root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 02:14:35 +00:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

llama : ggml-backend integration (#4766)

Browse Source 
* llama : ggml-backend integration

* ggml-backend : add names to buffers

* fix unmap after loading

* batched-bench : add tensor_split param

* llama : check for null tensor_split

* ggml-backend : increase GGML_MAX_BACKENDS

* improve graph splitting, partial fix for --no-kv-offload

* cuda : add ggml-backend split buffer support

* cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available)

* ggml : fix null backend dereference (#4807)

* ggml : fix null backend dereference

* ggml : also check ggml_backend_is_cpu

* test-backend-ops : check buffer allocation failures

* llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row)

* ggml : fix mul_mat_id work size

* llama : rewrite session kv load/set without graphs

* minor

* llama : only initialize used backends, free backends on context free

* llama : abort ctx if cuda backend init fails

* llama : rewrite lora with ggml-backend and compute on CPU

ggml-ci

* llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer

* opencl : add ggml-backend buffer type

* cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf)

* llama : on Metal, by default offload the full model

ggml-ci

* metal : page align the data ptr (#4854)

* Apply suggestions from code review

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* cuda : fix split buffer free

* address review comments

* llama-bench : add split-mode parameter

* fix whitespace

* opencl : fix double initialization

* server : add --split-mode parameter

* use async copy and compute to improve multi-gpu performance

ggml-ci

* use async memcpys to copy the graph outputs to the CPU

* fix opencl

* use a host buffer for the cpu compute buffer for faster copies to the gpu

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
slaren 2024-01-12 20:07:38 +01:00 committed by GitHub
parent 584d674be6
commit e7e4df031b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
21 changed files with 2533 additions and 2295 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

65
common/common.cpp
View File

@ -543,9 +543,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
invalid_param = true;
break;
}
#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
params.n_gpu_layers = std::stoi(argv[i]);
#else
#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
#endif
@ -554,9 +553,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
invalid_param = true;
break;
}
#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
params.n_gpu_layers_draft = std::stoi(argv[i]);
#else
#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
#endif
@ -565,25 +563,44 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
invalid_param = true;
break;
}
#ifdef GGML_USE_CUBLAS
params.main_gpu = std::stoi(argv[i]);
#else
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
#endif
#ifndef GGML_USE_CUBLAS
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the main GPU has no effect.\n");
#endif // GGML_USE_CUBLAS
} else if (arg == "--split-mode" || arg == "-sm") {
if (++i >= argc) {
invalid_param = true;
break;
}
std::string arg_next = argv[i];
if (arg_next == "none") {
params.split_mode = LLAMA_SPLIT_NONE;
} else if (arg_next == "layer") {
params.split_mode = LLAMA_SPLIT_LAYER;
} else if (arg_next == "row") {
params.split_mode = LLAMA_SPLIT_ROW;
} else {
invalid_param = true;
break;
}
#ifndef GGML_USE_CUBLAS
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
#endif // GGML_USE_CUBLAS
} else if (arg == "--tensor-split" || arg == "-ts") {
if (++i >= argc) {
invalid_param = true;
break;
}
#ifdef GGML_USE_CUBLAS
std::string arg_next = argv[i];
// split string by , and /
const std::regex regex{R"([,/]+)"};
std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
std::vector<std::string> split_arg{it, {}};
GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
if (split_arg.size() >= LLAMA_MAX_DEVICES) {
invalid_param = true;
break;
}
for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
if (i < split_arg.size()) {
params.tensor_split[i] = std::stof(split_arg[i]);
@ -591,14 +608,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
params.tensor_split[i] = 0.0f;
}
}
#else
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
#endif // GGML_USE_CUBLAS
} else if (arg == "--no-mul-mat-q" || arg == "-nommq") {
#ifdef GGML_USE_CUBLAS
params.mul_mat_q = false;
#else
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
#ifndef GGML_USE_CUBLAS
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting a tensor split has no effect.\n");
#endif // GGML_USE_CUBLAS
} else if (arg == "--no-mmap") {
params.use_mmap = false;
@ -915,14 +926,15 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" number of layers to store in VRAM\n");
printf(" -ngld N, --n-gpu-layers-draft N\n");
printf(" number of layers to store in VRAM for the draft model\n");
printf(" -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
printf(" how to split the model across multiple GPUs, one of:\n");
printf(" - none: use one GPU only\n");
printf(" - layer (default): split layers and KV across GPUs\n");
printf(" - row: split rows across GPUs\n");
printf(" -ts SPLIT, --tensor-split SPLIT\n");
printf(" how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
printf(" -mg i, --main-gpu i the GPU to use for scratch and small tensors\n");
#ifdef GGML_USE_CUBLAS
printf(" -nommq, --no-mul-mat-q\n");
printf(" use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
printf(" Not recommended since this is both slower and uses more VRAM.\n");
#endif // GGML_USE_CUBLAS
printf(" fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
printf(" -mg i, --main-gpu i the GPU to use for the model (with split-mode = none),\n");
printf(" or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
#endif
printf(" -gan N, --grp-attn-n N\n");
printf(" group-attention factor (default: %d)\n", params.grp_attn_n);
@ -1041,6 +1053,7 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
mparams.n_gpu_layers = params.n_gpu_layers;
}
mparams.main_gpu = params.main_gpu;
mparams.split_mode = params.split_mode;
mparams.tensor_split = params.tensor_split;
mparams.use_mmap = params.use_mmap;
mparams.use_mlock = params.use_mlock;

1
common/common.h
View File

@ -59,6 +59,7 @@ struct gpt_params {
float p_split = 0.1f; // speculative decoding split probability
int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default)
int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
llama_split_mode split_mode = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
float tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
int32_t n_beams = 0; // if non-zero then use beam search of given width.

