Skip to content

Commit

Permalink
Merge branch 'master' into gg/metal-feature-set
Browse files Browse the repository at this point in the history 
ggml-ci
  • Loading branch information
@ggerganov
ggerganov committed Jan 13, 2024
2 parents 5d33d3c + 7dc7876 commit f81e467
Show file tree
Hide file tree
Showing 30 changed files with 4,029 additions and 2,324 deletions.
65 changes: 39 additions & 26 deletions common/common.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -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
Expand All @@ -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
Expand All @@ -565,40 +563,53 @@ 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]);
} else {
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;
Expand Down Expand Up @@ -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);
Expand Down Expand Up @@ -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;
Expand Down
1 change: 1 addition & 0 deletions common/common.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -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.
Expand Down
39 changes: 27 additions & 12 deletions convert-hf-to-gguf.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,15 @@
import gguf


# check for any of the given keys in the dictionary and return the value of the first key found
def get_key_opts(d, keys):
for k in keys:
if k in d:
return d[k]
print(f"Could not find any of {keys}")
sys.exit()


###### MODEL DEFINITIONS ######

class SentencePieceTokenTypes(IntEnum):
Expand Down Expand Up @@ -257,10 +266,11 @@ def _set_vocab_gpt2(self):
toktypes.append(gguf.TokenType.USER_DEFINED)
elif reverse_vocab[i] in added_vocab:
tokens.append(reverse_vocab[i])
if tokenizer.added_tokens_decoder[i].special:
toktypes.append(gguf.TokenType.CONTROL)
else:
toktypes.append(gguf.TokenType.USER_DEFINED)
if hasattr(tokenizer, "added_tokens_decoder"):
if tokenizer.added_tokens_decoder[i].special:
toktypes.append(gguf.TokenType.CONTROL)
else:
toktypes.append(gguf.TokenType.USER_DEFINED)
else:
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.NORMAL)
Expand Down Expand Up @@ -1068,17 +1078,22 @@ def write_tensors(self):

class Phi2Model(Model):
def set_gguf_parameters(self):
block_count = self.hparams["n_layer"]
block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])

rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])

self.gguf_writer.add_name("Phi2")
self.gguf_writer.add_context_length(self.hparams["n_positions"])
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))

self.gguf_writer.add_embedding_length(n_embd)
self.gguf_writer.add_feed_forward_length(4 * n_embd)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(self.hparams["n_head"])
self.gguf_writer.add_head_count_kv(self.hparams["n_head"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_rope_dimension_count(self.hparams["rotary_dim"])
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head)
self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_add_bos_token(False)

Expand Down
3 changes: 3 additions & 0 deletions examples/batched-bench/batched-bench.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -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);

Expand Down
3 changes: 1 addition & 2 deletions examples/export-lora/export-lora.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -245,9 +245,8 @@ static struct lora_data * load_lora(struct lora_info * info) {
params_ggml.no_alloc = true;
result->ctx = ggml_init(params_ggml);

uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
uint32_t magic = file.read_u32();
if (magic != LLAMA_FILE_MAGIC_LORA) {
if (magic != LLAMA_FILE_MAGIC_GGLA) {
die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
}
uint32_t version = file.read_u32();
Expand Down
Loading

0 comments on commit f81e467

Please sign in to comment.