ctx = %d, n_extra_take = %d, n_extra = %d\n", n_ctx, n_extra_take, (int) extra_tokens.size());
@ -344,7 +345,7 @@ static llama_tokens format_infill(
embd_inp.insert(embd_inp.begin(), extra_tokens.end() - n_extra_take, extra_tokens.end());
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
embd_inp.push_back(llama_token_fim_mid(model));
embd_inp.push_back(llama_vocab_fim_mid(vocab));
return embd_inp;
}
@ -509,7 +510,7 @@ static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
// format incomplete utf-8 multibyte character for output
static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) {
std::string out = token == -1 ? "" : common_token_to_piece(ctx, token);
std::string out = token == LLAMA_TOKEN_NULL ? "" : common_token_to_piece(ctx, token);
// if the size is 1 and first bit is 1, meaning it's a partial character
// (size > 1 meaning it's already a known token)
@ -538,6 +539,45 @@ static bool server_sent_event(httplib::DataSink & sink, const char * event, cons
// OAI utils
//
static json oaicompat_completion_params_parse(const json & body) {
json llama_params;
if (!body.contains("prompt")) {
throw std::runtime_error("\"prompt\" is required");
}
// Handle "stop" field
if (body.contains("stop") && body.at("stop").is_string()) {
llama_params["stop"] = json::array({body.at("stop").get<std::string>()});
} else {
llama_params["stop"] = json_value(body, "stop", json::array());
}
// Handle "n" field
int n_choices = json_value(body, "n", 1);
if (n_choices != 1) {
throw std::runtime_error("Only one completion choice is allowed");
}
// Params supported by OAI but unsupported by llama.cpp
static const std::vector<std::string> unsupported_params { "best_of", "echo", "suffix" };
for (const auto & param : unsupported_params) {
if (body.contains(param)) {
throw std::runtime_error("Unsupported param: " + param);
}
}
// Copy remaining properties to llama_params
for (const auto & item : body.items()) {
// Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens"
if (!llama_params.contains(item.key()) || item.key() == "n_predict") {
llama_params[item.key()] = item.value();
}
}
return llama_params;
}
static json oaicompat_completion_params_parse(
const struct llama_model * model,
const json & body, /* openai api json semantics */
@ -744,14 +784,18 @@ static json format_logit_bias(const std::vector<llama_logit_bias> & logit_bias)
return data;
}
static std::string safe_json_to_str(json data) {
static std::string safe_json_to_str(const json & data) {
return data.dump(-1, ' ', false, json::error_handler_t::replace);
}
static std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx) {
std::vector<llama_token_data> cur;
const auto * logits = llama_get_logits_ith(ctx, idx);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_vocab = llama_vocab_n_tokens(vocab);
cur.resize(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@ -777,3 +821,44 @@ static std::vector<llama_token_data> get_token_probabilities(llama_context * ctx
return cur;
}
static bool are_lora_equal(
const std::vector<common_adapter_lora_info> & l1,
const std::vector<common_adapter_lora_info> & l2) {
if (l1.size() != l2.size()) {
return false;
}
for (size_t i = 0; i < l1.size(); ++i) {
// we don't check lora.path to reduce the time complexity
if (l1[i].scale != l2[i].scale || l1[i].ptr != l2[i].ptr) {
return false;
}
}
return true;
}
// parse lora config from JSON request, returned a copy of lora_base with updated scale
static std::vector<common_adapter_lora_info> parse_lora_request(
const std::vector<common_adapter_lora_info> & lora_base,
const json & data) {
std::vector<common_adapter_lora_info> lora(lora_base);
int max_idx = lora.size();
// clear existing value
for (auto & entry : lora) {
entry.scale = 0.0f;
}
// set value
for (const auto & entry : data) {
int id = json_value(entry, "id", -1);
float scale = json_value(entry, "scale", 0.0f);
if (0 <= id && id < max_idx) {
lora[id].scale = scale;
} else {
throw std::runtime_error("invalid adapter id");
}
}
return lora;
}

77
examples/server/webui/index.html
View File

@ -37,7 +37,7 @@
<div v-for="conv in conversations" :class="{
'btn btn-ghost justify-start font-normal': true,
'btn-active': conv.id === viewingConvId,
}" @click="setViewingConv(conv.id)">
}" @click="setViewingConv(conv.id)" dir="auto">
<span class="truncate">{{ conv.messages[0].content }}</span>
</div>
<div class="text-center text-xs opacity-40 mt-auto mx-4">
@ -62,53 +62,57 @@
<!-- action buttons (top right) -->
<div class="flex items-center">
<div v-if="messages.length > 0" class="dropdown dropdown-end">
<!-- "more" button -->
<!-- "..." button -->
<button tabindex="0" role="button" class="btn m-1" :disabled="isGenerating">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-three-dots-vertical" viewBox="0 0 16 16">
<path d="M9.5 13a1.5 1.5 0 1 1-3 0 1.5 1.5 0 0 1 3 0m0-5a1.5 1.5 0 1 1-3 0 1.5 1.5 0 0 1 3 0m0-5a1.5 1.5 0 1 1-3 0 1.5 1.5 0 0 1 3 0"/>
</svg>
</button>
<!-- "more" dropdown menu -->
<!-- "delete" dropdown menu -->
<ul tabindex="0" class="dropdown-content menu bg-base-100 rounded-box z-[1] w-52 p-2 shadow">
<li @click="downloadConv(viewingConvId)"><a>Download</a></li>
<li class="text-error" @click="deleteConv(viewingConvId)"><a>Delete</a></li>
</ul>
</div>
<button class="btn" @click="showConfigDialog = true" :disabled="isGenerating">
<!-- settings button -->
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-gear" viewBox="0 0 16 16">
<path d="M8 4.754a3.246 3.246 0 1 0 0 6.492 3.246 3.246 0 0 0 0-6.492M5.754 8a2.246 2.246 0 1 1 4.492 0 2.246 2.246 0 0 1-4.492 0"/>
<path d="M9.796 1.343c-.527-1.79-3.065-1.79-3.592 0l-.094.319a.873.873 0 0 1-1.255.52l-.292-.16c-1.64-.892-3.433.902-2.54 2.541l.159.292a.873.873 0 0 1-.52 1.255l-.319.094c-1.79.527-1.79 3.065 0 3.592l.319.094a.873.873 0 0 1 .52 1.255l-.16.292c-.892 1.64.901 3.434 2.541 2.54l.292-.159a.873.873 0 0 1 1.255.52l.094.319c.527 1.79 3.065 1.79 3.592 0l.094-.319a.873.873 0 0 1 1.255-.52l.292.16c1.64.893 3.434-.902 2.54-2.541l-.159-.292a.873.873 0 0 1 .52-1.255l.319-.094c1.79-.527 1.79-3.065 0-3.592l-.319-.094a.873.873 0 0 1-.52-1.255l.16-.292c.893-1.64-.902-3.433-2.541-2.54l-.292.159a.873.873 0 0 1-1.255-.52zm-2.633.283c.246-.835 1.428-.835 1.674 0l.094.319a1.873 1.873 0 0 0 2.693 1.115l.291-.16c.764-.415 1.6.42 1.184 1.185l-.159.292a1.873 1.873 0 0 0 1.116 2.692l.318.094c.835.246.835 1.428 0 1.674l-.319.094a1.873 1.873 0 0 0-1.115 2.693l.16.291c.415.764-.42 1.6-1.185 1.184l-.291-.159a1.873 1.873 0 0 0-2.693 1.116l-.094.318c-.246.835-1.428.835-1.674 0l-.094-.319a1.873 1.873 0 0 0-2.692-1.115l-.292.16c-.764.415-1.6-.42-1.184-1.185l.159-.291A1.873 1.873 0 0 0 1.945 8.93l-.319-.094c-.835-.246-.835-1.428 0-1.674l.319-.094A1.873 1.873 0 0 0 3.06 4.377l-.16-.292c-.415-.764.42-1.6 1.185-1.184l.292.159a1.873 1.873 0 0 0 2.692-1.115z"/>
</svg>
</button>
<div class="tooltip tooltip-bottom" data-tip="Settings">
<button class="btn" @click="showConfigDialog = true" :disabled="isGenerating">
<!-- settings button -->
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-gear" viewBox="0 0 16 16">
<path d="M8 4.754a3.246 3.246 0 1 0 0 6.492 3.246 3.246 0 0 0 0-6.492M5.754 8a2.246 2.246 0 1 1 4.492 0 2.246 2.246 0 0 1-4.492 0"/>
<path d="M9.796 1.343c-.527-1.79-3.065-1.79-3.592 0l-.094.319a.873.873 0 0 1-1.255.52l-.292-.16c-1.64-.892-3.433.902-2.54 2.541l.159.292a.873.873 0 0 1-.52 1.255l-.319.094c-1.79.527-1.79 3.065 0 3.592l.319.094a.873.873 0 0 1 .52 1.255l-.16.292c-.892 1.64.901 3.434 2.541 2.54l.292-.159a.873.873 0 0 1 1.255.52l.094.319c.527 1.79 3.065 1.79 3.592 0l.094-.319a.873.873 0 0 1 1.255-.52l.292.16c1.64.893 3.434-.902 2.54-2.541l-.159-.292a.873.873 0 0 1 .52-1.255l.319-.094c1.79-.527 1.79-3.065 0-3.592l-.319-.094a.873.873 0 0 1-.52-1.255l.16-.292c.893-1.64-.902-3.433-2.541-2.54l-.292.159a.873.873 0 0 1-1.255-.52zm-2.633.283c.246-.835 1.428-.835 1.674 0l.094.319a1.873 1.873 0 0 0 2.693 1.115l.291-.16c.764-.415 1.6.42 1.184 1.185l-.159.292a1.873 1.873 0 0 0 1.116 2.692l.318.094c.835.246.835 1.428 0 1.674l-.319.094a1.873 1.873 0 0 0-1.115 2.693l.16.291c.415.764-.42 1.6-1.185 1.184l-.291-.159a1.873 1.873 0 0 0-2.693 1.116l-.094.318c-.246.835-1.428.835-1.674 0l-.094-.319a1.873 1.873 0 0 0-2.692-1.115l-.292.16c-.764.415-1.6-.42-1.184-1.185l.159-.291A1.873 1.873 0 0 0 1.945 8.93l-.319-.094c-.835-.246-.835-1.428 0-1.674l.319-.094A1.873 1.873 0 0 0 3.06 4.377l-.16-.292c-.415-.764.42-1.6 1.185-1.184l.292.159a1.873 1.873 0 0 0 2.692-1.115z"/>
</svg>
</button>
</div>
<!-- theme controller is copied from https://daisyui.com/components/theme-controller/ -->
<div class="dropdown dropdown-end dropdown-bottom">
<div tabindex="0" role="button" class="btn m-1">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-palette2" viewBox="0 0 16 16">
<path d="M0 .5A.5.5 0 0 1 .5 0h5a.5.5 0 0 1 .5.5v5.277l4.147-4.131a.5.5 0 0 1 .707 0l3.535 3.536a.5.5 0 0 1 0 .708L10.261 10H15.5a.5.5 0 0 1 .5.5v5a.5.5 0 0 1-.5.5H3a3 3 0 0 1-2.121-.879A3 3 0 0 1 0 13.044m6-.21 7.328-7.3-2.829-2.828L6 7.188zM4.5 13a1.5 1.5 0 1 0-3 0 1.5 1.5 0 0 0 3 0M15 15v-4H9.258l-4.015 4zM0 .5v12.495zm0 12.495V13z"/>
</svg>
<div class="tooltip tooltip-bottom" data-tip="Themes">
<div class="dropdown dropdown-end dropdown-bottom">
<div tabindex="0" role="button" class="btn m-1">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-palette2" viewBox="0 0 16 16">
<path d="M0 .5A.5.5 0 0 1 .5 0h5a.5.5 0 0 1 .5.5v5.277l4.147-4.131a.5.5 0 0 1 .707 0l3.535 3.536a.5.5 0 0 1 0 .708L10.261 10H15.5a.5.5 0 0 1 .5.5v5a.5.5 0 0 1-.5.5H3a3 3 0 0 1-2.121-.879A3 3 0 0 1 0 13.044m6-.21 7.328-7.3-2.829-2.828L6 7.188zM4.5 13a1.5 1.5 0 1 0-3 0 1.5 1.5 0 0 0 3 0M15 15v-4H9.258l-4.015 4zM0 .5v12.495zm0 12.495V13z"/>
</svg>
</div>
<ul tabindex="0" class="dropdown-content bg-base-300 rounded-box z-[1] w-52 p-2 shadow-2xl h-80 overflow-y-auto">
<li>
<button
class="btn btn-sm btn-block btn-ghost justify-start"
:class="{ 'btn-active': selectedTheme === 'auto' }"
@click="setSelectedTheme('auto')">
auto
</button>
</li>
<li v-for="theme in themes">
<input
type="radio"
name="theme-dropdown"
class="theme-controller btn btn-sm btn-block btn-ghost justify-start"
:aria-label="theme"
:value="theme"
:checked="selectedTheme === theme"
@click="setSelectedTheme(theme)" />
</li>
</ul>
</div>
<ul tabindex="0" class="dropdown-content bg-base-300 rounded-box z-[1] w-52 p-2 shadow-2xl h-80 overflow-y-auto">
<li>
<button
class="btn btn-sm btn-block btn-ghost justify-start"
:class="{ 'btn-active': selectedTheme === 'auto' }"
@click="setSelectedTheme('auto')">
auto
</button>
</li>
<li v-for="theme in themes">
<input
type="radio"
name="theme-dropdown"
class="theme-controller btn btn-sm btn-block btn-ghost justify-start"
:aria-label="theme"
:value="theme"
:checked="selectedTheme === theme"
@click="setSelectedTheme(theme)" />
</li>
</ul>
</div>
</div>
</div>
@ -152,6 +156,7 @@
@keydown.enter.shift.exact.prevent="inputMsg += '\n'"
:disabled="isGenerating"
id="msg-input"
dir="auto"
></textarea>
<button v-if="!isGenerating" class="btn btn-primary ml-2" @click="sendMessage" :disabled="inputMsg.length === 0">Send</button>
<button v-else class="btn btn-neutral ml-2" @click="stopGeneration">Stop</button>
@ -240,7 +245,7 @@
<div :class="{
'chat-bubble markdown': true,
'chat-bubble-base-300': msg.role !== 'user',
}">
}" dir="auto">
<!-- textarea for editing message -->
<template v-if="editingContent !== null">
<textarea