3
examples/batched-bench/batched-bench.cpp
View File

@ -88,7 +88,10 @@ int main(int argc, char ** argv) {
llama_model_params model_params = llama_model_default_params();
const std::vector<float> t_split (LLAMA_MAX_DEVICES, 0.0f);
model_params.n_gpu_layers = n_gpu_layers;
model_params.tensor_split = t_split.data();
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);

146
examples/llama-bench/llama-bench.cpp
View File

@ -128,6 +128,25 @@ static std::string get_gpu_info() {
// command line params
enum output_formats {CSV, JSON, MARKDOWN, SQL};
static const char * output_format_str(output_formats format) {
switch (format) {
case CSV: return "csv";
case JSON: return "json";
case MARKDOWN: return "md";
case SQL: return "sql";
default: GGML_ASSERT(!"invalid output format");
}
}
static const char * split_mode_str(llama_split_mode mode) {
switch (mode) {
case LLAMA_SPLIT_NONE: return "none";
case LLAMA_SPLIT_LAYER: return "layer";
case LLAMA_SPLIT_ROW: return "row";
default: GGML_ASSERT(!"invalid split mode");
}
}
struct cmd_params {
std::vector<std::string> model;
std::vector<int> n_prompt;
@ -137,6 +156,7 @@ struct cmd_params {
std::vector<ggml_type> type_v;
std::vector<int> n_threads;
std::vector<int> n_gpu_layers;
std::vector<llama_split_mode> split_mode;
std::vector<int> main_gpu;
std::vector<bool> no_kv_offload;
std::vector<bool> mul_mat_q;
@ -155,6 +175,7 @@ static const cmd_params cmd_params_defaults = {
/* type_v */ {GGML_TYPE_F16},
/* n_threads */ {get_num_physical_cores()},
/* n_gpu_layers */ {99},
/* split_mode */ {LLAMA_SPLIT_LAYER},
/* main_gpu */ {0},
/* no_kv_offload */ {false},
/* mul_mat_q */ {true},
@ -169,21 +190,22 @@ static void print_usage(int /* argc */, char ** argv) {
printf("\n");
printf("options:\n");
printf(" -h, --help\n");
printf(" -m, --model <filename> (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
printf(" -p, --n-prompt <n> (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
printf(" -n, --n-gen <n> (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
printf(" -b, --batch-size <n> (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
printf(" -ctk <t>, --cache-type-k <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
printf(" -ctv <t>, --cache-type-v <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
printf(" -t, --threads <n> (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
printf(" -ngl, --n-gpu-layers <n> (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
printf(" -mg, --main-gpu <i> (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
printf(" -mmq, --mul-mat-q <0|1> (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
printf(" -ts, --tensor_split <ts0/ts1/..> \n");
printf(" -r, --repetitions <n> (default: %d)\n", cmd_params_defaults.reps);
printf(" -o, --output <csv|json|md|sql> (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
printf(" -v, --verbose (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
printf(" -m, --model <filename> (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
printf(" -p, --n-prompt <n> (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
printf(" -n, --n-gen <n> (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
printf(" -b, --batch-size <n> (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
printf(" -ctk <t>, --cache-type-k <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
printf(" -ctv <t>, --cache-type-v <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
printf(" -t, --threads <n> (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
printf(" -ngl, --n-gpu-layers <n> (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
printf(" -sm, --split-mode <none|layer|row> (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
printf(" -mg, --main-gpu <i> (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
printf(" -mmq, --mul-mat-q <0|1> (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
printf(" -ts, --tensor_split <ts0/ts1/..> (default: 0)\n");
printf(" -r, --repetitions <n> (default: %d)\n", cmd_params_defaults.reps);
printf(" -o, --output <csv|json|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
printf(" -v, --verbose (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
printf("\n");
printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
}
@ -306,6 +328,28 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
}
auto p = split<int>(argv[i], split_delim);
params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
} else if (arg == "-sm" || arg == "--split-mode") {
if (++i >= argc) {
invalid_param = true;
break;
}
auto p = split<std::string>(argv[i], split_delim);
std::vector<llama_split_mode> modes;
for (const auto & m : p) {
llama_split_mode mode;
if (m == "none") {
mode = LLAMA_SPLIT_NONE;
} else if (m == "layer") {
mode = LLAMA_SPLIT_LAYER;
} else if (m == "row") {
mode = LLAMA_SPLIT_ROW;
} else {
invalid_param = true;
break;
}
modes.push_back(mode);
}
params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
} else if (arg == "-mg" || arg == "--main-gpu") {
if (++i >= argc) {
invalid_param = true;
@ -392,6 +436,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
if (params.type_k.empty()) { params.type_k = cmd_params_defaults.type_k; }
