root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 20:04:35 +00:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/bitnet-ternary

Browse Source
This commit is contained in:
Francis Couture-Harpin 2024-08-13 17:36:09 -04:00
commit 82b240406d
22 changed files with 75 additions and 58 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -129,6 +129,8 @@ jobs:
- name: Server bench
id: server_bench
env:
HEAD_REF: ${{ github.head_ref || github.ref_name }}
run: |
set -eux
@ -137,7 +139,7 @@ jobs:
python bench.py \
--runner-label ${{ env.RUNNER_LABEL }} \
--name ${{ github.job }} \
--branch ${{ github.head_ref || github.ref_name }} \
--branch $HEAD_REF \
--commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
--scenario script.js \
--duration ${{ github.event.inputs.duration || env.DURATION }} \

22
.github/workflows/build.yml vendored
View File

@ -47,7 +47,7 @@ jobs:
sysctl -a
mkdir build
cd build
cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -105,7 +105,7 @@ jobs:
sysctl -a
# Metal is disabled due to intermittent failures with Github runners not having a GPU:
# https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -222,7 +222,7 @@ jobs:
run: |
mkdir build
cd build
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
cmake --build . --config Release -j $(nproc)
- name: Test
@ -696,22 +696,20 @@ jobs:
strategy:
matrix:
include:
- build: 'rpc-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
- build: 'noavx-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
- build: 'avx2-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
- build: 'avx-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
- build: 'avx512-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
- build: 'openblas-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
- build: 'kompute-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=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 -DBUILD_SHARED_LIBS=ON'
- build: 'vulkan-x64'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=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'
- build: 'msvc-arm64'

6
.github/workflows/python-check-requirements.yml vendored
View File

@ -6,15 +6,13 @@ on:
- '.github/workflows/python-check-requirements.yml'
- 'scripts/check-requirements.sh'
- 'convert*.py'
- 'requirements.txt'
- 'requirements/*.txt'
- '**/requirements*.txt'
pull_request:
paths:
- '.github/workflows/python-check-requirements.yml'
- 'scripts/check-requirements.sh'
- 'convert*.py'
- 'requirements.txt'
- 'requirements/*.txt'
- '**/requirements*.txt'
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

2
README.md
View File

@ -186,10 +186,12 @@ Unless otherwise noted these projects are open-source with permissive licensing:
- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
**Infrastructure:**
- [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
**Games:**
- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.

3
common/grammar-parser.cpp
View File

@ -369,6 +369,9 @@ namespace grammar_parser {
}
// Validate the state to ensure that all rules are defined
for (const auto & rule : state.rules) {
if (rule.empty()) {
throw std::runtime_error("Undefined rule");
}
for (const auto & elem : rule) {
if (elem.type == LLAMA_GRETYPE_RULE_REF) {
// Ensure that the rule at that location exists

6
examples/export-lora/README.md
View File

@ -17,9 +17,9 @@ For example:
```bash
./bin/llama-export-lora \
-m open-llama-3b-v2-q8_0.gguf \
-o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
--lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.gguf
-m open-llama-3b-v2.gguf \
-o open-llama-3b-v2-english2tokipona-chat.gguf \
--lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
```
Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:

35
examples/export-lora/export-lora.cpp
View File

@ -10,6 +10,12 @@
static bool g_verbose = false;
struct tensor_transformation {
struct ggml_tensor * in;
struct ggml_tensor * out;
bool is_copy;
};
static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
int id = gguf_find_key(ctx_gguf, key.c_str());
return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
@ -198,8 +204,7 @@ struct lora_merge_ctx {
}
// mapping base tensor to out tensor (same shape with base, but different type)
// if out_tensor == nullptr, we only copy it
std::vector<std::pair<struct ggml_tensor *, struct ggml_tensor *>> base_to_out_tensors;
std::vector<tensor_transformation> trans;
for (auto & it : base_model.tensors) {
bool t_a = true;
bool t_b = true;
@ -212,14 +217,22 @@ struct lora_merge_ctx {
// only copy
struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
ggml_set_name(cpy_tensor, base_tensor->name);
base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
trans.push_back({
cpy_tensor,
cpy_tensor,
true,
});
gguf_add_tensor(ctx_out, cpy_tensor);
} else if (t_a && t_b) {
// need merging
struct ggml_tensor * out_tensor = ggml_new_tensor(
ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
ggml_set_name(out_tensor, base_tensor->name);
base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
trans.push_back({
base_tensor,
out_tensor,
false,
});
gguf_add_tensor(ctx_out, out_tensor);
} else {
throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
@ -234,12 +247,12 @@ struct lora_merge_ctx {
// process base model tensors
size_t n_merged = 0;
for (auto & it : base_to_out_tensors) {
if (it.second != nullptr) {
merge_tensor(it.first, it.second);
for (auto & it : trans) {
if (!it.is_copy) {
merge_tensor(it.in, it.out);
n_merged++;
} else {
copy_tensor(it.first);
copy_tensor(it.in);
}
}
@ -252,7 +265,7 @@ struct lora_merge_ctx {
}
printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
}
void copy_tensor(struct ggml_tensor * base) {
@ -285,6 +298,10 @@ struct lora_merge_ctx {
for (size_t i = 0; i < adapters.size(); ++i) {
auto t_a = adapters[i]->get_tensor(name_lora_a);
auto t_b = adapters[i]->get_tensor(name_lora_b);
// TODO: add support for quantized lora
if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) {
throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32");
}
inp_a[i] = ggml_dup_tensor(ctx, t_a);
inp_b[i] = ggml_dup_tensor(ctx, t_b);
}

2
examples/llava/requirements.txt
View File

@ -2,4 +2,4 @@
--extra-index-url https://download.pytorch.org/whl/cpu
pillow~=10.2.0
torch~=2.2.1
torchvision==0.17.1
torchvision~=0.17.1

5
examples/server/server.cpp
View File

@ -631,6 +631,7 @@ struct server_context {
bool clean_kv_cache = true;
bool add_bos_token = true;
bool has_eos_token = false;
int32_t n_ctx; // total context for all clients / slots
@ -693,7 +694,7 @@ struct server_context {
n_ctx = llama_n_ctx(ctx);
add_bos_token = llama_should_add_bos_token(model);
GGML_ASSERT(llama_add_eos_token(model) != 1);
has_eos_token = llama_add_eos_token(model) != 1;
return true;
}
@ -1031,7 +1032,7 @@ struct server_context {
{
slot.sparams.logit_bias.clear();
if (json_value(data, "ignore_eos", false)) {
if (json_value(data, "ignore_eos", false) && has_eos_token) {
slot.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
}

6
ggml/include/ggml.h
View File

@ -244,6 +244,8 @@
#define GGML_EXIT_SUCCESS 0
#define GGML_EXIT_ABORTED 1
#define GGML_ROPE_TYPE_NEOX 2
#define GGUF_MAGIC "GGUF"
#define GGUF_VERSION 3
@ -1455,8 +1457,8 @@ extern "C" {
struct ggml_tensor * b);
// rotary position embedding
// if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
// if mode & 2 == 1, GPT-NeoX style
// if (mode & 1) - skip n_past elements (NOT SUPPORTED)
// if (mode & GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
//
// b is an int32 vector with size a->ne[2], it contains the positions
GGML_API struct ggml_tensor * ggml_rope(

2
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -2881,7 +2881,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast,
beta_slow, corr_dims);
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
// init cos/sin cache
ggml_cann_pool_alloc sin_allocator(

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

@ -226,7 +226,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const int32_t * pos = (const int32_t *) src1_d;

2
ggml/src/ggml-metal.m
View File

@ -2313,7 +2313,7 @@ static enum ggml_status ggml_metal_graph_compute(
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
id<MTLComputePipelineState> pipeline = nil;