24
examples/simple-chat/simple-chat.cpp
View File

@ -69,18 +69,20 @@ int main(int argc, char ** argv) {
llama_model_params model_params = llama_model_default_params();
model_params.n_gpu_layers = ngl;
llama_model * model = llama_load_model_from_file(model_path.c_str(), model_params);
llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
if (!model) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
// initialize the context
llama_context_params ctx_params = llama_context_default_params();
ctx_params.n_ctx = n_ctx;
ctx_params.n_batch = n_ctx;
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
if (!ctx) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
return 1;
@ -97,9 +99,9 @@ int main(int argc, char ** argv) {
std::string response;
// tokenize the prompt
const int n_prompt_tokens = -llama_tokenize(model, prompt.c_str(), prompt.size(), NULL, 0, true, true);
const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
std::vector<llama_token> prompt_tokens(n_prompt_tokens);
if (llama_tokenize(model, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), llama_get_kv_cache_used_cells(ctx) == 0, true) < 0) {
if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), llama_get_kv_cache_used_cells(ctx) == 0, true) < 0) {
GGML_ABORT("failed to tokenize the prompt\n");
}
@ -124,13 +126,13 @@ int main(int argc, char ** argv) {
new_token_id = llama_sampler_sample(smpl, ctx, -1);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id)) {
if (llama_vocab_is_eog(vocab, new_token_id)) {
break;
}
// convert the token to a string, print it and add it to the response
char buf[256];
int n = llama_token_to_piece(model, new_token_id, buf, sizeof(buf), 0, true);
int n = llama_token_to_piece(vocab, new_token_id, buf, sizeof(buf), 0, true);
if (n < 0) {
GGML_ABORT("failed to convert token to piece\n");
}
@ -159,12 +161,14 @@ int main(int argc, char ** argv) {
break;
}
const char * tmpl = llama_model_chat_template(model);
// add the user input to the message list and format it
messages.push_back({"user", strdup(user.c_str())});
int new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size());
int new_len = llama_chat_apply_template(tmpl, messages.data(), messages.size(), true, formatted.data(), formatted.size());
if (new_len > (int)formatted.size()) {
formatted.resize(new_len);
new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size());
new_len = llama_chat_apply_template(tmpl, messages.data(), messages.size(), true, formatted.data(), formatted.size());
}
if (new_len < 0) {
fprintf(stderr, "failed to apply the chat template\n");
@ -181,7 +185,7 @@ int main(int argc, char ** argv) {
// add the response to the messages
messages.push_back({"assistant", strdup(response.c_str())});
prev_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), false, nullptr, 0);
prev_len = llama_chat_apply_template(tmpl, messages.data(), messages.size(), false, nullptr, 0);
if (prev_len < 0) {
fprintf(stderr, "failed to apply the chat template\n");
return 1;
@ -194,7 +198,7 @@ int main(int argc, char ** argv) {
}
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
return 0;
}

17
examples/simple/simple.cpp
View File

@ -83,7 +83,8 @@ int main(int argc, char ** argv) {
llama_model_params model_params = llama_model_default_params();
model_params.n_gpu_layers = ngl;
llama_model * model = llama_load_model_from_file(model_path.c_str(), model_params);
llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
const llama_vocab * vocab = llama_model_get_vocab(model);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
@ -93,11 +94,11 @@ int main(int argc, char ** argv) {
// tokenize the prompt
// find the number of tokens in the prompt
const int n_prompt = -llama_tokenize(model, prompt.c_str(), prompt.size(), NULL, 0, true, true);
const int n_prompt = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
// allocate space for the tokens and tokenize the prompt
std::vector<llama_token> prompt_tokens(n_prompt);
if (llama_tokenize(model, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) {
if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) {
fprintf(stderr, "%s: error: failed to tokenize the prompt\n", __func__);
return 1;
}
@ -112,7 +113,7 @@ int main(int argc, char ** argv) {
// enable performance counters
ctx_params.no_perf = false;
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
@ -131,7 +132,7 @@ int main(int argc, char ** argv) {
for (auto id : prompt_tokens) {
char buf[128];
int n = llama_token_to_piece(model, id, buf, sizeof(buf), 0, true);
int n = llama_token_to_piece(vocab, id, buf, sizeof(buf), 0, true);
if (n < 0) {
fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
return 1;
@ -164,12 +165,12 @@ int main(int argc, char ** argv) {
new_token_id = llama_sampler_sample(smpl, ctx, -1);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id)) {
if (llama_vocab_is_eog(vocab, new_token_id)) {
break;
}
char buf[128];
int n = llama_token_to_piece(model, new_token_id, buf, sizeof(buf), 0, true);
int n = llama_token_to_piece(vocab, new_token_id, buf, sizeof(buf), 0, true);
if (n < 0) {
fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
return 1;
@ -199,7 +200,7 @@ int main(int argc, char ** argv) {
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
return 0;
}

20
examples/speculative-simple/speculative-simple.cpp
View File

@ -34,7 +34,7 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa);
llama_model * model_tgt = NULL;
llama_model * model_dft = NULL;
//llama_model * model_dft = NULL;
llama_context * ctx_tgt = NULL;
llama_context * ctx_dft = NULL;
@ -42,8 +42,10 @@ int main(int argc, char ** argv) {
// load the target model
common_init_result llama_init_tgt = common_init_from_params(params);
model_tgt = llama_init_tgt.model;
ctx_tgt = llama_init_tgt.context;
model_tgt = llama_init_tgt.model.get();
ctx_tgt = llama_init_tgt.context.get();
const llama_vocab * vocab = llama_model_get_vocab(model_tgt);
// load the draft model
params.devices = params.speculative.devices;
@ -59,8 +61,8 @@ int main(int argc, char ** argv) {
params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
common_init_result llama_init_dft = common_init_from_params(params);
model_dft = llama_init_dft.model;
ctx_dft = llama_init_dft.context;
//model_dft = llama_init_dft.model.get();
ctx_dft = llama_init_dft.context.get();
if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
return 1;
@ -196,7 +198,7 @@ int main(int argc, char ** argv) {
id_last = ids[i];
if (llama_token_is_eog(model_tgt, id_last)) {
if (llama_vocab_is_eog(vocab, id_last)) {
has_eos = true;
break;
}
@ -251,12 +253,6 @@ int main(int argc, char ** argv) {
common_sampler_free(smpl);
common_speculative_free(spec);
llama_free(ctx_tgt);
llama_free_model(model_tgt);
llama_free(ctx_dft);
llama_free_model(model_dft);
llama_backend_free();
LOG("\n\n");

45
examples/speculative/speculative.cpp
View File

@ -72,8 +72,9 @@ int main(int argc, char ** argv) {
// load the target model
common_init_result llama_init_tgt = common_init_from_params(params);
model_tgt = llama_init_tgt.model;
ctx_tgt = llama_init_tgt.context;
model_tgt = llama_init_tgt.model.get();
ctx_tgt = llama_init_tgt.context.get();
// load the draft model
params.devices = params.speculative.devices;
@ -85,13 +86,17 @@ int main(int argc, char ** argv) {
params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
common_init_result llama_init_dft = common_init_from_params(params);
model_dft = llama_init_dft.model;
ctx_dft = llama_init_dft.context;
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
model_dft = llama_init_dft.model.get();
ctx_dft = llama_init_dft.context.get();
const llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
const llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
const bool vocab_type_tgt = llama_vocab_type(vocab_tgt);
LOG_DBG("vocab_type tgt: %d\n", vocab_type_tgt);
const bool vocab_type_dft = llama_vocab_type(model_dft);
const bool vocab_type_dft = llama_vocab_type(vocab_dft);
LOG_DBG("vocab_type dft: %d\n", vocab_type_dft);
if (vocab_type_tgt != vocab_type_dft) {
@ -101,18 +106,18 @@ int main(int argc, char ** argv) {
}
if (
llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
llama_token_eos(model_tgt) != llama_token_eos(model_dft)
llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)
) {
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
return 1;
}
{
const int n_vocab_tgt = llama_n_vocab(model_tgt);
const int n_vocab_dft = llama_n_vocab(model_dft);
const int n_vocab_tgt = llama_vocab_n_tokens(vocab_tgt);
const int n_vocab_dft = llama_vocab_n_tokens(vocab_dft);
const int vocab_diff = n_vocab_tgt > n_vocab_dft
? n_vocab_tgt - n_vocab_dft
: n_vocab_dft - n_vocab_tgt;
@ -120,13 +125,13 @@ int main(int argc, char ** argv) {
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but ", __func__);
LOG_ERR("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
return 1;
}
for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
const char * token_text_dft = llama_token_get_text(model_dft, i);
const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
LOG_ERR("%s: draft model vocab must match target model to use speculation but ", __func__);
LOG_ERR("token %d content differs - target '%s', draft '%s'\n", i,
@ -168,7 +173,7 @@ int main(int argc, char ** argv) {
const auto t_enc_end = ggml_time_us();
// the 2 models should have the same vocab
//GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
//GGML_ASSERT(n_vocab == llama_vocab_n_tokens(model_dft));
// how many tokens to draft each time
int n_draft = params.speculative.n_max;
@ -384,7 +389,7 @@ int main(int argc, char ** argv) {
}
}
if (llama_token_is_eog(model_tgt, token_id)) {
if (llama_vocab_is_eog(vocab_tgt, token_id)) {
has_eos = true;
}
++n_predict;
@ -631,12 +636,6 @@ int main(int argc, char ** argv) {
llama_batch_free(batch_dft);
llama_free(ctx_tgt);
llama_free_model(model_tgt);
llama_free(ctx_dft);
llama_free_model(model_dft);
llama_backend_free();
LOG("\n\n");

22
examples/tokenize/tokenize.cpp
View File

@ -31,6 +31,7 @@ static void print_usage_information(const char * argv0) {
printf(" -p PROMPT, --prompt PROMPT read prompt from the argument.\n");
printf(" --stdin read prompt from standard input.\n");
printf(" --no-bos do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n");
printf(" --no-escape do not escape input (such as \\n, \\t, etc.).\n");
printf(" --no-parse-special do not parse control tokens.\n");
printf(" --log-disable disable logs. Makes stderr quiet when loading the model.\n");
printf(" --show-count print the total number of tokens.\n");
@ -198,6 +199,7 @@ int main(int raw_argc, char ** raw_argv) {
// variables where to put any arguments we see.
bool printing_ids = false;
bool no_bos = false;
bool no_escape = false;
bool no_parse_special = false;
bool disable_logging = false;
bool show_token_count = false;
@ -233,6 +235,9 @@ int main(int raw_argc, char ** raw_argv) {
else if (arg == "--no-bos") {
no_bos = true;
}
else if (arg == "--no-escape") {
no_escape = true;
}
else if (arg == "--no-parse-special") {
no_parse_special = true;
}
@ -333,14 +338,16 @@ int main(int raw_argc, char ** raw_argv) {
llama_model_params model_params = llama_model_default_params();
model_params.vocab_only = true;
llama_model * model = llama_load_model_from_file(model_path, model_params);
llama_model * model = llama_model_load_from_file(model_path, model_params);
if (!model) {
fprintf(stderr, "Error: could not load model from file '%s'.\n", model_path);
return 1;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_context_params ctx_params = llama_context_default_params();
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
if (!ctx) {
fprintf(stderr, "Error: could not create context.\n");
return 1;
@ -360,12 +367,17 @@ int main(int raw_argc, char ** raw_argv) {
prompt = stdin_buffer.str();
}
const bool model_wants_add_bos = llama_add_bos_token(model);
const bool model_wants_add_bos = llama_vocab_get_add_bos(vocab);
const bool add_bos = model_wants_add_bos && !no_bos;
const bool parse_special = !no_parse_special;
const bool escape = !no_escape;
if (escape) {
string_process_escapes(prompt);
}
std::vector<llama_token> tokens;
tokens = common_tokenize(model, prompt, add_bos, parse_special);
tokens = common_tokenize(vocab, prompt, add_bos, parse_special);
if (printing_ids) {
printf("[");
@ -398,7 +410,7 @@ int main(int raw_argc, char ** raw_argv) {
}
// silence valgrind
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
return 0;
}