if (params.type_v.empty()) { params.type_v = cmd_params_defaults.type_v; }
if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
if (params.split_mode.empty()) { params.split_mode = cmd_params_defaults.split_mode; }
if (params.main_gpu.empty()) { params.main_gpu = cmd_params_defaults.main_gpu; }
if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
if (params.mul_mat_q.empty()) { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
@ -410,6 +455,7 @@ struct cmd_params_instance {
ggml_type type_v;
int n_threads;
int n_gpu_layers;
llama_split_mode split_mode;
int main_gpu;
bool no_kv_offload;
bool mul_mat_q;
@ -419,6 +465,7 @@ struct cmd_params_instance {
llama_model_params mparams = llama_model_default_params();
mparams.n_gpu_layers = n_gpu_layers;
mparams.split_mode = split_mode;
mparams.main_gpu = main_gpu;
mparams.tensor_split = tensor_split.data();
@ -428,6 +475,7 @@ struct cmd_params_instance {
bool equal_mparams(const cmd_params_instance & other) const {
return model == other.model &&
n_gpu_layers == other.n_gpu_layers &&
split_mode == other.split_mode &&
main_gpu == other.main_gpu &&
tensor_split == other.tensor_split;
}
@ -446,45 +494,13 @@ struct cmd_params_instance {
}
};
static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_params & params, int n_gen, int n_prompt) {
std::vector<cmd_params_instance> instances;
for (const auto & m : params.model)
for (const auto & nl : params.n_gpu_layers)
for (const auto & mg : params.main_gpu)
for (const auto & ts : params.tensor_split)
for (const auto & nb : params.n_batch)
for (const auto & tk : params.type_k)
for (const auto & tv : params.type_v)
for (const auto & mmq : params.mul_mat_q)
for (const auto & nkvo : params.no_kv_offload)
for (const auto & nt : params.n_threads) {
cmd_params_instance instance = {
/* .model = */ m,
/* .n_prompt = */ n_prompt,
/* .n_gen = */ n_gen,
/* .n_batch = */ nb,
/* .type_k = */ tk,
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .main_gpu = */ mg,
/* .no_kv_offload= */ nkvo,
/* .mul_mat_q = */ mmq,
/* .tensor_split = */ ts,
};
instances.push_back(instance);
}
return instances;
}
static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
std::vector<cmd_params_instance> instances;
#if 1
// this ordering minimizes the number of times that each model needs to be reloaded
for (const auto & m : params.model)
for (const auto & nl : params.n_gpu_layers)
for (const auto & sm : params.split_mode)
for (const auto & mg : params.main_gpu)
for (const auto & ts : params.tensor_split)
for (const auto & nb : params.n_batch)
@ -506,6 +522,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .split_mode = */ sm,
/* .main_gpu = */ mg,
/* .no_kv_offload= */ nkvo,
/* .mul_mat_q = */ mmq,
@ -527,6 +544,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .split_mode = */ sm,
/* .main_gpu = */ mg,
/* .no_kv_offload= */ nkvo,
/* .mul_mat_q = */ mmq,
@ -535,24 +553,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
instances.push_back(instance);
}
}
#else
// this ordering separates the prompt and generation tests
for (const auto & n_prompt : params.n_prompt) {
if (n_prompt == 0) {
continue;
}
auto instances_prompt = get_cmd_params_instances_int(params, 0, n_prompt);
instances.insert(instances.end(), instances_prompt.begin(), instances_prompt.end());
}
for (const auto & n_gen : params.n_gen) {
if (n_gen == 0) {
continue;
}
auto instances_gen = get_cmd_params_instances_int(params, n_gen, 0);
instances.insert(instances.end(), instances_gen.begin(), instances_gen.end());
}
#endif
return instances;
}
@ -576,6 +576,7 @@ struct test {
ggml_type type_k;
ggml_type type_v;
int n_gpu_layers;
llama_split_mode split_mode;
int main_gpu;
bool no_kv_offload;
bool mul_mat_q;
@ -597,6 +598,7 @@ struct test {
type_k = inst.type_k;
type_v = inst.type_v;
n_gpu_layers = inst.n_gpu_layers;
split_mode = inst.split_mode;
main_gpu = inst.main_gpu;
no_kv_offload = inst.no_kv_offload;
mul_mat_q = inst.mul_mat_q;
@ -660,7 +662,8 @@ struct test {
"cpu_info", "gpu_info",
"model_filename", "model_type", "model_size", "model_n_params",
"n_batch", "n_threads", "type_k", "type_v",
"n_gpu_layers", "main_gpu", "no_kv_offload",
"n_gpu_layers", "split_mode",
"main_gpu", "no_kv_offload",
"mul_mat_q", "tensor_split",
"n_prompt", "n_gen", "test_time",
"avg_ns", "stddev_ns",
@ -711,7 +714,8 @@ struct test {
cpu_info, gpu_info,
model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(no_kv_offload),
std::to_string(n_gpu_layers), split_mode_str(split_mode),
std::to_string(main_gpu), std::to_string(no_kv_offload),
std::to_string(mul_mat_q), tensor_split_str,
std::to_string(n_prompt), std::to_string(n_gen), test_time,
std::to_string(avg_ns()), std::to_string(stdev_ns()),
@ -867,6 +871,9 @@ struct markdown_printer : public printer {
if (field == "n_gpu_layers") {
return "ngl";
}
if (field == "split_mode") {
return "sm";
}
if (field == "n_threads") {
return "threads";
}
@ -907,6 +914,9 @@ struct markdown_printer : public printer {
if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
fields.push_back("main_gpu");
}
if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
fields.push_back("split_mode");
}
if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
fields.push_back("mul_mat_q");
}