2
ggml/src/ggml-sycl/rope.cpp
View File

@ -226,7 +226,7 @@ void ggml_sycl_op_rope(
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const int32_t * pos = (const int32_t *) src1_dd;

2
ggml/src/ggml-vulkan.cpp
View File

@ -4053,7 +4053,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
case GGML_OP_ROPE:
{
const int mode = ((const int32_t *) dst->op_params)[2];
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
if (is_neox) {
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {

6
ggml/src/ggml.c
View File

@ -14131,7 +14131,7 @@ static void ggml_compute_forward_rope_f32(
float corr_dims[2];
ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const float * freq_factors = NULL;
if (src2 != NULL) {
@ -14256,7 +14256,7 @@ static void ggml_compute_forward_rope_f16(
float corr_dims[2];
ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const float * freq_factors = NULL;
if (src2 != NULL) {
@ -21168,7 +21168,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
(int64_t) info->ne[2] *
(int64_t) info->ne[3];
if (ne % ggml_blck_size(info->type) != 0) {
if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
__func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
fclose(file);

2
ggml/src/kompute-shaders/op_rope_f16.comp
View File

@ -11,7 +11,7 @@ void main() {
const uint i2 = gl_WorkGroupID.y;
const uint i1 = gl_WorkGroupID.x;
const bool is_neox = (pcs.mode & 2) != 0;
const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;
float corr_dims[2];
rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

2
ggml/src/kompute-shaders/op_rope_f32.comp
View File

@ -11,7 +11,7 @@ void main() {
const uint i2 = gl_WorkGroupID.y;
const uint i1 = gl_WorkGroupID.x;
const bool is_neox = (pcs.mode & 2) != 0;
const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;
float corr_dims[2];
rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

2
ggml/src/kompute-shaders/rope_common.comp
View File

@ -1,5 +1,7 @@
#include "common.comp"
#define GGML_ROPE_TYPE_NEOX 2
// TODO: use a local size of 32 or more (Metal uses 1024)
layout(local_size_x = 1) in;

7
include/llama.h
View File

@ -95,13 +95,10 @@ extern "C" {
LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22,
};
// note: these values should be synchronized with ggml_rope
// TODO: maybe move this enum to ggml.h (ggml_rope_type)
enum llama_rope_type {
LLAMA_ROPE_TYPE_NONE = -1,
LLAMA_ROPE_TYPE_NORM = 0,
LLAMA_ROPE_TYPE_NEOX = 2,
LLAMA_ROPE_TYPE_GLM = 4,
LLAMA_ROPE_TYPE_NORM = 0,
LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX,
};
enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file

4
src/llama-sampling.cpp
View File

@ -85,14 +85,14 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra
constexpr float bucket_low = -10.0f;
constexpr float bucket_high = 10.0f;
constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
constexpr float bucker_inter = -bucket_low * bucket_scale;
constexpr float bucket_inter = -bucket_low * bucket_scale;
std::vector<int> bucket_idx(candidates->size);
std::vector<int> histo(nbuckets, 0);
for (int i = 0; i < (int)candidates->size; ++i) {
const float val = candidates->data[i].logit;
int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
ib = std::max(0, std::min(nbuckets-1, ib));
bucket_idx[i] = ib;
++histo[ib];

9
src/llama.cpp
View File

@ -3575,13 +3575,8 @@ namespace GGUFMeta {
using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
// TODO: update when needed or think of some clever automatic way to do this
static size_t llama_model_max_nodes(const llama_model & /*model*/) {
//if (model.arch == LLM_ARCH_LLAMA && model.hparams.n_layer > ??) { // llama-3 405B
// return 32768;
//}
return 8192;
static size_t llama_model_max_nodes(const llama_model & model) {
return std::max<size_t>(8192, model.tensors_by_name.size()*5);
}
struct llama_model_loader {