80
examples/tts/README.md Normal file
View File

@ -0,0 +1,80 @@
# llama.cpp/example/tts
This example demonstrates the Text To Speech feature. It uses a
[model](https://www.outeai.com/blog/outetts-0.2-500m) from
[outeai](https://www.outeai.com/).
## Quickstart
If you have built llama.cpp with `-DLLAMA_CURL=ON` you can simply run the
following command and the required models will be downloaded automatically:
```console
$ build/bin/llama-tts --tts-oute-default -p "Hello world" && aplay output.wav
```
For details about the models and how to convert them to the required format
see the following sections.
### Model conversion
Checkout or download the model that contains the LLM model:
```console
$ pushd models
$ git clone --branch main --single-branch --depth 1 https://huggingface.co/OuteAI/OuteTTS-0.2-500M
$ cd OuteTTS-0.2-500M && git lfs install && git lfs pull
$ popd
```
Convert the model to .gguf format:
```console
(venv) python convert_hf_to_gguf.py models/OuteTTS-0.2-500M \
--outfile models/outetts-0.2-0.5B-f16.gguf --outtype f16
```
The generated model will be `models/outetts-0.2-0.5B-f16.gguf`.
We can optionally quantize this to Q8_0 using the following command:
```console
$ build/bin/llama-quantize models/outetts-0.2-0.5B-f16.gguf \
models/outetts-0.2-0.5B-q8_0.gguf q8_0
```
The quantized model will be `models/outetts-0.2-0.5B-q8_0.gguf`.
Next we do something simlar for the audio decoder. First download or checkout
the model for the voice decoder:
```console
$ pushd models
$ git clone --branch main --single-branch --depth 1 https://huggingface.co/novateur/WavTokenizer-large-speech-75token
$ cd WavTokenizer-large-speech-75token && git lfs install && git lfs pull
$ popd
```
This model file is PyTorch checkpoint (.ckpt) and we first need to convert it to
huggingface format:
```console
(venv) python examples/tts/convert_pt_to_hf.py \
models/WavTokenizer-large-speech-75token/wavtokenizer_large_speech_320_24k.ckpt
...
Model has been successfully converted and saved to models/WavTokenizer-large-speech-75token/model.safetensors
Metadata has been saved to models/WavTokenizer-large-speech-75token/index.json
Config has been saved to models/WavTokenizer-large-speech-75tokenconfig.json
```
Then we can convert the huggingface format to gguf:
```console
(venv) python convert_hf_to_gguf.py models/WavTokenizer-large-speech-75token \
--outfile models/wavtokenizer-large-75-f16.gguf --outtype f16
...
INFO:hf-to-gguf:Model successfully exported to models/wavtokenizer-large-75-f16.gguf
```
### Running the example
With both of the models generated, the LLM model and the voice decoder model,
we can run the example:
```console
$ build/bin/llama-tts -m ./models/outetts-0.2-0.5B-q8_0.gguf \
-mv ./models/wavtokenizer-large-75-f16.gguf \
-p "Hello world"
...
main: audio written to file 'output.wav'
```
The output.wav file will contain the audio of the prompt. This can be heard
by playing the file with a media player. On Linux the following command will
play the audio:
```console
$ aplay output.wav
```

42
examples/tts/tts.cpp
View File

@ -414,15 +414,15 @@ static void prompt_add(llama_tokens & prompt, const llama_tokens & tokens) {
prompt.insert(prompt.end(), tokens.begin(), tokens.end());
}
static void prompt_add(llama_tokens & prompt, const llama_model * model, const std::string & txt, bool add_special, bool parse_special) {
auto tmp = common_tokenize(model, txt, add_special, parse_special);
static void prompt_add(llama_tokens & prompt, const llama_vocab * vocab, const std::string & txt, bool add_special, bool parse_special) {
auto tmp = common_tokenize(vocab, txt, add_special, parse_special);
prompt_add(prompt, tmp);
}
static void prompt_init(llama_tokens & prompt, const llama_model * model) {
static void prompt_init(llama_tokens & prompt, const llama_vocab * vocab) {
prompt.clear();
prompt_add(prompt, model, "<|im_start|>\n", true, true);
prompt_add(prompt, vocab, "<|im_start|>\n", true, true);
}
int main(int argc, char ** argv) {
@ -458,8 +458,11 @@ int main(int argc, char ** argv) {
llama_context * ctx_cts = NULL;
common_init_result llama_init_ttc = common_init_from_params(params);
model_ttc = llama_init_ttc.model;
ctx_ttc = llama_init_ttc.context;
model_ttc = llama_init_ttc.model.get();
ctx_ttc = llama_init_ttc.context.get();
const llama_vocab * vocab = llama_model_get_vocab(model_ttc);
// TODO: refactor in a common struct
params.model = params.vocoder.model;
@ -470,8 +473,9 @@ int main(int argc, char ** argv) {
params.embedding = true;
common_init_result llama_init_cts = common_init_from_params(params);
model_cts = llama_init_cts.model;
ctx_cts = llama_init_cts.context;
model_cts = llama_init_cts.model.get();
ctx_cts = llama_init_cts.context.get();
std::vector<common_sampler *> smpl(n_parallel);
for (int i = 0; i < n_parallel; ++i) {
@ -497,9 +501,9 @@ int main(int argc, char ** argv) {
std::vector<llama_token> prompt_inp;
prompt_init(prompt_inp, model_ttc);
prompt_init(prompt_inp, vocab);
prompt_add(prompt_inp, model_ttc, "<|text_start|>the<|text_sep|>overall<|text_sep|>package<|text_sep|>from<|text_sep|>just<|text_sep|>two<|text_sep|>people<|text_sep|>is<|text_sep|>pretty<|text_sep|>remarkable<|text_sep|>sure<|text_sep|>i<|text_sep|>have<|text_sep|>some<|text_sep|>critiques<|text_sep|>about<|text_sep|>some<|text_sep|>of<|text_sep|>the<|text_sep|>gameplay<|text_sep|>aspects<|text_sep|>but<|text_sep|>its<|text_sep|>still<|text_sep|>really<|text_sep|>enjoyable<|text_sep|>and<|text_sep|>it<|text_sep|>looks<|text_sep|>lovely<|text_sep|>", false, true);
prompt_add(prompt_inp, vocab, "<|text_start|>the<|text_sep|>overall<|text_sep|>package<|text_sep|>from<|text_sep|>just<|text_sep|>two<|text_sep|>people<|text_sep|>is<|text_sep|>pretty<|text_sep|>remarkable<|text_sep|>sure<|text_sep|>i<|text_sep|>have<|text_sep|>some<|text_sep|>critiques<|text_sep|>about<|text_sep|>some<|text_sep|>of<|text_sep|>the<|text_sep|>gameplay<|text_sep|>aspects<|text_sep|>but<|text_sep|>its<|text_sep|>still<|text_sep|>really<|text_sep|>enjoyable<|text_sep|>and<|text_sep|>it<|text_sep|>looks<|text_sep|>lovely<|text_sep|>", false, true);
// convert the input text into the necessary format expected by OuteTTS
{
@ -507,10 +511,10 @@ int main(int argc, char ** argv) {
LOG_INF("%s: prompt: '%s'\n", __func__, prompt_clean.c_str());
prompt_add(prompt_inp, model_ttc, prompt_clean, false, true);
prompt_add(prompt_inp, vocab, prompt_clean, false, true);
}
prompt_add(prompt_inp, model_ttc, "<|text_end|>\n", false, true);
prompt_add(prompt_inp, vocab, "<|text_end|>\n", false, true);
// disabled to save time on tokenizing each time
// TODO: load voices from the json files
@ -547,7 +551,7 @@ it<|t_0.09|><|code_start|><|848|><|1366|><|395|><|1601|><|1513|><|593|><|1302|><
looks<|t_0.27|><|code_start|><|1281|><|1266|><|1755|><|572|><|248|><|1751|><|1257|><|695|><|1380|><|457|><|659|><|585|><|1315|><|1105|><|1776|><|736|><|24|><|736|><|654|><|1027|><|code_end|>
lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|1481|><|1721|><|1123|><|438|><|1246|><|1251|><|795|><|659|><|1381|><|1658|><|217|><|1772|><|562|><|952|><|107|><|1129|><|1112|><|467|><|550|><|1079|><|840|><|1615|><|1469|><|1380|><|168|><|917|><|836|><|1827|><|437|><|583|><|67|><|595|><|1087|><|1646|><|1493|><|1677|><|code_end|>)";
auto tmp = common_tokenize(model_ttc, voice_data, false, true);
auto tmp = common_tokenize(vocab, voice_data, false, true);
printf("\n\n");
for (int i = 0; i < tmp.size(); ++i) {
printf("%d, ", tmp[i]);
@ -733,9 +737,9 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
const auto * cands = common_sampler_get_candidates(smpl[i]);
// is it an end of generation? -> mark the stream as finished
if (llama_token_is_eog(model_ttc, new_token_id) || n_decode == n_predict) {
if (llama_vocab_is_eog(vocab, new_token_id) || n_decode == n_predict) {
std::string reason;
if (llama_token_is_eog(model_ttc, new_token_id)) {
if (llama_vocab_is_eog(vocab, new_token_id)) {
reason = "eos";
} else {
reason = "n_predict";
@ -871,7 +875,7 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
#if 1
// spectral operations
const int n_embd = llama_n_embd(model_cts);
const int n_embd = llama_model_n_embd(model_cts);
const float * embd = llama_get_embeddings(ctx_cts);
auto audio = embd_to_audio(embd, n_codes, n_embd, params.cpuparams.n_threads);
@ -920,12 +924,6 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
LOG_INF("%s: audio written to file '%s'\n", __func__, fname.c_str());
llama_free(ctx_ttc);
llama_free_model(model_ttc);
llama_free(ctx_cts);
llama_free_model(model_cts);
llama_backend_free();
return 0;

23
ggml/CMakeLists.txt
View File

@ -243,7 +243,8 @@ set(GGML_PUBLIC_HEADERS
include/ggml-metal.h
include/ggml-rpc.h
include/ggml-sycl.h
include/ggml-vulkan.h)
include/ggml-vulkan.h
include/gguf.h)
set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
#if (GGML_METAL)
@ -252,26 +253,6 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
install(TARGETS ggml LIBRARY PUBLIC_HEADER)
install(TARGETS ggml-base LIBRARY)
# FIXME: this should be done in the backend cmake files
if (GGML_METAL)
# FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
install(
FILES src/ggml-metal/ggml-metal.metal
PERMISSIONS
OWNER_READ
OWNER_WRITE
GROUP_READ
WORLD_READ
DESTINATION ${CMAKE_INSTALL_BINDIR})
if (NOT GGML_METAL_EMBED_LIBRARY)
install(
FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
DESTINATION ${CMAKE_INSTALL_BINDIR}
)
endif()
endif()
if (GGML_STANDALONE)
configure_file(${CMAKE_CURRENT_SOURCE_DIR}/ggml.pc.in
${CMAKE_CURRENT_BINARY_DIR}/ggml.pc

1
ggml/include/ggml-cpp.h
View File

@ -7,6 +7,7 @@
#include "ggml.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#include "gguf.h"
#include <memory>
// Smart pointers for ggml types

150
ggml/include/ggml.h
View File

@ -241,12 +241,6 @@
#define GGML_ROPE_TYPE_MROPE 8
#define GGML_ROPE_TYPE_VISION 24
#define GGUF_MAGIC "GGUF"
#define GGUF_VERSION 3
#define GGUF_DEFAULT_ALIGNMENT 32
#define GGML_UNUSED(x) (void)(x)
#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
@ -403,12 +397,6 @@ extern "C" {
GGML_PREC_F32,
};
enum ggml_backend_type {
GGML_BACKEND_TYPE_CPU = 0,
GGML_BACKEND_TYPE_GPU = 10,
GGML_BACKEND_TYPE_GPU_SPLIT = 20,
};
// model file types
enum ggml_ftype {
GGML_FTYPE_UNKNOWN = -1,
@ -513,6 +501,7 @@ extern "C" {
GGML_OP_GET_REL_POS,
GGML_OP_ADD_REL_POS,
GGML_OP_RWKV_WKV6,
GGML_OP_GATED_LINEAR_ATTN,
GGML_OP_UNARY,
@ -587,8 +576,6 @@ extern "C" {
struct ggml_tensor {
enum ggml_type type;
GGML_DEPRECATED(enum ggml_backend_type backend, "use the buffer type to find the storage location of the tensor");
struct ggml_backend_buffer * buffer;
int64_t ne[GGML_MAX_DIMS]; // number of elements
@ -1873,6 +1860,15 @@ extern "C" {
struct ggml_tensor * td,
struct ggml_tensor * state);
GGML_API struct ggml_tensor * ggml_gated_linear_attn(
struct ggml_context * ctx,
struct ggml_tensor * k,
struct ggml_tensor * v,
struct ggml_tensor * q,
struct ggml_tensor * g,
struct ggml_tensor * state,
float scale);
// custom operators
typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
@ -2111,132 +2107,6 @@ extern "C" {
int64_t n_per_row,
const float * imatrix);
//
// gguf
//
enum gguf_type {
GGUF_TYPE_UINT8 = 0,
GGUF_TYPE_INT8 = 1,
GGUF_TYPE_UINT16 = 2,
GGUF_TYPE_INT16 = 3,
GGUF_TYPE_UINT32 = 4,
GGUF_TYPE_INT32 = 5,
GGUF_TYPE_FLOAT32 = 6,
GGUF_TYPE_BOOL = 7,
GGUF_TYPE_STRING = 8,
GGUF_TYPE_ARRAY = 9,
GGUF_TYPE_UINT64 = 10,
GGUF_TYPE_INT64 = 11,
GGUF_TYPE_FLOAT64 = 12,
GGUF_TYPE_COUNT, // marks the end of the enum
};
struct gguf_context;
struct gguf_init_params {
bool no_alloc;
// if not NULL, create a ggml_context and allocate the tensor data in it
struct ggml_context ** ctx;
};
GGML_API struct gguf_context * gguf_init_empty(void);
GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params);
//GGML_API struct gguf_context * gguf_init_from_buffer(..);
GGML_API void gguf_free(struct gguf_context * ctx);
GGML_API const char * gguf_type_name(enum gguf_type type);
GGML_API int gguf_get_version (const struct gguf_context * ctx);
GGML_API size_t gguf_get_alignment (const struct gguf_context * ctx);
GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx);
GGML_API void * gguf_get_data (const struct gguf_context * ctx);
GGML_API int gguf_get_n_kv(const struct gguf_context * ctx);
GGML_API int gguf_find_key(const struct gguf_context * ctx, const char * key);
GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int key_id);
GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int key_id);
GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id);
// will abort if the wrong type is used for the key
GGML_API uint8_t gguf_get_val_u8 (const struct gguf_context * ctx, int key_id);
GGML_API int8_t gguf_get_val_i8 (const struct gguf_context * ctx, int key_id);
GGML_API uint16_t gguf_get_val_u16 (const struct gguf_context * ctx, int key_id);
GGML_API int16_t gguf_get_val_i16 (const struct gguf_context * ctx, int key_id);
GGML_API uint32_t gguf_get_val_u32 (const struct gguf_context * ctx, int key_id);
GGML_API int32_t gguf_get_val_i32 (const struct gguf_context * ctx, int key_id);
GGML_API float gguf_get_val_f32 (const struct gguf_context * ctx, int key_id);
GGML_API uint64_t gguf_get_val_u64 (const struct gguf_context * ctx, int key_id);
GGML_API int64_t gguf_get_val_i64 (const struct gguf_context * ctx, int key_id);
GGML_API double gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
GGML_API bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id);
GGML_API int gguf_get_arr_n (const struct gguf_context * ctx, int key_id);
GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
GGML_API int gguf_get_n_tensors (const struct gguf_context * ctx);
GGML_API int gguf_find_tensor (const struct gguf_context * ctx, const char * name);
GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
GGML_API char * gguf_get_tensor_name (const struct gguf_context * ctx, int i);
GGML_API enum ggml_type gguf_get_tensor_type (const struct gguf_context * ctx, int i);
// removes key if it exists
GGML_API void gguf_remove_key(struct gguf_context * ctx, const char * key);
// overrides existing values or adds a new one
GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val);
GGML_API void gguf_set_val_i8 (struct gguf_context * ctx, const char * key, int8_t val);
GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val);
GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t val);
GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);
GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t val);
GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float val);
GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val);
GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t val);
GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double val);
GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val);
GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);
GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n);
GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, int n);
// set or add KV pairs from another context
GGML_API void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src);
// manage tensor info
GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor);
GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type);
GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size);
// writing gguf files can be done in 2 ways:
//
// - write the entire gguf_context to a binary file in a single pass:
//
// gguf_write_to_file(ctx, fname);
//
// - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data:
//
// FILE * f = fopen(fname, "wb");
// fseek(f, gguf_get_meta_size(ctx), SEEK_SET);
// fwrite(f, ...);
// void * data = gguf_meta_get_meta_data(ctx);
// fseek(f, 0, SEEK_SET);
// fwrite(f, data, gguf_get_meta_size(ctx));
// free(data);
// fclose(f);
//
// write the entire context to a binary file
GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta);
// get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data);
#ifdef __cplusplus
// restrict not standard in C++
# if defined(__GNUC__)