40
examples/server/server.cpp
View File

@ -2005,12 +2005,15 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
printf(" -ngl N, --n-gpu-layers N\n");
printf(" number of layers to store in VRAM\n");
printf(" -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
printf(" how to split the model across multiple GPUs, one of:\n");
printf(" - none: use one GPU only\n");
printf(" - layer (default): split layers and KV across GPUs\n");
printf(" - row: split rows across GPUs\n");
printf(" -ts SPLIT --tensor-split SPLIT\n");
printf(" how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
printf(" -mg i, --main-gpu i the GPU to use for scratch and small tensors\n");
printf(" -nommq, --no-mul-mat-q\n");
printf(" use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
printf(" Not recommended since this is both slower and uses more VRAM.\n");
printf(" fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
printf(" -mg i, --main-gpu i the GPU to use for the model (with split-mode = none),\n");
printf(" or for intermediate results and KV (with split-mode = row)\n");
#endif
printf(" -m FNAME, --model FNAME\n");
printf(" model path (default: %s)\n", params.model.c_str());
@ -2253,6 +2256,33 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
"See main README.md for information on enabling GPU BLAS support",
{{"n_gpu_layers", params.n_gpu_layers}});
#endif
}
else if (arg == "--split-mode" || arg == "-sm")
{
if (++i >= argc) {
invalid_param = true;
break;
}
std::string arg_next = argv[i];
if (arg_next == "none")
{
params.split_mode = LLAMA_SPLIT_NONE;
}
else if (arg_next == "layer")
{
params.split_mode = LLAMA_SPLIT_LAYER;
}
else if (arg_next == "row")
{
params.split_mode = LLAMA_SPLIT_ROW;
}
else {
invalid_param = true;
break;
}
#ifndef GGML_USE_CUBLAS
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
#endif // GGML_USE_CUBLAS
}
else if (arg == "--tensor-split" || arg == "-ts")
{

34
ggml-alloc.c
View File

@ -102,8 +102,6 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
}
}
AT_PRINTF("block %d\n", best_fit_block);
if (best_fit_block == -1) {
// the last block is our last resort
struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
@ -117,6 +115,7 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
return;
}
}
struct free_block * block = &alloc->free_blocks[best_fit_block];
void * addr = block->addr;
block->addr = (char*)block->addr + size;
@ -129,6 +128,8 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
}
}
AT_PRINTF("block %d, addr %p\n", best_fit_block, addr);
tensor->data = addr;
tensor->buffer = alloc->buffer;
if (!alloc->measure) {
@ -229,6 +230,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
} else {
alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset;
ggml_backend_buffer_reset(alloc->buffer);
}
}
@ -263,9 +265,9 @@ ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment) {
return alloc;
}
ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) {
ggml_tallocr_t ggml_tallocr_new_measure_from_buft(struct ggml_backend_buffer_type * buft) {
// create a backend buffer to get the correct tensor allocation sizes
ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, 1);
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, 1);
// TODO: move alloc initialization to a common ggml_tallocr_new_impl function
ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
@ -275,13 +277,22 @@ ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backe
return alloc;
}
ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) {
ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, size);
ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) {
return ggml_tallocr_new_measure_from_buft(ggml_backend_get_default_buffer_type(backend));
}
ggml_tallocr_t ggml_tallocr_new_from_buft(struct ggml_backend_buffer_type * buft, size_t size) {
// create a backend buffer to get the correct tensor allocation sizes
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
alloc->buffer_owned = true;
return alloc;
}
ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) {
return ggml_tallocr_new_from_buft(ggml_backend_get_default_buffer_type(backend), size);
}
ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
@ -779,10 +790,21 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
if (nbytes == 0) {
// all the tensors in the context are already allocated
#ifndef NDEBUG
fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
#endif
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
if (buffer == NULL) {
// failed to allocate buffer
#ifndef NDEBUG
fprintf(stderr, "%s: failed to allocate buffer\n", __func__);
#endif
return NULL;
}
ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {

4
ggml-alloc.h
View File

@ -52,8 +52,10 @@ typedef struct ggml_tallocr * ggml_tallocr_t;
GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
GGML_API ggml_tallocr_t ggml_tallocr_new_from_buft(struct ggml_backend_buffer_type * buft, size_t size);
GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_buft(struct ggml_backend_buffer_type * buft);
GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);