202
ggml/include/gguf.h Normal file
View File

@ -0,0 +1,202 @@
// This file contains functionality related to "GGUF" files, the binary file format used by ggml.
// GGUF files have the following structure:
//
// 1. File magic "GGUF" (4 bytes).
// 2. File version (uint32_t).
// 3. Number of ggml tensors in file (int64_t).
// 4. Number of key-value-pairs in file (int64_t).
// 5. For each KV pair:
// 1. The key (string).
// 2. The value type (gguf_type).
// 3a. If the value type is GGUF_TYPE_ARRAY:
// 1. The type of the array (gguf_type).
// 2. The number of elements in the array (uint64_t).
// 3. The binary representation of each element in the array.
// 3b. Otherwise:
// 1. The binary representation of the value.
// 6. For each ggml tensor:
// 1. The tensor name (string).
// 2. The number of dimensions of the tensor (uint32_t).
// 3. For each dimension:
// 1. The size of the tensor in the dimension (int64_t).
// 4. The tensor data type (ggml_type).
// 5. The tensor data offset in the tensor data binary blob (uint64_t).
// 7. The tensor data binary blob (optional, aligned).
//
// Strings are serialized as the string length (uint64_t) followed by the C string without the null terminator.
// All enums are stored as int32_t.
// All bool values are stored as int8_t.
// If the special key "general.alignment" (uint32_t) is defined it is used for alignment,
// otherwise GGUF_DEFAULT_ALIGNMENT is used.
//
// Module maintainer: Johannes Gäßler (@JohannesGaessler, johannesg@5d6.de)
#pragma once
#include "ggml.h"
#include <stdbool.h>
#include <stdint.h>
#define GGUF_MAGIC "GGUF"
#define GGUF_VERSION 3
#define GGUF_KEY_GENERAL_ALIGNMENT "general.alignment"
#define GGUF_DEFAULT_ALIGNMENT 32
#ifdef __cplusplus
extern "C" {
#endif
// types that can be stored as GGUF KV data
enum gguf_type {
GGUF_TYPE_UINT8 = 0,
GGUF_TYPE_INT8 = 1,
GGUF_TYPE_UINT16 = 2,
GGUF_TYPE_INT16 = 3,
GGUF_TYPE_UINT32 = 4,
GGUF_TYPE_INT32 = 5,
GGUF_TYPE_FLOAT32 = 6,
GGUF_TYPE_BOOL = 7,
GGUF_TYPE_STRING = 8,
GGUF_TYPE_ARRAY = 9,
GGUF_TYPE_UINT64 = 10,
GGUF_TYPE_INT64 = 11,
GGUF_TYPE_FLOAT64 = 12,
GGUF_TYPE_COUNT, // marks the end of the enum
};
struct gguf_context;
struct gguf_init_params {
bool no_alloc;
// if not NULL, create a ggml_context and allocate the tensor data in it
struct ggml_context ** ctx;
};
GGML_API struct gguf_context * gguf_init_empty(void);
GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params);
//GGML_API struct gguf_context * gguf_init_from_buffer(..);
GGML_API void gguf_free(struct gguf_context * ctx);
GGML_API const char * gguf_type_name(enum gguf_type type);
GGML_API uint32_t gguf_get_version (const struct gguf_context * ctx);
GGML_API size_t gguf_get_alignment (const struct gguf_context * ctx);
GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx);
GGML_API int64_t gguf_get_n_kv(const struct gguf_context * ctx);
GGML_API int64_t gguf_find_key(const struct gguf_context * ctx, const char * key); // returns -1 if key is not found
GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int64_t key_id);
GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int64_t key_id);
GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int64_t key_id);
// will abort if the wrong type is used for the key
GGML_API uint8_t gguf_get_val_u8 (const struct gguf_context * ctx, int64_t key_id);
GGML_API int8_t gguf_get_val_i8 (const struct gguf_context * ctx, int64_t key_id);
GGML_API uint16_t gguf_get_val_u16 (const struct gguf_context * ctx, int64_t key_id);
GGML_API int16_t gguf_get_val_i16 (const struct gguf_context * ctx, int64_t key_id);
GGML_API uint32_t gguf_get_val_u32 (const struct gguf_context * ctx, int64_t key_id);
GGML_API int32_t gguf_get_val_i32 (const struct gguf_context * ctx, int64_t key_id);
GGML_API float gguf_get_val_f32 (const struct gguf_context * ctx, int64_t key_id);
GGML_API uint64_t gguf_get_val_u64 (const struct gguf_context * ctx, int64_t key_id);
GGML_API int64_t gguf_get_val_i64 (const struct gguf_context * ctx, int64_t key_id);
GGML_API double gguf_get_val_f64 (const struct gguf_context * ctx, int64_t key_id);
GGML_API bool gguf_get_val_bool(const struct gguf_context * ctx, int64_t key_id);
GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int64_t key_id);
GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int64_t key_id);
GGML_API size_t gguf_get_arr_n (const struct gguf_context * ctx, int64_t key_id);
// get raw pointer to the first element of the array with the given key_id
// for bool arrays, note that they are always stored as int8 on all platforms (usually this makes no difference)
GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int64_t key_id);
// get ith C string from array with given key_id
GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int64_t key_id, size_t i);
GGML_API int64_t gguf_get_n_tensors (const struct gguf_context * ctx);
GGML_API int64_t gguf_find_tensor (const struct gguf_context * ctx, const char * name); // returns -1 if the tensor is not found
GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int64_t tensor_id);
GGML_API const char * gguf_get_tensor_name (const struct gguf_context * ctx, int64_t tensor_id);
GGML_API enum ggml_type gguf_get_tensor_type (const struct gguf_context * ctx, int64_t tensor_id);
GGML_API size_t gguf_get_tensor_size (const struct gguf_context * ctx, int64_t tensor_id);
// removes key if it exists, returns id that the key had prior to removal (-1 if it didn't exist)
GGML_API int64_t gguf_remove_key(struct gguf_context * ctx, const char * key);
// overrides an existing KV pair or adds a new one, the new KV pair is always at the back
GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val);
GGML_API void gguf_set_val_i8 (struct gguf_context * ctx, const char * key, int8_t val);
GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val);
GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t val);
GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);
GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t val);
GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float val);
GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val);
GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t val);
GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double val);
GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val);
GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);
// creates a new array with n elements of the given type and copies the corresponding number of bytes from data
GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, size_t n);
// creates a new array with n strings and copies the corresponding strings from data
GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, size_t n);
// set or add KV pairs from another context
GGML_API void gguf_set_kv(struct gguf_context * ctx, const struct gguf_context * src);
// add tensor to GGUF context, tensor name must be unique
GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor);
// after changing a tensor's type, the offsets of all tensors with higher indices are immediately recalculated
// in such a way that the tensor data remains as one contiguous block (except for padding)
GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type);
// assumes that at least gguf_get_tensor_size bytes can be read from data
GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data);
// writing gguf files can be done in 3 ways:
//
// - write the entire gguf_context to a binary file in a single pass:
//
// gguf_write_to_file(ctx, fname, /*only_meta =*/ false);
//
// - write only the meta data to a file, then re-open the file and append the tensor data:
//
// gguf_write_to_file(ctx, fname, /*only_meta =*/ true);
// FILE * f = fopen(fname, "ab");
// fwrite(f, ...); // write tensor data
// fclose(f);
//
// - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data:
//
// FILE * f = fopen(fname, "wb");
// const size_t size_meta = gguf_get_meta_size(ctx);
// fseek(f, size_meta, SEEK_SET);
// fwrite(f, ...); // write tensor data
// void * data = malloc(size_meta);
// gguf_get_meta_data(ctx, data);
// rewind(f);
// fwrite(data, 1, data, f);
// free(data);
// fclose(f);
//
// write the entire context to a binary file
GGML_API bool gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta);
// get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
// writes the meta data to pointer "data"
GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data);
#ifdef __cplusplus
}
#endif

8
ggml/src/CMakeLists.txt
View File

@ -208,6 +208,7 @@ add_library(ggml-base
../include/ggml-backend.h
../include/ggml-cpp.h
../include/ggml-opt.h
../include/gguf.h
ggml.c
ggml-alloc.c
ggml-backend.cpp
@ -215,7 +216,8 @@ add_library(ggml-base
ggml-threading.cpp
ggml-threading.h
ggml-quants.c
ggml-quants.h)
ggml-quants.h
gguf.cpp)
target_include_directories(ggml-base PRIVATE .)
@ -290,9 +292,9 @@ if (GGML_CPU_ALL_VARIANTS)
ggml_add_cpu_backend_variant(haswell AVX F16C AVX2 FMA)
ggml_add_cpu_backend_variant(skylakex AVX F16C AVX2 FMA AVX512)
ggml_add_cpu_backend_variant(icelake AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
ggml_add_cpu_backend_variant(alderlake AVX F16C AVX2 FMA AVX_VNNI)
if (NOT MSVC)
# MSVC doesn't support AVX-VNNI or AMX
ggml_add_cpu_backend_variant(alderlake AVX F16C AVX2 FMA AVX_VNNI)
# MSVC doesn't support AMX
ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
endif()
else ()

5
ggml/src/ggml-backend-reg.cpp
View File

@ -574,4 +574,9 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
ggml_backend_load_best("opencl", silent, dir_path);
ggml_backend_load_best("musa", silent, dir_path);
ggml_backend_load_best("cpu", silent, dir_path);
// check the environment variable GGML_BACKEND_PATH to load an out-of-tree backend
const char * backend_path = std::getenv("GGML_BACKEND_PATH");
if (backend_path) {
ggml_backend_load(backend_path);
}
}

7
ggml/src/ggml-backend.cpp
View File

@ -764,7 +764,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
// check if a backend with higher prio wants to offload the op
if (src_backend_id == sched->n_backends - 1) {
if (src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
for (int b = 0; b < src_backend_id; b++) {
if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
SET_CAUSE(tensor, "1.off");
@ -795,9 +795,12 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
for (int i = 0; i < graph->n_nodes; i++) {
if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs", cur_split, ggml_backend_name(split_backend),
sched->splits[cur_split].n_inputs);
for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
if (j == 0) {
GGML_LOG_DEBUG(": ");
}
GGML_LOG_DEBUG("[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
}

3
ggml/src/ggml-cpu/CMakeLists.txt
View File

@ -215,8 +215,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
list(APPEND ARCH_DEFINITIONS GGML_SSE42)
endif()
if (GGML_AVX_VNNI)
# MSVC generates AVX512 with AVX-VNNI intrinsics even with /arch:AVX2
#list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
endif()
else ()
if (GGML_NATIVE)

7
ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
View File

@ -194,9 +194,12 @@ static inline __m256i sum_i16_pairs_int32x8(const __m256i x) {
}
static inline __m256i mul_sum_us8_pairs_int32x8(const __m256i ax, const __m256i sy) {
#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
const __m256i zero = _mm256_setzero_si256();
return _mm256_dpbusd_epi32(zero, ax, sy);
#elif defined(__AVXVNNI__)
const __m256i zero = _mm256_setzero_si256();
return _mm256_dpbusd_avx_epi32(zero, ax, sy);
#else
// Perform multiplication and create 16-bit values
const __m256i dot = _mm256_maddubs_epi16(ax, sy);
@ -4166,6 +4169,8 @@ static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(g
buffer->buft = buft;
buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
buffer->iface.set_tensor = ggml_backend_cpu_aarch64_buffer_set_tensor;
buffer->iface.get_tensor = nullptr;
buffer->iface.cpy_tensor = nullptr;
return buffer;
}

6
ggml/src/ggml-cpu/ggml-cpu-quants.c
View File

@ -103,10 +103,14 @@ static inline __m256 sum_i16_pairs_float(const __m256i x) {
}
static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
const __m256i zero = _mm256_setzero_si256();
const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
return _mm256_cvtepi32_ps(summed_pairs);
#elif defined(__AVXVNNI__)
const __m256i zero = _mm256_setzero_si256();
const __m256i summed_pairs = _mm256_dpbusd_avx_epi32(zero, ax, sy);
return _mm256_cvtepi32_ps(summed_pairs);
#else
// Perform multiplication and create 16-bit values
const __m256i dot = _mm256_maddubs_epi16(ax, sy);

200
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -11803,9 +11803,9 @@ static void ggml_compute_forward_add_rel_pos(
static void ggml_compute_forward_rwkv_wkv6_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
const int64_t T = dst->src[1]->ne[3];
const int64_t T = dst->src[1]->ne[2];
const int64_t C = dst->ne[0];
const int64_t HEADS = dst->src[1]->ne[2];
const int64_t HEADS = dst->src[1]->ne[1];
const int64_t n_seqs = dst->src[5]->ne[1];
const int64_t head_size = C / HEADS;
@ -12000,6 +12000,197 @@ static void ggml_compute_forward_rwkv_wkv6(
}
}
// ggml_compute_forward_gla
static void ggml_compute_forward_gla_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
const int64_t T = dst->src[1]->ne[2];
const int64_t C = dst->ne[0];
const int64_t HEADS = dst->src[1]->ne[1];
const int64_t n_seqs = dst->src[4]->ne[1];
const int64_t head_size = C / HEADS;
const float scale = ggml_get_op_params_f32(dst, 0);
float * dst_data = (float *) dst->data;
float * state = ((float *) dst->data) + C * T;
const int ith = params->ith;
const int nth = params->nth;
if (ith >= HEADS) {
return;
}
const int h_start = (HEADS * ith) / nth;
const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
(HEADS * (ith + 1)) / nth : HEADS;
float * k = (float *) dst->src[0]->data;
float * v = (float *) dst->src[1]->data;
float * q = (float *) dst->src[2]->data;
float * g = (float *) dst->src[3]->data;
size_t t_stride = HEADS * head_size; // Same to C
size_t h_stride = C / HEADS;
GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS
size_t h_stride_2d = head_size * head_size;
if (ith == 0) {
memset(dst_data, 0, T * C * sizeof(float));
}
ggml_barrier(params->threadpool);
#if defined(__AVX__) && !defined(__AVX512F__)
#define GGML_F32X GGML_F32x8
#define GGML_F32X_SET1 GGML_F32x8_SET1
#define GGML_F32X_LOAD GGML_F32x8_LOAD
#define GGML_F32X_STORE GGML_F32x8_STORE
#define GGML_F32X_MUL GGML_F32x8_MUL
#define GGML_F32X_FMA GGML_F32x8_FMA
#define GLA_VECTOR_SIZE 8
#elif defined(__AVX512F__)
#define GGML_F32X GGML_F32x16
#define GGML_F32X_SET1 GGML_F32x16_SET1
#define GGML_F32X_LOAD GGML_F32x16_LOAD
#define GGML_F32X_STORE GGML_F32x16_STORE
#define GGML_F32X_MUL GGML_F32x16_MUL
#define GGML_F32X_FMA GGML_F32x16_FMA
#define GLA_VECTOR_SIZE 16
#elif defined(__ARM_NEON) && defined(__aarch64__)
#define GGML_F32X GGML_F32x4
#define GGML_F32X_SET1 GGML_F32x4_SET1
#define GGML_F32X_LOAD GGML_F32x4_LOAD
#define GGML_F32X_STORE GGML_F32x4_STORE
#define GGML_F32X_MUL GGML_F32x4_MUL
#define GGML_F32X_FMA GGML_F32x4_FMA
#define GLA_VECTOR_SIZE 4
#endif
#ifdef GLA_VECTOR_SIZE
const int64_t vec_count = head_size / GLA_VECTOR_SIZE;
for (int64_t t = 0; t < T; t++) {
size_t t_offset = t * t_stride;
size_t state_offset = head_size * C * (t / (T / n_seqs));
float * state_cur = state + state_offset;
float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[4]->data + state_offset;
for (int64_t h = h_start; h < h_end; h++) {
size_t h_offset = h * h_stride;
size_t t_h_offset = t_offset + h_offset;
size_t h_2d_offset = h * h_stride_2d;
for (int64_t i = 0; i < head_size; i++) {
size_t t_h_i_offset = t_h_offset + i;
size_t h_2d_i_offset = h_2d_offset + i * h_stride;
float k_val = k[t_h_i_offset];
float q_val = q[t_h_i_offset] * scale;
float g_val = g[t_h_i_offset];
// Broadcast scalar values to vectors
GGML_F32X k_vec = GGML_F32X_SET1(k_val);
GGML_F32X q_vec = GGML_F32X_SET1(q_val);
GGML_F32X g_vec = GGML_F32X_SET1(g_val);
for (int64_t j = 0; j < vec_count; j++) {
size_t base_j = j * GLA_VECTOR_SIZE;
size_t t_h_j_offset = t_h_offset + base_j;
size_t h_2d_i_j_offset = h_2d_i_offset + base_j;
// Load x elements at once
GGML_F32X v_vec = GGML_F32X_LOAD(&v[t_h_j_offset]);
GGML_F32X prev_state_vec = GGML_F32X_LOAD(&state_prev[h_2d_i_j_offset]);
GGML_F32X dst_vec = GGML_F32X_LOAD(&dst_data[t_h_j_offset]);
// Compute kv = v * k
GGML_F32X kv_vec = GGML_F32X_MUL(v_vec, k_vec);
// Compute temp = prev_state * g + kv
GGML_F32X temp_vec = GGML_F32X_FMA(kv_vec, prev_state_vec, g_vec);
// Update dst: dst += temp * q
dst_vec = GGML_F32X_FMA(dst_vec, temp_vec, q_vec);
GGML_F32X_STORE(&dst_data[t_h_j_offset], dst_vec);
// Update state
GGML_F32X_STORE(&state_cur[h_2d_i_j_offset], temp_vec);
}
// Handle remaining elements, this will not be used.
for (int64_t j = vec_count * GLA_VECTOR_SIZE; j < head_size; j++) {
size_t t_h_j_offset = t_h_offset + j;
size_t h_2d_i_j_offset = h_2d_i_offset + j;
float v_val = v[t_h_j_offset];
float kv_val = v_val * k_val;
float prev_state_val = state_prev[h_2d_i_j_offset];
float temp_val = kv_val + prev_state_val * g_val;
dst_data[t_h_j_offset] += temp_val * q_val;
state_cur[h_2d_i_j_offset] = temp_val;
}
}
}
}
#else
for (int64_t t = 0; t < T; t++) {
size_t t_offset = t * t_stride;
size_t state_offset = head_size * C * (t / (T / n_seqs));
float * state_cur = state + state_offset;
float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[4]->data + state_offset;
for (int64_t h = h_start; h < h_end; h++) {
size_t h_offset = h * h_stride;
size_t t_h_offset = t_offset + h_offset;
size_t h_2d_offset = h * h_stride_2d;
for (int64_t i = 0; i < head_size; i++) {
size_t t_h_i_offset = t_h_offset + i;
size_t h_2d_i_offset = h_2d_offset + i * h_stride;
float k_val = k[t_h_i_offset];
float q_val = q[t_h_i_offset] * scale;
float g_val = g[t_h_i_offset];
for (int64_t j = 0; j < head_size; j++) {
size_t t_h_j_offset = t_h_offset + j;
size_t h_2d_i_j_offset = h_2d_i_offset + j;
float v_val = v[t_h_j_offset];
float kv_val = v_val * k_val;
float prev_state_val = state_prev[h_2d_i_j_offset];
float temp_val = prev_state_val * g_val + kv_val;
dst_data[t_h_j_offset] += temp_val * q_val;
state_cur[h_2d_i_j_offset] = temp_val;
}
}
}
}
#endif
}
static void ggml_compute_forward_gla(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_gla_f32(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_map_unary
static void ggml_compute_forward_map_unary_f32(
@ -12749,6 +12940,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_rwkv_wkv6(params, tensor);
} break;
case GGML_OP_GATED_LINEAR_ATTN:
{
ggml_compute_forward_gla(params, tensor);
} break;
case GGML_OP_MAP_UNARY:
{
ggml_unary_op_f32_t fun;
@ -13047,6 +13242,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_WIN_UNPART:
case GGML_OP_GET_REL_POS:
case GGML_OP_RWKV_WKV6:
case GGML_OP_GATED_LINEAR_ATTN:
case GGML_OP_MAP_UNARY:
case GGML_OP_MAP_BINARY:
case GGML_OP_MAP_CUSTOM1_F32:

840
ggml/src/ggml-cpu/llamafile/sgemm.cpp
View File

@ -54,6 +54,7 @@
#include "ggml-quants.h"
#include <atomic>
#include <array>
#ifdef _MSC_VER
#define NOINLINE __declspec(noinline)
@ -1000,8 +1001,10 @@ class tinyBLAS_Q0_AVX {
inline __m256 updot(__m256i u, __m256i s) {
__m256i res;
#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
res = _mm256_dpbusd_epi32(_mm256_setzero_si256(), u, s);
#elif defined(__AVXVNNI__)
res = _mm256_dpbusd_avx_epi32(_mm256_setzero_si256(), u, s);
#else
res = _mm256_madd_epi16(_mm256_set1_epi16(1), _mm256_maddubs_epi16(u, s));
#endif
@ -1049,6 +1052,704 @@ class tinyBLAS_Q0_AVX {
} \
} \
template <typename TA, typename TB, typename TC>
class tinyBLAS_Q0_PPC {
public:
tinyBLAS_Q0_PPC(int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
private:
template<int RM, int RN>
inline void save_res(int ii, int jj, int idx, vector float* fin_res) {
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
}
}
}
template<int size>
inline void compute(acc_t* ACC, int c_idx, int s_idx, std::array<int, size>& comparray, vector float* vs, vector float* fin_res) {
vector signed int vec_C[4];
vector float CA[4] = {0};
vector float res[4] = {0};
__builtin_mma_disassemble_acc(vec_C, ACC);
for (int i = 0; i < 4; i++) {
CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
}
}
template<typename VA, typename VB>
void packNormal(const TA* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
int64_t i, j;
TA *aoffset = NULL;
VA *vecOffset = NULL;
TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
__vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2]={0};
VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2]={0};
VB t1, t2, t3, t4, t5, t6, t7, t8;
vector unsigned char xor_vector;
uint8_t flip_vec = 0x80;
xor_vector = vec_splats(flip_vec);
vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
aoffset = const_cast<TA*>(a);
vecOffset = vec;
j = (rows >> 3);
if (j > 0) {
do {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset5 = aoffset4 + lda;
aoffset6 = aoffset5 + lda;
aoffset7 = aoffset6 + lda;
aoffset8 = aoffset7 + lda;
aoffset += 8 * lda;
i = (cols >> 3);
if (i > 0) {
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4->qs);
C5 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset5->qs);
C6 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset6->qs);
C7 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset7->qs);
C8 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset8->qs);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
__builtin_vsx_disassemble_pair(c5, &C5);
__builtin_vsx_disassemble_pair(c6, &C6);
__builtin_vsx_disassemble_pair(c7, &C7);
__builtin_vsx_disassemble_pair(c8, &C8);
t1 = vec_perm(c1[0], c2[0], swiz1);
t2 = vec_perm(c1[0], c2[0], swiz2);
t3 = vec_perm(c3[0], c4[0], swiz1);
t4 = vec_perm(c3[0], c4[0], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset);
vec_xst(t6, 0, vecOffset+16);
vec_xst(t7, 0, vecOffset+32);
vec_xst(t8, 0, vecOffset+48);
t1 = vec_perm(c1[1], c2[1], swiz1);
t2 = vec_perm(c1[1], c2[1], swiz2);
t3 = vec_perm(c3[1], c4[1], swiz1);
t4 = vec_perm(c3[1], c4[1], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+64);
vec_xst(t6, 0, vecOffset+80);
vec_xst(t7, 0, vecOffset+96);
vec_xst(t8, 0, vecOffset+112);
t1 = vec_perm(c5[0], c6[0], swiz1);
t2 = vec_perm(c5[0], c6[0], swiz2);
t3 = vec_perm(c7[0], c8[0], swiz1);
t4 = vec_perm(c7[0], c8[0], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+128);
vec_xst(t6, 0, vecOffset+144);
vec_xst(t7, 0, vecOffset+160);
vec_xst(t8, 0, vecOffset+176);
t1 = vec_perm(c5[1], c6[1], swiz1);
t2 = vec_perm(c5[1], c6[1], swiz2);
t3 = vec_perm(c7[1], c8[1], swiz1);
t4 = vec_perm(c7[1], c8[1], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+192);
vec_xst(t6, 0, vecOffset+208);
vec_xst(t7, 0, vecOffset+224);
vec_xst(t8, 0, vecOffset+240);
aoffset1 += lda;
aoffset2 += lda;
aoffset3 += lda;
aoffset4 += lda;
aoffset5 += lda;
aoffset6 += lda;
aoffset7 += lda;
aoffset8 += lda;
vecOffset += 256;
i--;
} while(i > 0);
}
j--;
} while(j > 0);
}
if (rows & 4) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset += 4 * lda;
i = (cols >> 3);
if (i > 0) {
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4->qs);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
t1 = vec_perm(c1[0], c2[0], swiz1);
t2 = vec_perm(c1[0], c2[0], swiz2);
t3 = vec_perm(c3[0], c4[0], swiz1);
t4 = vec_perm(c3[0], c4[0], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset);
vec_xst(t6, 0, vecOffset+16);
vec_xst(t7, 0, vecOffset+32);
vec_xst(t8, 0, vecOffset+48);
t1 = vec_perm(c1[1], c2[1], swiz1);
t2 = vec_perm(c1[1], c2[1], swiz2);
t3 = vec_perm(c3[1], c4[1], swiz1);
t4 = vec_perm(c3[1], c4[1], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+64);
vec_xst(t6, 0, vecOffset+80);
vec_xst(t7, 0, vecOffset+96);
vec_xst(t8, 0, vecOffset+112);
aoffset1 += lda;
aoffset2 += lda;
aoffset3 += lda;
aoffset4 += lda;
vecOffset += 128;
i--;
} while(i > 0);
}
}
if (rows & 3) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
i = (cols >> 3);
if (i > 0) {
do {
switch(rows) {
case 3: C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
__builtin_vsx_disassemble_pair(c3, &C3);
case 2: C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs);
__builtin_vsx_disassemble_pair(c2, &C2);
case 1: C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs);
__builtin_vsx_disassemble_pair(c1, &C1);
break;
}
t1 = vec_perm(c1[0], c2[0], swiz1);
t2 = vec_perm(c1[0], c2[0], swiz2);
t3 = vec_perm(c3[0], c4[0], swiz1);
t4 = vec_perm(c3[0], c4[0], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset);
vec_xst(t6, 0, vecOffset+16);
vec_xst(t7, 0, vecOffset+32);
vec_xst(t8, 0, vecOffset+48);
t1 = vec_perm(c1[1], c2[1], swiz1);
t2 = vec_perm(c1[1], c2[1], swiz2);
t3 = vec_perm(c3[1], c4[1], swiz1);
t4 = vec_perm(c3[1], c4[1], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+64);
vec_xst(t6, 0, vecOffset+80);
vec_xst(t7, 0, vecOffset+96);
vec_xst(t8, 0, vecOffset+112);
aoffset1 += lda;
aoffset2 += lda;
aoffset3 += lda;
vecOffset += 128;
i--;
} while(i > 0);
}
}
}
void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
int m_rem = MIN(m - m0, 8);
int n_rem = MIN(n - n0, 8);
// TO-DO: KERNEL_16x8 and KERNEL_8x16 are having some performance
// issues. After resolving them, below code will be enabled.
/*if (m_rem >= 16 && n_rem >= 8) {
mc = 16;
nc = 8;
gemm<16,8>(m0, m, n0, n);
} else if(m_rem >= 8 && n_rem >= 16) {
mc = 8;
nc = 16;
gemm<8,16>(m0, m, n0, n);
}*/
if (m_rem >= 8 && n_rem >= 8) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 8) {
mc = 4;
nc = 8;
gemm<4,8>(m0, m, n0, n);
} else if (m_rem >= 8 && n_rem >= 4) {
mc = 8;
nc = 4;
gemm<8,4>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 4) {
mc = 4;
nc = 4;
gemm_small<4, 4>(m0, m, n0, n);
} else if ((m_rem < 4) && (n_rem > 4)) {
nc = 4;
switch(m_rem) {
case 1:
mc = 1;
gemm_small<1, 4>(m0, m, n0, n);
break;
case 2:
mc = 2;
gemm_small<2, 4>(m0, m, n0, n);
break;
case 3:
mc = 3;
gemm_small<3, 4>(m0, m, n0, n);
break;
default:
return;
}
} else if ((m_rem > 4) && (n_rem < 4)) {
mc = 4;
switch(n_rem) {
case 1:
nc = 1;
gemm_small<4, 1>(m0, m, n0, n);
break;
case 2:
nc = 2;
gemm_small<4, 2>(m0, m, n0, n);
break;
case 3:
nc = 3;
gemm_small<4, 3>(m0, m, n0, n);
break;
default:
return;
}
} else {
switch((m_rem << 4) | n_rem) {
case 0x43:
mc = 4;
nc = 3;
gemm_small<4, 3>(m0, m, n0, n);
break;
case 0x42:
mc = 4;
nc = 2;
gemm_small<4, 2>(m0, m, n0, n);
break;
case 0x41:
mc = 4;
nc = 1;
gemm_small<4, 1>(m0, m, n0, n);
break;
case 0x34:
mc = 3;
nc = 4;
gemm_small<3, 4>(m0, m, n0, n);
break;
case 0x33:
mc = 3;
nc = 3;
gemm_small<3, 3>(m0, m, n0, n);
break;
case 0x32:
mc = 3;
nc = 2;
gemm_small<3, 2>(m0, m, n0, n);
break;
case 0x31:
mc = 3;
nc = 1;
gemm_small<3, 1>(m0, m, n0, n);
break;
case 0x24:
mc = 2;
nc = 4;
gemm_small<2, 4>(m0, m, n0, n);
break;
case 0x23:
mc = 2;
nc = 3;
gemm_small<2, 3>(m0, m, n0, n);
break;
case 0x22:
mc = 2;
nc = 2;
gemm_small<2, 2>(m0, m, n0, n);
break;
case 0x21:
mc = 2;
nc = 1;
gemm_small<2, 1>(m0, m, n0, n);
break;
case 0x14:
mc = 1;
nc = 4;
gemm_small<1, 4>(m0, m, n0, n);
break;
case 0x13:
mc = 1;
nc = 3;
gemm_small<1, 3>(m0, m, n0, n);
break;
case 0x12:
mc = 1;
nc = 2;
gemm_small<1, 2>(m0, m, n0, n);
break;
case 0x11:
mc = 1;
nc = 1;
gemm_small<1, 1>(m0, m, n0, n);
break;
default:
return;
}
}
mp = m0 + (m - m0) / mc * mc;
np = n0 + (n - n0) / nc * nc;
mnpack(mp, m, n0, np);
mnpack(m0, m, np, n);
}
void KERNEL_4x8(int64_t ii, int64_t jj) {
vec_t vec_A[8], vec_B[16] = {0};
acc_t acc_0, acc_1;
std::array<int, 4> comparray;
vector float fin_res[8] = {0};
vector float vs[8] = {0};
for (int l = 0; l < k; l++) {
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
for(int x = 0; x < 8; x++) {
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_1, vec_A[x], vec_B[x+8]);
}
for (int I = 0; I<4; I++) {
for (int J = 0; J<4; J++) {
*((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
*((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
}
}
auto aoffset = A+(ii*lda)+l;
for (int i = 0; i < 4; i++) {
comparray[i] = 0;
int ca = 0;
const int8_t *at = aoffset->qs;
for (int j = 0; j < 32; j++)
ca += (int)*at++;
comparray[i] = ca;
aoffset += lda;
}
compute<4>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<4>(&acc_1, 0, 4, comparray, vs, fin_res);
}
save_res<4, 4>(ii, jj, 0, fin_res);
save_res<4, 4>(ii, jj+4, 4, fin_res);
}
void KERNEL_8x4(int64_t ii, int64_t jj) {
vec_t vec_A[16], vec_B[8] = {0};
acc_t acc_0, acc_1;
std::array<int, 8> comparray;
vector float fin_res[8] = {0};
vector float vs[8] = {0};
for (int l = 0; l < k; l++) {
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true);
for(int x = 0; x < 8; x++) {
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
}
for (int I = 0; I<8; I++) {
for (int J = 0; J<4; J++) {
*((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
}
}
auto aoffset = A+(ii*lda)+l;
for (int i = 0; i < 8; i++) {
comparray[i] = 0;
int ca = 0;
const int8_t *at = aoffset->qs;
for (int j = 0; j < 32; j++)
ca += (int)*at++;
comparray[i] = ca;
aoffset += lda;
}
compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
}
save_res<4, 4>(ii, jj, 0, fin_res);
save_res<4, 4>(ii+4, jj, 4, fin_res);
}
void KERNEL_8x8(int64_t ii, int64_t jj) {
vec_t vec_A[16], vec_B[16] = {0};
acc_t acc_0, acc_1, acc_2, acc_3;
std::array<int, 8> comparray;
vector float fin_res[16] = {0};
vector float vs[16] = {0};
for (int l = 0; l < k; l++) {
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
__builtin_mma_xxsetaccz(&acc_2);
__builtin_mma_xxsetaccz(&acc_3);
packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
for(int x = 0; x < 8; x++) {
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_2, vec_A[x], vec_B[x+8]);
__builtin_mma_xvi8ger4pp(&acc_3, vec_A[x+8], vec_B[x+8]);
}
for (int I = 0; I<8; I++) {
for (int J = 0; J<4; J++) {
*((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
*((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
}
}
auto aoffset = A+(ii*lda)+l;
for (int i = 0; i < 8; i++) {
comparray[i] = 0;
int ca = 0;
const int8_t *at = aoffset->qs;
for (int j = 0; j < 32; j++)
ca += (int)*at++;
comparray[i] = ca;
aoffset += lda;
}
compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
compute<8>(&acc_2, 0, 8, comparray, vs, fin_res);
compute<8>(&acc_3, 4, 12, comparray, vs, fin_res);
}
save_res<4, 4>(ii, jj, 0, fin_res);
save_res<4, 4>(ii+4, jj, 4, fin_res);
save_res<4, 4>(ii, jj+4, 8, fin_res);
save_res<4, 4>(ii+4, jj+4, 12, fin_res);
}
template<int RM, int RN>
void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
vec_t vec_A[8], vec_B[8] = {0};
vector signed int vec_C[4];
acc_t acc_0;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
std::array<int, RM> comparray;
vector float res[4] = {0};
vector float fin_res[4] = {0};
vector float vs[4] = {0};
vector float CA[4] = {0};
__builtin_prefetch((A+(ii*lda)+0)->qs, 0, 1); // prefetch first value
__builtin_prefetch((B+(jj*ldb)+0)->qs, 0, 1); // prefetch first value
for (int l = 0; l < k; l++) {
__builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead
__builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead
__builtin_mma_xxsetaccz(&acc_0);
packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true);
for(int x = 0; x < 8; x+=4) {
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+1], vec_B[x+1]);
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+2], vec_B[x+2]);
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+3], vec_B[x+3]);
}
for (int I = 0; I<RM; I++) {
for (int J = 0; J<RN; J++) {
*((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
}
}
__builtin_mma_disassemble_acc(vec_C, &acc_0);
auto aoffset = A+(ii*lda)+l;
for (int i = 0; i < RM; i++) {
comparray[i] = 0;
int ca = 0;
const int8_t *at = aoffset->qs;
for (int j = 0; j < 32; j++)
ca += (int)*at++;
comparray[i] = ca;
aoffset += lda;
}
for (int i = 0; i < RM; i++) {
CA[i] = vec_splats((float)(((double)comparray[i]) * -128.0));
res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
fin_res[i] = vec_madd(res[i], vs[i], fin_res[i]);
}
}
save_res<RM, RN>(ii, jj, 0, fin_res);
}
}
template<int RM, int RN>
inline void kernel(int64_t ii, int64_t jj) {
if constexpr(RM == 4 && RN == 8) {
KERNEL_4x8(ii,jj);
} else if constexpr(RM == 8 && RN == 4) {
KERNEL_8x4(ii,jj);
} else if constexpr(RM == 8 && RN == 8) {
KERNEL_8x8(ii,jj);
} else {
static_assert(false, "RN/RM values not supported");
}
}
template <int RM, int RN>
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
kernel<RM, RN>(ii, jj);
}
}
const TA *const A;
const TB *const B;
TC *C;
TA *At;
TB *Bt;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
template <typename TA, typename TB, typename TC>
class tinyBLAS_PPC {
public:
@ -1068,13 +1769,17 @@ class tinyBLAS_PPC {
void (tinyBLAS_PPC::*kernel)(int64_t, int64_t);
void READ_BLOCK(const float* a, int64_t lda, int rows, int cols, float* vec) {
template<typename VA>
void packTranspose(const TA* a, int64_t lda, int rows, int cols, TA* vec) {
int64_t i, j;
float *aoffset = NULL, *boffset = NULL;
float *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
float *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
aoffset = const_cast<float*>(a);
TA *aoffset = NULL, *boffset = NULL;
TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
__vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
VA c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
VA c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
VA t1, t2, t3, t4, t5, t6, t7, t8;
aoffset = const_cast<TA*>(a);
boffset = vec;
j = (rows >> 3);
if (j > 0) {
@ -1090,9 +1795,6 @@ class tinyBLAS_PPC {
aoffset += 8 * lda;
i = (cols >> 3);
if (i > 0) {
__vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
vector float c1[2], c2[2], c3[2], c4[2], c5[2], c6[2], c7[2], c8[2];
vector float t1, t2, t3, t4, t5, t6, t7, t8;
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
@ -1172,21 +1874,19 @@ class tinyBLAS_PPC {
} while(i > 0);
}
if (cols & 4) {
vector float c1, c2, c3, c4, c5, c6, c7, c8;
vector float t1, t2, t3, t4, t5, t6, t7, t8;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
c4 = vec_xl(0, aoffset4);
c5 = vec_xl(0, aoffset5);
c6 = vec_xl(0, aoffset6);
c7 = vec_xl(0, aoffset7);
c8 = vec_xl(0, aoffset8);
c1[0] = vec_xl(0, aoffset1);
c2[0] = vec_xl(0, aoffset2);
c3[0] = vec_xl(0, aoffset3);
c4[0] = vec_xl(0, aoffset4);
c5[0] = vec_xl(0, aoffset5);
c6[0] = vec_xl(0, aoffset6);
c7[0] = vec_xl(0, aoffset7);
c8[0] = vec_xl(0, aoffset8);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t3 = vec_mergeh(c5, c6);
t4 = vec_mergeh(c7, c8);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_mergeh(c5[0], c6[0]);
t4 = vec_mergeh(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
@ -1196,10 +1896,10 @@ class tinyBLAS_PPC {
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t3 = vec_mergel(c5, c6);
t4 = vec_mergel(c7, c8);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_mergel(c5[0], c6[0]);
t4 = vec_mergel(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
@ -1221,9 +1921,6 @@ class tinyBLAS_PPC {
aoffset += 4 * lda;
i = (cols >> 3);
if (i > 0) {
__vector_pair C1, C2, C3, C4;
vector float c1[2], c2[2], c3[2], c4[2];
vector float t1, t2, t3, t4, t5, t6, t7, t8;
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
@ -1270,22 +1967,20 @@ class tinyBLAS_PPC {
}
if (cols & 4) {
vector float c1, c2, c3, c4;
vector float t1, t2, t3, t4;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
c4 = vec_xl(0, aoffset4);
c1[0] = vec_xl(0, aoffset1);
c2[0] = vec_xl(0, aoffset2);
c3[0] = vec_xl(0, aoffset3);
c4[0] = vec_xl(0, aoffset4);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset);
vec_xst(t4, 0, boffset+4);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset+8);
@ -1297,21 +1992,19 @@ class tinyBLAS_PPC {
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
if (cols & 4) {
vector float c1, c2, c3, c4 = {0};
vector float t1, t2, t3, t4;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
c1[0] = vec_xl(0, aoffset1);
c2[0] = vec_xl(0, aoffset2);
c3[0] = vec_xl(0, aoffset3);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset);
vec_xst(t4, 0, boffset+4);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset+8);
@ -1319,14 +2012,13 @@ class tinyBLAS_PPC {
}
}
}
void KERNEL_4x4(int64_t ii, int64_t jj) {
vec_t vec_A[4], vec_B[4], vec_C[4];
acc_t acc_0;
__builtin_mma_xxsetaccz(&acc_0);
for (int l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
packTranspose<vector float>(A+(ii*lda)+l, lda, 4, 4, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
@ -1341,8 +2033,8 @@ class tinyBLAS_PPC {
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
for (int64_t l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 4, (float*)vec_B);
packTranspose<vector float>(A+(ii*lda)+l, lda, 4, 4, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 8, 4, (TA*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]);
@ -1362,8 +2054,8 @@ class tinyBLAS_PPC {
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
for (int64_t l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 8, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
packTranspose<vector float>(A+(ii*lda)+l, lda, 8, 4, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]);
@ -1385,8 +2077,8 @@ class tinyBLAS_PPC {
__builtin_mma_xxsetaccz(&acc_2);
__builtin_mma_xxsetaccz(&acc_3);
for (int l = 0; l < k; l+=8) {
READ_BLOCK(A+(ii*lda)+l, lda, 8, 8, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 8, (float*)vec_B);
packTranspose<vector float>(A+(ii*lda)+l, lda, 8, 8, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 8, 8, (TA*)vec_B);
for(int x = 0; x < 16; x+=2) {
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]);
@ -1569,15 +2261,15 @@ class tinyBLAS_PPC {
vec_t vec_A[4], vec_B[4];
for (int l=0; l<k; l+=4) {
if (RN >= 4 && RM == 1) {
float* a = const_cast<float*>(A+(ii)*lda+l);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
TA* a = const_cast<TA*>(A+(ii)*lda+l);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
vec_A[0] = (vec_t)vec_xl(0,a);
vec_A[1] = (vec_t)vec_splats(*((float*)&vec_A+1));
vec_A[2] = (vec_t)vec_splats(*((float*)&vec_A+2));
vec_A[3] = (vec_t)vec_splats(*((float*)&vec_A+3));
vec_A[1] = (vec_t)vec_splats(*((TA*)&vec_A+1));
vec_A[2] = (vec_t)vec_splats(*((TA*)&vec_A+2));
vec_A[3] = (vec_t)vec_splats(*((TA*)&vec_A+3));
} else {
READ_BLOCK(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
packTranspose<vector float>(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B);
}
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
@ -1587,7 +2279,7 @@ class tinyBLAS_PPC {
__builtin_mma_disassemble_acc(vec_C, &acc_0);
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J);
*((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
}
}
}
@ -1810,6 +2502,20 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
params->ith, params->nth};
tb.matmul(m, n);
return true;
#elif defined(__MMA__)
if (n < 8 && n != 4)
return false;
if (m < 8 && m != 4)
return false;
tinyBLAS_Q0_PPC<block_q8_0, block_q8_0, float> tb{
k, (const block_q8_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif

2
ggml/src/ggml-cuda/concat.cu
View File

@ -124,7 +124,7 @@ static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
uint64_t nb1,
uint64_t nb2,
uint64_t nb3){
static_assert(dim >= 0 && dim <= 3);
static_assert(dim >= 0 && dim <= 3, "dim must be in [0, 3]");
const int64_t i3 = blockIdx.z;
const int64_t i2 = blockIdx.y;

2
ggml/src/ggml-cuda/convert.cu
View File

@ -680,6 +680,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
return dequantize_row_iq3_s_cuda;
case GGML_TYPE_F16:
return convert_unary_cuda<half>;
case GGML_TYPE_BF16:
return convert_unary_cuda<nv_bfloat16>;
default:
return nullptr;
}

434
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -37,6 +37,7 @@
#include "ggml-cuda/unary.cuh"
#include "ggml-cuda/upscale.cuh"
#include "ggml-cuda/wkv6.cuh"
#include "ggml-cuda/gla.cuh"
#include <algorithm>
#include <array>
@ -1728,7 +1729,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
bool use_mul_mat_vec = src0->type == GGML_TYPE_F16
bool use_mul_mat_vec = (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
@ -2167,6 +2168,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_OP_RWKV_WKV6:
ggml_cuda_op_rwkv_wkv6(ctx, dst);
break;
case GGML_OP_GATED_LINEAR_ATTN:
ggml_cuda_op_gated_linear_attn(ctx, dst);
break;
case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
ggml_cuda_cross_entropy_loss_back(ctx, dst);
break;
@ -2285,6 +2289,66 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
}
#ifdef USE_CUDA_GRAPH
static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool use_cuda_graph) {
// Loop over nodes in GGML graph to obtain info needed for CUDA graph
cuda_ctx->cuda_graph->updated_kernel_arg.clear();
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue;
}
if (node->src[0] && node->src[0]->buffer && ggml_backend_buft_is_cuda_split(node->src[0]->buffer->buft)) {
use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to split buffer\n", __func__);
#endif
}
if (node->op == GGML_OP_MUL_MAT_ID) {
use_cuda_graph = false; // This node type is not supported by CUDA graph capture
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
#endif
}
if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
// disable CUDA graphs for batch size > 1 for now.
// Changes in batch size or context size can cause changes to the grid size of some kernels.
use_cuda_graph = false;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
#endif
}
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));
// 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 (!ptr) {
use_cuda_graph = false;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
#endif
} else {
if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
ggml_cuda_cpy_fn_ptrs.push_back(ptr);
}
}
}
if (!use_cuda_graph) {
break;
}
}
return use_cuda_graph;
}
static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
graph_node_properties->node_address = node->data;
graph_node_properties->node_op = node->op;
@ -2335,149 +2399,105 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
return true;
}
#endif
static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
static void maintain_cuda_graph(ggml_backend_cuda_context * cuda_ctx, std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool cuda_graph_update_required) {
ggml_cuda_set_device(cuda_ctx->device);
if (cuda_graph_update_required) {
// Extract nodes from graph
// First call with null argument gets number of nodes in graph
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
// Subsequent call with non-null argument gets nodes
cuda_ctx->cuda_graph->nodes.clear();
cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
cuda_ctx->cuda_graph->params.clear();
cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
if (cuda_ctx->cuda_graph->num_nodes > 0) {
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));
#ifdef USE_CUDA_GRAPH
static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
// Objects required for CUDA Graph
if (cuda_ctx->cuda_graph == nullptr) {
cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
}
bool use_cuda_graph = true;
bool cuda_graph_update_required = false;
// vector of pointers to CUDA cpy kernels, which are required to identify
// kernel parameters which need updated in the graph for each token
std::vector<void *> ggml_cuda_cpy_fn_ptrs;
if (cuda_ctx->cuda_graph->graph == nullptr) {
if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_AMPERE) {
cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
#endif
}
}
// Disable CUDA graphs in presence of env var, old GPU, use-case which is changing too rapidly,
// or previous graph capture failure.
// Also disable for multi-gpu for now. TO DO investigate
if (disable_cuda_graphs_due_to_env
|| cuda_ctx->cuda_graph->disable_due_to_gpu_arch
|| cuda_ctx->cuda_graph->disable_due_to_too_many_updates
|| cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture) {
use_cuda_graph = false;
}
if (use_cuda_graph) {
if (cuda_ctx->cuda_graph->instance == nullptr) {
cuda_graph_update_required = true;
}
// Check if the graph size has changed
if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
cuda_graph_update_required = true;
cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
}
// Loop over nodes in GGML graph to determine if CUDA graph update is required
// and store properties to allow this comparison for the next token
for (int i = 0; i < cgraph->n_nodes; i++) {
bool has_matching_properties = true;
if (!cuda_graph_update_required) {
has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
}
if (!has_matching_properties) {
cuda_graph_update_required = true;
}
set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
}
// Loop over nodes in GGML graph to obtain info needed for CUDA graph
cuda_ctx->cuda_graph->updated_kernel_arg.clear();
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue;
}
if (node->src[0] && node->src[0]->buffer && ggml_backend_buft_is_cuda_split(node->src[0]->buffer->buft)) {
use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to split buffer\n", __func__);
#endif
}
if (node->op == GGML_OP_MUL_MAT_ID) {
use_cuda_graph = false; // This node type is not supported by CUDA graph capture
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
#endif
}
if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
// disable CUDA graphs for batch size > 1 for now.
// Changes in batch size or context size can cause changes to the grid size of some kernels.
use_cuda_graph = false;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
#endif
}
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));
// 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 (!ptr) {
use_cuda_graph = false;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
#endif
} else {
if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
ggml_cuda_cpy_fn_ptrs.push_back(ptr);
// Loop over nodes, and extract kernel parameters from each node
for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
cudaGraphNodeType node_type;
CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type));
if (node_type == cudaGraphNodeTypeKernel) {
cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime
if (stat == cudaErrorInvalidDeviceFunction) {
// Fails due to incorrect handling by CUDA runtime of CUDA BLAS node.
// We don't need to update blas nodes, so clear error and move on.
cudaGetLastError();
} else {
GGML_ASSERT(stat == cudaSuccess);
}
}
}
if (!use_cuda_graph) {
break;
}
} else {
// One of the arguments to the copy kernel is updated for each token, hence we need to
// replace that argument with the updated value in the CUDA graph
// 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(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]));
}
}
// Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
if (use_cuda_graph && cuda_graph_update_required) {
cuda_ctx->cuda_graph->number_consecutive_updates++;
} else {
cuda_ctx->cuda_graph->number_consecutive_updates = 0;
}
if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
#endif
}
}
}
if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
}
static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
#else
bool use_cuda_graph = false;
bool cuda_graph_update_required = false;
#endif // USE_CUDA_GRAPH
bool graph_evaluated_or_captured = false;
if (cuda_ctx->cuda_graph->instance == nullptr) {
cuda_graph_update_required = true;
}
// Check if the graph size has changed
if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
cuda_graph_update_required = true;
cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
}
// Loop over nodes in GGML graph to determine if CUDA graph update is required
// and store properties to allow this comparison for the next token
for (int i = 0; i < cgraph->n_nodes; i++) {
bool has_matching_properties = true;
if (!cuda_graph_update_required) {
has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
}
if (!has_matching_properties) {
cuda_graph_update_required = true;
}
set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
}
return cuda_graph_update_required;
}
static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
cudaGraphExecUpdateResultInfo result_info;
cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
if (stat == cudaErrorGraphExecUpdateFailure) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: CUDA graph update failed\n", __func__);
#endif
// The pre-existing graph exec cannot be updated due to violated constraints
// so instead clear error and re-instantiate
cudaGetLastError();
CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
cuda_ctx->cuda_graph->instance = nullptr;
CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
} else {
GGML_ASSERT(stat == cudaSuccess);
}
}
#endif
static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
[[maybe_unused]] std::vector<void *> & ggml_cuda_cpy_fn_ptrs, bool & graph_evaluated_or_captured, bool & use_cuda_graph,
bool & cuda_graph_update_required) {
while (!graph_evaluated_or_captured) {
// Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
@ -2515,19 +2535,8 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
CUDA_CHECK(cudaGraphDestroy(cuda_ctx->cuda_graph->graph));
cuda_ctx->cuda_graph->graph = nullptr;
}
CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
#if 0
if (disable_cuda_graphs_due_to_failed_capture) {
use_cuda_graph = false;
cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
#endif
} else {
graph_evaluated_or_captured = true; // CUDA graph has been captured
}
#endif
CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
graph_evaluated_or_captured = true; // CUDA graph has been captured
} else {
graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
@ -2540,72 +2549,91 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
}
// Perform update to graph (if required for this token), and change copy parameter (required for every token)
if (cuda_graph_update_required) {
// Extract nodes from graph
// First call with null argument gets number of nodes in graph
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
// Subsequent call with non-null argument gets nodes
cuda_ctx->cuda_graph->nodes.clear();
cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
cuda_ctx->cuda_graph->params.clear();
cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
if (cuda_ctx->cuda_graph->num_nodes > 0) {
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));
// Loop over nodes, and extract kernel parameters from each node
for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
cudaGraphNodeType node_type;
CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type));
if (node_type == cudaGraphNodeTypeKernel) {
cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime
if (stat == cudaErrorInvalidDeviceFunction) {
// Fails due to incorrect handling by CUDA runtime of CUDA BLAS node.
// We don't need to update blas nodes, so clear error and move on.
cudaGetLastError();
} else {
GGML_ASSERT(stat == cudaSuccess);
}
}
}
}
}
// One of the arguments to the copy kernel is updated for each token, hence we need to
// replace that argument with the updated value in the CUDA graph
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(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]));
}
}
}
maintain_cuda_graph(cuda_ctx, ggml_cuda_cpy_fn_ptrs, cuda_graph_update_required);
// Update graph executable
cudaGraphExecUpdateResultInfo result_info;
cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
if (stat == cudaErrorGraphExecUpdateFailure) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: CUDA graph update failed\n", __func__);
#endif
// The pre-existing graph exec cannot be updated due to violated constraints
// so instead clear error and re-instantiate
cudaGetLastError();
CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
cuda_ctx->cuda_graph->instance = nullptr;
CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
} else {
GGML_ASSERT(stat == cudaSuccess);
}
update_cuda_graph_executable(cuda_ctx);
// Launch graph
CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
#else
graph_evaluated_or_captured = true;
#endif // USE_CUDA_GRAPH
#endif // USE_CUDA_GRAPH
}
}
static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
ggml_cuda_set_device(cuda_ctx->device);
// vector of pointers to CUDA cpy kernels, which are required to identify
// kernel parameters which need updated in the graph for each token
std::vector<void *> ggml_cuda_cpy_fn_ptrs;
#ifdef USE_CUDA_GRAPH
static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
// Objects required for CUDA Graph
if (cuda_ctx->cuda_graph == nullptr) {
cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
}
bool use_cuda_graph = true;
bool cuda_graph_update_required = false;
if (cuda_ctx->cuda_graph->graph == nullptr) {
if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_AMPERE) {
cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
#endif
}
}
// Disable CUDA graphs in presence of env var, old GPU, use-case which is changing too rapidly,
// or previous graph capture failure.
// Also disable for multi-gpu for now. TO DO investigate
if (disable_cuda_graphs_due_to_env
|| cuda_ctx->cuda_graph->disable_due_to_gpu_arch
|| cuda_ctx->cuda_graph->disable_due_to_too_many_updates
|| cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture) {
use_cuda_graph = false;
}
if (use_cuda_graph) {
cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph,
ggml_cuda_cpy_fn_ptrs, use_cuda_graph);
// Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
if (use_cuda_graph && cuda_graph_update_required) {
cuda_ctx->cuda_graph->number_consecutive_updates++;
} else {
cuda_ctx->cuda_graph->number_consecutive_updates = 0;
}
if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
#endif
}
}
if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
}
#else
bool use_cuda_graph = false;
bool cuda_graph_update_required = false;
#endif // USE_CUDA_GRAPH
bool graph_evaluated_or_captured = false;
evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, ggml_cuda_cpy_fn_ptrs, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);
return GGML_STATUS_SUCCESS;
}
@ -2869,6 +2897,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_BF16:
#ifdef GGML_USE_MUSA
if (a->type == GGML_TYPE_Q3_K) {
return false;
@ -3010,6 +3039,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_LEAKY_RELU:
case GGML_OP_RWKV_WKV6:
case GGML_OP_GATED_LINEAR_ATTN:
return true;
case GGML_OP_FLASH_ATTN_EXT: {
#ifndef FLASH_ATTN_AVAILABLE