38
ggml-backend-impl.h
View File

@ -16,9 +16,10 @@ extern "C" {
typedef void * ggml_backend_buffer_type_context_t;
struct ggml_backend_buffer_type_i {
const char * (*get_name) (ggml_backend_buffer_type_t buft);
ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
size_t (*get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment
size_t (*get_alloc_size) (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
size_t (*get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
bool (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
// check if tensor data is in host memory
// should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
@ -34,16 +35,15 @@ extern "C" {
typedef void * ggml_backend_buffer_context_t;
struct ggml_backend_buffer_i {
void (*free_buffer) (ggml_backend_buffer_t buffer);
//void (*reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
void * (*get_base) (ggml_backend_buffer_t buffer);
void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
// (optional) copy tensor between different buffer-type, allow for single-copy tranfers
void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
void (*cpy_tensor_to) (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
const char * (*get_name) (ggml_backend_buffer_t buffer);
void (*free_buffer)(ggml_backend_buffer_t buffer);
void * (*get_base) (ggml_backend_buffer_t buffer);
void (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
void (*reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
};
struct ggml_backend_buffer {
@ -51,6 +51,7 @@ extern "C" {
ggml_backend_buffer_type_t buft;
ggml_backend_buffer_context_t context;
size_t size;
enum ggml_backend_buffer_usage usage;
};
ggml_backend_buffer_t ggml_backend_buffer_init(
@ -59,6 +60,8 @@ extern "C" {
ggml_backend_buffer_context_t context,
size_t size);
// do not use directly, use ggml_backend_tensor_copy instead
bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
//
// Backend
@ -74,22 +77,20 @@ extern "C" {
// buffer allocation
ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
// (optional) asynchroneous tensor data access
// (optional) asynchronous tensor data access
void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
bool (*cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
// (optional) asynchroneous tensor copy
void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
void (*cpy_tensor_to_async) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
// (optional) complete all pending operations
void (*synchronize)(ggml_backend_t backend);
// compute graph with a plan
ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
// compute graph without a plan
// compute graph without a plan (async)
bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
// check if the backend supports an operation
@ -102,7 +103,6 @@ extern "C" {
ggml_backend_context_t context;
};
//
// Backend registry
//

693
ggml-backend.c
View File

File diff suppressed because it is too large Load Diff

60
ggml-backend.h
View File

@ -17,22 +17,31 @@ extern "C" {
//
// buffer type
GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft);
GGML_API const char * ggml_backend_buft_name (ggml_backend_buffer_type_t buft);
GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer (ggml_backend_buffer_type_t buft, size_t size);
GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
GGML_API size_t ggml_backend_buft_get_alloc_size (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft);
// buffer
GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer);
GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value);
GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer);
GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
enum ggml_backend_buffer_usage {
GGML_BACKEND_BUFFER_USAGE_ANY = 0,
GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
};
GGML_API const char * ggml_backend_buffer_name (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer);
GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value);
GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type (ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_buffer_reset (ggml_backend_buffer_t buffer);
//
// Backend
@ -140,23 +149,24 @@ extern "C" {
typedef struct ggml_backend_sched * ggml_backend_sched_t;
// Initialize a backend scheduler
GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
// Initialize backend buffers from a measure graph
GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
// Get the number of splits of the last graph
GGML_API int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
GGML_API ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
GGML_API ggml_backend_t ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
// Allocate a graph on the backend scheduler
GGML_API void ggml_backend_sched_graph_compute(
ggml_backend_sched_t sched,
struct ggml_cgraph * graph);
// Allocate and compute graph on the backend scheduler
GGML_API void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
// Reset all assignments and allocators - must be called before using the sched allocators to allocate inputs
GGML_API void ggml_backend_sched_reset(ggml_backend_sched_t sched);
//
// Utils
@ -176,7 +186,7 @@ extern "C" {
typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
// Compare the output of two backends
GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
// Tensor initialization
GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

907
ggml-cuda.cu
View File

File diff suppressed because it is too large Load Diff

26
ggml-cuda.h
View File

@ -27,22 +27,6 @@ GGML_API void * ggml_cuda_host_malloc(size_t size);
GGML_API void ggml_cuda_host_free(void * ptr);
GGML_API bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API void ggml_cuda_set_tensor_split(const float * tensor_split);
GGML_API void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
GGML_API void ggml_cuda_free_data(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
GGML_API void ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
GGML_API void ggml_cuda_set_main_device(int main_device);
GGML_API void ggml_cuda_set_mul_mat_q(bool mul_mat_q);
GGML_API void ggml_cuda_set_scratch_size(size_t scratch_size);
GGML_API void ggml_cuda_free_scratch(void);
GGML_API bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
GGML_API int ggml_cuda_get_device_count(void);
@ -52,13 +36,17 @@ GGML_API void ggml_cuda_get_device_description(int device, char * description,
GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
GGML_API int ggml_backend_cuda_get_device(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
// split tensor buffer that splits matrices by rows across multiple devices
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
GGML_API int ggml_backend_cuda_get_device_count(void);
GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
#ifdef __cplusplus
}
#endif

2
ggml-impl.h
View File

@ -228,6 +228,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
#define GGML_HASHTABLE_FULL ((size_t)-1)
#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
struct ggml_hash_set ggml_hash_set_new(size_t size);
bool ggml_hash_contains (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted

55
ggml-metal.m
View File

@ -2520,10 +2520,10 @@ static void ggml_backend_metal_free_device(void) {
}
}
static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
return "Metal";
return ctx->all_data;
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
@ -2541,6 +2541,12 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
free(ctx);
}
static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
return ctx->all_data;
}
static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
memcpy((char *)tensor->data + offset, data, size);
@ -2553,14 +2559,12 @@ static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, c
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
if (ggml_backend_buffer_is_host(src->buffer)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
return true;
}
return false;
UNUSED(buffer);
}
@ -2572,18 +2576,25 @@ static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_
}
static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
/* .get_name = */ ggml_backend_metal_buffer_get_name,
/* .free_buffer = */ ggml_backend_metal_buffer_free_buffer,
/* .get_base = */ ggml_backend_metal_buffer_get_base,
/* .init_tensor = */ NULL,
/* .set_tensor = */ ggml_backend_metal_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_metal_buffer_get_tensor,
/* .cpy_tensor_from = */ ggml_backend_metal_buffer_cpy_tensor_from,
/* .cpy_tensor_to = */ ggml_backend_metal_buffer_cpy_tensor_to,
/* .cpy_tensor = */ ggml_backend_metal_buffer_cpy_tensor,
/* .clear = */ ggml_backend_metal_buffer_clear,
/* .reset = */ NULL,
};
// default buffer type
static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "Metal";
UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
@ -2656,6 +2667,7 @@ static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t bu
ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
/* .iface = */ {
/* .get_name = */ ggml_backend_metal_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_metal_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_metal_buffer_type_get_alignment,
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
@ -2679,6 +2691,14 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
ctx->n_buffers = 0;
const size_t size_page = sysconf(_SC_PAGESIZE);
// page-align the data ptr
{
const uintptr_t offs = (uintptr_t) data % size_page;
data = (void *) ((char *) data - offs);
size += offs;
}
size_t size_aligned = size;
if ((size_aligned % size_page) != 0) {
size_aligned += (size_page - (size_aligned % size_page));
@ -2779,14 +2799,13 @@ static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct
UNUSED(backend);
}
static struct ggml_backend_i metal_backend_i = {
static struct ggml_backend_i ggml_backend_metal_i = {
/* .get_name = */ ggml_backend_metal_name,
/* .free = */ ggml_backend_metal_free,
/* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_from_async = */ NULL,
/* .cpy_tensor_to_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
@ -2805,7 +2824,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
*metal_backend = (struct ggml_backend) {
/* .interface = */ metal_backend_i,
/* .interface = */ ggml_backend_metal_i,
/* .context = */ ctx,
};
@ -2813,7 +2832,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
}
bool ggml_backend_is_metal(ggml_backend_t backend) {
return backend->iface.get_name == ggml_backend_metal_name;
return backend && backend->iface.get_name == ggml_backend_metal_name;
}
void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {

335
ggml-opencl.cpp
View File

@ -1,5 +1,6 @@
#include "ggml.h"
#include "ggml-opencl.h"
#include "ggml-backend-impl.h"
#include <array>
#include <atomic>
@ -10,7 +11,7 @@
#include <sstream>
#include <vector>
#define CL_TARGET_OPENCL_VERSION 110
#define CL_TARGET_OPENCL_VERSION 120
#include <clblast.h>
#if defined(_MSC_VER)
@ -929,6 +930,12 @@ static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, co
}
void ggml_cl_init(void) {
static bool initialized = false;
if (initialized) {
return;
}
initialized = true;
cl_int err;
struct cl_device;
@ -1483,8 +1490,8 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
} else {
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
cl_mem d_Y = src1->backend == GGML_BACKEND_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = dst->backend == GGML_BACKEND_GPU ? (cl_mem) dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
size_t x_offset = 0;
@ -1501,7 +1508,9 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
if (src1->backend == GGML_BACKEND_CPU) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
}
CL_CHECK(clFinish(queue));
@ -1522,8 +1531,10 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
if (dst->backend == GGML_BACKEND_CPU) {
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
}
}
}
@ -1532,8 +1543,12 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
if (src0->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
if (src1->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_Y, y_size);
}
if (dst->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_D, d_size);
}
}
static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
@ -1598,6 +1613,8 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
}
// FIXME: convert on device
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
// convert src1 to fp16
// TODO: use multiple threads
@ -1643,11 +1660,13 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
}
// copy dst to host, then convert to float
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
ggml_fp16_to_fp32_row(tmp, d, d_ne);
if (dst->backend == GGML_BACKEND_CPU) {
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
ggml_fp16_to_fp32_row(tmp, d, d_ne);
} else {
// FIXME: convert dst to fp32 on device
}
}
}
}
@ -1801,7 +1820,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
}
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst) {
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = dst->ne[0];
@ -1895,3 +1914,291 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
tensor->extra = dst;
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
}
// ggml-backend
// buffer
struct ggml_backend_opencl_buffer_context {
~ggml_backend_opencl_buffer_context() {
if (buffer) {
clReleaseMemObject(buffer);
}
for (auto * sub_buffer : sub_buffers) {
clReleaseMemObject(sub_buffer);
}
}
cl_mem buffer;
std::vector<cl_mem> sub_buffers;
};
static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000;
static const char * ggml_backend_opencl_buffer_get_name(ggml_backend_buffer_t buffer) {
return "OpenCL";
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
delete ctx;
}
static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
return cl_ptr_base;
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
if (tensor->view_src != NULL && tensor->view_offs == 0) {
tensor->extra = tensor->view_src->extra;
} else {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
cl_buffer_region region = {(size_t)((char *)tensor->data - (char *)cl_ptr_base), ggml_nbytes(tensor)};
cl_int err;
cl_mem sub_buffer = clCreateSubBuffer(ctx->buffer, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
CL_CHECK(err);
ctx->sub_buffers.push_back(sub_buffer);
tensor->extra = sub_buffer;
}
tensor->backend = GGML_BACKEND_GPU;
}
static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
cl_mem tensor_buffer = (cl_mem) tensor->extra;
CL_CHECK(clEnqueueWriteBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
cl_mem tensor_buffer = (cl_mem) tensor->extra;
CL_CHECK(clEnqueueReadBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
CL_CHECK(clEnqueueFillBuffer(queue, ctx->buffer, &value, sizeof(value), 0, buffer->size, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
}
static void ggml_backend_opencl_buffer_reset(ggml_backend_buffer_t buffer) {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
for (auto * sub_buffer : ctx->sub_buffers) {
clReleaseMemObject(sub_buffer);
}
ctx->sub_buffers.clear();
}
static ggml_backend_buffer_i ggml_backend_opencl_buffer_interface = {
/* .get_name = */ ggml_backend_opencl_buffer_get_name,
/* .free_buffer = */ ggml_backend_opencl_buffer_free_buffer,
/* .get_base = */ ggml_backend_opencl_buffer_get_base,
/* .init_tensor = */ ggml_backend_opencl_buffer_init_tensor,
/* .set_tensor = */ ggml_backend_opencl_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_opencl_buffer_get_tensor,
/* .cpy_tensor = */ NULL,
/* .clear = */ ggml_backend_opencl_buffer_clear,
/* .reset = */ ggml_backend_opencl_buffer_reset,
};
// buffer type
static const char * ggml_backend_opencl_buffer_type_name(ggml_backend_buffer_type_t buffer_type) {
return "OpenCL";
GGML_UNUSED(buffer_type);
}
static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buffer_type, size_t size) {
ggml_cl_init();
cl_int err;
cl_mem mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err);
if (err != CL_SUCCESS) {
fprintf(stderr, "%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);
return nullptr;
}
ggml_backend_opencl_buffer_context * ctx = new ggml_backend_opencl_buffer_context{mem, {}};
return ggml_backend_buffer_init(buffer_type, ggml_backend_opencl_buffer_interface, ctx, size);
}
static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_type_t buffer_type) {
// FIXME: not thread safe, device may not be initialized yet
static cl_uint alignment = -1;
if (alignment == (cl_uint)-1) {
ggml_cl_init();
clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &alignment, NULL);
}
return alignment;
GGML_UNUSED(buffer_type);
}
static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buffer_type, ggml_backend_t backend) {
//return ggml_backend_is_opencl(backend); // opencl must be used through the cpu backend
return ggml_backend_is_cpu(backend);
GGML_UNUSED(buffer_type);
}
static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
/* .get_name = */ ggml_backend_opencl_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_opencl_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_opencl_buffer_type_get_alignment,
/* .get_alloc_size = */ NULL,
/* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
/* .is_host = */ NULL,
};
ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type() {
static ggml_backend_buffer_type buffer_type = {
/* .iface = */ ggml_backend_opencl_buffer_type_interface,
/* .context = */ nullptr,
};
return &buffer_type;
}
#if 0
// host buffer type
static const char * ggml_backend_opencl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
return "CL_Host";
GGML_UNUSED(buft);
}
static const char * ggml_backend_opencl_host_buffer_name(ggml_backend_buffer_t buffer) {
return "CL_Host";
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_cl_host_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_opencl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * ptr = ggml_cl_host_malloc(size);
if (ptr == nullptr) {
// fallback to cpu buffer
return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.get_name = ggml_backend_opencl_host_buffer_name;
buffer->iface.free_buffer = ggml_backend_opencl_host_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type() {
static struct ggml_backend_buffer_type ggml_backend_opencl_buffer_type_host = {
/* .iface = */ {
/* .get_name = */ ggml_backend_opencl_host_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_opencl_host_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
/* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
},
/* .context = */ nullptr,
};
return &ggml_backend_opencl_buffer_type_host;
}
// backend
static const char * ggml_backend_opencl_name(ggml_backend_t backend) {
return "OpenCL";
GGML_UNUSED(backend);
}
static void ggml_backend_opencl_free(ggml_backend_t backend) {
GGML_UNUSED(backend);
}
static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_opencl_buffer_type();
GGML_UNUSED(backend);
}
static bool ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
for (int i = 0; i < graph->n_nodes; ++i) {
ggml_tensor * node = graph->nodes[i];
switch (node->op) {
case GGML_OP_MUL_MAT:
ggml_cl_mul_mat(node->src[0], node->src[1], node, nullptr, 0);
break;
case GGML_OP_MUL:
ggml_cl_mul(node->src[0], node->src[1], node);
break;
default:
GGML_ASSERT(false);
}
}
return true;
GGML_UNUSED(backend);
}
static bool ggml_backend_opencl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
switch (op->op) {
case GGML_OP_MUL_MAT:
return ggml_cl_can_mul_mat(op->src[0], op->src[1], op);
case GGML_OP_MUL:
// return ggml_can_repeat_rows(op->src[1], op->src[0]);
return true;
default:
return false;
}
GGML_UNUSED(backend);
}
static ggml_backend_i opencl_backend_i = {
/* .get_name = */ ggml_backend_opencl_name,
/* .free = */ ggml_backend_opencl_free,
/* .get_default_buffer_type = */ ggml_backend_opencl_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_from_async = */ NULL,
/* .cpy_tensor_to_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_opencl_graph_compute,
/* .supports_op = */ ggml_backend_opencl_supports_op,
};
ggml_backend_t ggml_backend_opencl_init() {
ggml_backend_t backend = new ggml_backend {
/* .interface = */ opencl_backend_i,
/* .context = */ nullptr
};
return backend;
}
bool ggml_backend_is_opencl(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_opencl_name;
}
#endif

16
ggml-opencl.h
View File

@ -1,6 +1,7 @@
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
@ -9,17 +10,26 @@ extern "C" {
GGML_API void ggml_cl_init(void);
GGML_API void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
GGML_API void * ggml_cl_host_malloc(size_t size);
GGML_API void ggml_cl_host_free(void * ptr);
// GGML_API void * ggml_cl_host_malloc(size_t size);
// GGML_API void ggml_cl_host_free(void * ptr);
GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
// backend API
// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
#ifdef __cplusplus
}
#endif