93
ggml/src/ggml-cuda/gla.cu Normal file
View File

@ -0,0 +1,93 @@
#include "common.cuh"
#include "gla.cuh"
template<int HEAD_SIZE>
static __global__ void gated_linear_attn_f32(const int B, const int T, const int C, const int H, const float scale,
const float * k, const float * v, const float * r, const float * td, const float * s, float * dst) {
const int tid = threadIdx.x;
const int bid = blockIdx.x;
const int head_size = HEAD_SIZE;
const int batch_i = bid / H;
const int head_i = bid % H;
const int state_size = C * head_size;
const int n_seq_tokens = T / B;
float state[head_size];
__shared__ float _k[head_size], _r[head_size], _td[head_size];
#pragma unroll
for (int i = 0; i < head_size; i++) {
state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
}
for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
__syncthreads();
_k[tid] = k[t];
_r[tid] = r[t];
_td[tid] = td[t];
__syncthreads();
const float _v = v[t];
float y = 0;
for (int j = 0; j < head_size; j += 4) {
const float4 & k = (float4 &)(_k[j]);
const float4 & r = (float4 &)(_r[j]);
const float4 & td = (float4 &)(_td[j]);
float4 & s = (float4 &)(state[j]);
float4 kv;
kv.x = k.x * _v;
kv.y = k.y * _v;
kv.z = k.z * _v;
kv.w = k.w * _v;
s.x = s.x * td.x + kv.x;
s.y = s.y * td.y + kv.y;
s.z = s.z * td.z + kv.z;
s.w = s.w * td.w + kv.w;
y += r.x * s.x;
y += r.y * s.y;
y += r.z * s.z;
y += r.w * s.w;
}
dst[t] = y * scale;
}
#pragma unroll
for (int i = 0; i < head_size; i++) {
dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
}
}
void ggml_cuda_op_gated_linear_attn(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const float * k_d = (const float *)dst->src[0]->data;
const float * v_d = (const float *)dst->src[1]->data;
const float * r_d = (const float *)dst->src[2]->data;
const float * td_d = (const float *)dst->src[3]->data;
const float * s_d = (const float *)dst->src[4]->data;
const int64_t B = dst->src[4]->ne[1];
const int64_t T = dst->src[0]->ne[2];
const int64_t C = dst->ne[0];
const int64_t H = dst->src[0]->ne[1];
float scale;
memcpy(&scale, (float*)dst->op_params, sizeof(float));
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(dst->src[4]->type == GGML_TYPE_F32);
GGML_ASSERT(C % H == 0);
GGML_ASSERT(C / H == 64 || C / H == 128);
if (C / H == 64) {
gated_linear_attn_f32<64><<<B * H, C / H, 0, stream>>>(B, T, C, H, scale, k_d, v_d, r_d, td_d, s_d, dst_d);
} else {
gated_linear_attn_f32<128><<<B * H, C / H, 0, stream>>>(B, T, C, H, scale, k_d, v_d, r_d, td_d, s_d, dst_d);
}
}

3
ggml/src/ggml-cuda/gla.cuh Normal file
View File

@ -0,0 +1,3 @@
#include "common.cuh"
void ggml_cuda_op_gated_linear_attn(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

114
ggml/src/ggml-cuda/mmv.cu
View File

@ -1,9 +1,9 @@
#include "common.cuh"
#include "mmv.cuh"
template <typename type_acc, int block_size>
template <typename T, typename type_acc, int block_size>
static __global__ void mul_mat_vec(
const half * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
const T * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst) {
const int64_t row = blockIdx.x;
const int64_t channel = blockIdx.z;
@ -13,7 +13,6 @@ static __global__ void mul_mat_vec(
y += channel *stride_channel_y;
dst += channel *stride_channel_dst;
const half2 * x2 = (const half2 *) x;
const float2 * y2 = (const float2 *) y;
extern __shared__ char data_mmv[];
@ -28,28 +27,44 @@ static __global__ void mul_mat_vec(
float sumf;
if (std::is_same<type_acc, float>::value) {
if constexpr (std::is_same<T, half>::value) {
const half2 * x2 = (const half2 *) x;
if (std::is_same<type_acc, float>::value) {
sumf = 0.0f;
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const float2 tmpx = __half22float2(x2[col2]);
const float2 tmpy = y2[col2];
sumf += tmpx.x * tmpy.x;
sumf += tmpx.y * tmpy.y;
}
} else {
#ifdef FP16_AVAILABLE
half2 sumh2 = make_half2(0.0f, 0.0f);
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const float2 tmp = y2[col2];
sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
}
sumf = __low2float(sumh2) + __high2float(sumh2);
#else
NO_DEVICE_CODE;
#endif // FP16_AVAILABLE
}
} else if constexpr (std::is_same<T, nv_bfloat16>::value) {
const int * x2 = (const int *) x;
sumf = 0.0f;
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const float2 tmpx = __half22float2(x2[col2]);
const int tmpx = x2[col2];
const float2 tmpy = y2[col2];
sumf += tmpx.x * tmpy.x;
sumf += tmpx.y * tmpy.y;
sumf += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
sumf += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
}
} else {
#ifdef FP16_AVAILABLE
half2 sumh2 = make_half2(0.0f, 0.0f);
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const float2 tmp = y2[col2];
sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
}
sumf = __low2float(sumh2) + __high2float(sumh2);
#else
NO_DEVICE_CODE;
#endif // FP16_AVAILABLE
static_assert(std::is_same<T, void>::value, "unsupported type");
}
sumf = warp_reduce_sum(sumf);
@ -71,9 +86,9 @@ static __global__ void mul_mat_vec(
dst[row] = sumf;
}
template <typename type_acc>
template <typename T, typename type_acc>
static void launch_mul_mat_vec_cuda(
const half * x, const float * y, float * dst,
const T * x, const float * y, float * dst,
const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
cudaStream_t stream) {
@ -97,35 +112,35 @@ static void launch_mul_mat_vec_cuda(
const dim3 block_dims(block_size_best, 1, 1);
switch (block_size_best) {
case 32: {
mul_mat_vec<type_acc, 32><<<block_nums, block_dims, smem, stream>>>
mul_mat_vec<T, type_acc, 32><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 64: {
mul_mat_vec<type_acc, 64><<<block_nums, block_dims, smem, stream>>>
mul_mat_vec<T, type_acc, 64><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 96: {
mul_mat_vec<type_acc, 96><<<block_nums, block_dims, smem, stream>>>
mul_mat_vec<T, type_acc, 96><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 128: {
mul_mat_vec<type_acc, 128><<<block_nums, block_dims, smem, stream>>>
mul_mat_vec<T, type_acc, 128><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 160: {
mul_mat_vec<type_acc, 160><<<block_nums, block_dims, smem, stream>>>
mul_mat_vec<T, type_acc, 160><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 192: {
mul_mat_vec<type_acc, 192><<<block_nums, block_dims, smem, stream>>>
mul_mat_vec<T, type_acc, 192><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 224: {
mul_mat_vec<type_acc, 224><<<block_nums, block_dims, smem, stream>>>
mul_mat_vec<T, type_acc, 224><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 256: {
mul_mat_vec<type_acc, 256><<<block_nums, block_dims, smem, stream>>>
mul_mat_vec<T, type_acc, 256><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
default: {
@ -134,25 +149,25 @@ static void launch_mul_mat_vec_cuda(
}
}
template<typename T>
static void mul_mat_vec_cuda(
const half * x, const float * y, float * dst,
const T * x, const float * y, float * dst,
const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
enum ggml_prec prec, cudaStream_t stream) {
switch (prec) {
case GGML_PREC_DEFAULT: {
launch_mul_mat_vec_cuda<half>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
launch_mul_mat_vec_cuda<T, half>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
stride_channel_x, stride_channel_y, stride_channel_dst, stream);
} break;
case GGML_PREC_F32: {
launch_mul_mat_vec_cuda<float>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
launch_mul_mat_vec_cuda<T, float>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
stride_channel_x, stride_channel_y, stride_channel_dst, stream);
} break;
}
}
void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
@ -164,7 +179,6 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
const half * src0_d = (const half *) src0->data;
const float * src1_d = (const float *) src1->data;
float * dst_d = (float *) dst->data;
@ -181,7 +195,20 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
const int64_t channel_stride_y = src1->nb[2] / ggml_type_size(src1->type);
const int64_t channel_stride_dst = dst->nb[2] / ggml_type_size( dst->type);
mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, stride_row, ne02, ne12, channel_stride_x, channel_stride_y, channel_stride_dst, prec, ctx.stream());
switch (src0->type) {
case GGML_TYPE_F16: {
const half * src0_d = (const half *) src0->data;
mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, stride_row, ne02, ne12,
channel_stride_x, channel_stride_y, channel_stride_dst, prec, ctx.stream());
} break;
case GGML_TYPE_BF16: {
const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, stride_row, ne02, ne12,
channel_stride_x, channel_stride_y, channel_stride_dst, prec, ctx.stream());
} break;
default:
GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
}
}
void ggml_cuda_op_mul_mat_vec(
@ -190,7 +217,6 @@ void ggml_cuda_op_mul_mat_vec(
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
const int64_t src1_padded_row_size, cudaStream_t stream) {
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
@ -211,8 +237,20 @@ void ggml_cuda_op_mul_mat_vec(
const int64_t channel_stride_y = 0;
const int64_t channel_stride_dst = 0;
mul_mat_vec_cuda((const half *) src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, channel_stride_x, channel_stride_y, channel_stride_dst, prec, stream);
switch (src0->type) {
case GGML_TYPE_F16: {
const half * src0_d = (const half *) src0_dd_i;
mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, channel_stride_x, channel_stride_y, channel_stride_dst, prec, stream);
} break;
case GGML_TYPE_BF16: {
const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, channel_stride_x, channel_stride_y, channel_stride_dst, prec, stream);
} break;
default:
GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
}
GGML_UNUSED(ctx);
GGML_UNUSED(src1);

1
ggml/src/ggml-cuda/vendors/cuda.h vendored
View File

@ -3,6 +3,7 @@
#include <cuda_runtime.h>
#include <cuda.h>
#include <cublas_v2.h>
#include <cuda_bf16.h>
#include <cuda_fp16.h>
#if CUDART_VERSION < 11020

3
ggml/src/ggml-cuda/vendors/hip.h vendored
View File

@ -3,6 +3,7 @@
#include <hip/hip_runtime.h>
#include <hipblas/hipblas.h>
#include <hip/hip_fp16.h>
#include <hip/hip_bfloat16.h>
#ifdef __HIP_PLATFORM_AMD__
// for rocblas_initialize()
#include "rocblas/rocblas.h"
@ -121,6 +122,8 @@
#define __has_builtin(x) 0
#endif
typedef hip_bfloat16 nv_bfloat16;
typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
static __device__ __forceinline__ int __vsubss4(const int a, const int b) {

3
ggml/src/ggml-cuda/vendors/musa.h vendored
View File

@ -3,6 +3,7 @@
#include <musa_runtime.h>
#include <musa.h>
#include <mublas.h>
#include <musa_bf16.h>
#include <musa_fp16.h>
#define CUBLAS_COMPUTE_16F CUDA_R_16F
#define CUBLAS_COMPUTE_32F CUDA_R_32F
@ -132,3 +133,5 @@
#define cudaKernelNodeParams musaKernelNodeParams
#define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
#define cudaStreamEndCapture musaStreamEndCapture
typedef mt_bfloat16 nv_bfloat16;

Some files were not shown because too many files have changed in this diff Show More