30
ggml.c
View File

@ -2354,6 +2354,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
}
void ggml_free(struct ggml_context * ctx) {
if (ctx == NULL) {
return;
}
// make this function thread safe
ggml_critical_section_start();
@ -4362,6 +4366,23 @@ struct ggml_tensor * ggml_cpy(
return ggml_cpy_impl(ctx, a, b);
}
struct ggml_tensor * ggml_cast(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_type type) {
bool is_node = false;
struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
ggml_format_name(result, "%s (copy)", a->name);
result->op = GGML_OP_CPY;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
result->src[1] = result;
return result;
}
// ggml_cont
static struct ggml_tensor * ggml_cont_impl(
@ -14871,7 +14892,7 @@ size_t ggml_hash_find_or_insert(struct ggml_hash_set hash_set, struct ggml_tenso
return i;
}
static struct ggml_hash_set ggml_hash_set_new(size_t size) {
struct ggml_hash_set ggml_hash_set_new(size_t size) {
size = ggml_hash_size(size);
struct ggml_hash_set result;
result.size = size;
@ -16620,7 +16641,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
return GGML_EXIT_SUCCESS;
}
struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threads) {
if (n_threads <= 0) {
n_threads = GGML_DEFAULT_N_THREADS;
}
@ -16682,14 +16703,15 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
} break;
case GGML_OP_MUL_MAT_ID:
{
cur = 0;
const struct ggml_tensor * src0 = node->src[2];
const struct ggml_tensor * src1 = node->src[1];
const enum ggml_type vec_dot_type = type_traits[src0->type].vec_dot_type;
if (src1->type != vec_dot_type) {
cur = ggml_row_size(vec_dot_type, ggml_nelements(src1));
cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
}
const int n_as = ggml_get_op_params_i32(node, 1);
cur = GGML_PAD(cur, sizeof(int64_t)); // align
cur += GGML_PAD(cur, sizeof(int64_t)); // align
cur += n_as * sizeof(int64_t); // matrix_row_counts
cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows
} break;

9
ggml.h
View File

@ -1165,6 +1165,11 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_cast(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_type type);
// make contiguous
GGML_API struct ggml_tensor * ggml_cont(
struct ggml_context * ctx,
@ -1842,8 +1847,8 @@ extern "C" {
// ggml_graph_plan() has to be called before ggml_graph_compute()
// when plan.work_size > 0, caller must allocate memory for plan.work_data
GGML_API struct ggml_cplan ggml_graph_plan (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
GGML_API int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
GGML_API struct ggml_cplan ggml_graph_plan (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
GGML_API int ggml_graph_compute( struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
// same as ggml_graph_compute() but the work data is allocated as a part of the context
// note: the drawback of this API is that you must have ensured that the context has enough memory for the work data

2320
llama.cpp
View File

File diff suppressed because it is too large Load Diff

18
llama.h
View File

@ -118,6 +118,12 @@ extern "C" {
LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN,
};
enum llama_split_mode {
LLAMA_SPLIT_NONE = 0, // single GPU
LLAMA_SPLIT_LAYER = 1, // split layers and KV across GPUs
LLAMA_SPLIT_ROW = 2, // split rows across GPUs
};
typedef struct llama_token_data {
llama_token id; // token id
float logit; // log-odds of the token
@ -180,8 +186,16 @@ extern "C" {
struct llama_model_params {
int32_t n_gpu_layers; // number of layers to store in VRAM
int32_t main_gpu; // the GPU that is used for scratch and small tensors
const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
enum llama_split_mode split_mode; // how to split the model across multiple GPUs
// main_gpu interpretation depends on split_mode:
// LLAMA_SPLIT_NONE: the GPU that is used for the entire model
// LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
// LLAMA_SPLIT_LAYER: ignored
int32_t main_gpu;
// proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES
const float * tensor_split;
// Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
// If the provided progress_callback returns true, model loading continues.

26
tests/test-backend-ops.cpp
View File

@ -376,6 +376,11 @@ struct test_case {
// allocate
ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend1);
if (buf == NULL) {
printf("failed to allocate tensors [%s] ", ggml_backend_name(backend1));
ggml_free(ctx);
return false;
}
// build graph
ggml_build_forward_expand(gf, out);
@ -463,19 +468,23 @@ struct test_case {
GGML_UNUSED(index);
};
ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
const bool cmp_ok = ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
if (ud.ok) {
printf("\033[1;32mOK\033[0m\n");
} else {
printf("\033[1;31mFAIL\033[0m\n");
if (!cmp_ok) {
printf("compare failed ");
}
ggml_backend_buffer_free(buf);
ggml_free(ctx);
return ud.ok;
if (ud.ok && cmp_ok) {
printf("\033[1;32mOK\033[0m\n");
return true;
}
printf("\033[1;31mFAIL\033[0m\n");
return false;
}
bool eval_perf(ggml_backend_t backend, const char * op_name) {
@ -519,6 +528,11 @@ struct test_case {
// allocate
ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend);
if (buf == NULL) {
printf("failed to allocate tensors\n");
ggml_free(ctx);
return false;
}
// randomize tensors
initialize_tensors(ctx);