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

code : normalize enum names (#5697)

Browse source 
* coda : normalize enum names

ggml-ci

* code : cont

* code : cont
This commit is contained in:
Georgi Gerganov 2024-02-25 12:09:09 +02:00 • committed by GitHub
parent 69917dfa55
commit ab336a9d5e
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
20 changed files with 502 additions and 502 deletions
Download patch file Download diff file Expand all files Collapse all files

18
common/common.cpp
View file

@ -295,9 +295,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
break;
}
std::string value(argv[i]);
/**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
/**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
else { invalid_param = true; break; }
} else if (arg == "--rope-scale") {
if (++i >= argc) {
@ -630,11 +630,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
}
std::string arg_next = argv[i];
if (arg_next == "none") {
params.split_mode = LLAMA_SPLIT_NONE;
params.split_mode = LLAMA_SPLIT_MODE_NONE;
} else if (arg_next == "layer") {
params.split_mode = LLAMA_SPLIT_LAYER;
params.split_mode = LLAMA_SPLIT_MODE_LAYER;
} else if (arg_next == "row") {
params.split_mode = LLAMA_SPLIT_ROW;
params.split_mode = LLAMA_SPLIT_MODE_ROW;
} else {
invalid_param = true;
break;
@ -837,15 +837,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
sep++;
if (strncmp(sep, "int:", 4) == 0) {
sep += 4;
kvo.tag = LLAMA_KV_OVERRIDE_INT;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
kvo.int_value = std::atol(sep);
} else if (strncmp(sep, "float:", 6) == 0) {
sep += 6;
kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
kvo.float_value = std::atof(sep);
} else if (strncmp(sep, "bool:", 5) == 0) {
sep += 5;
kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
if (std::strcmp(sep, "true") == 0) {
kvo.bool_value = true;
} else if (std::strcmp(sep, "false") == 0) {

4
common/common.h
View file

@ -61,7 +61,7 @@ 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
llama_split_mode split_mode = LLAMA_SPLIT_MODE_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[128] = {0}; // how split tensors should be distributed across GPUs
int32_t n_beams = 0; // if non-zero then use beam search of given width.
@ -75,7 +75,7 @@ struct gpt_params {
float yarn_beta_fast = 32.0f; // YaRN low correction dim
float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length
int32_t rope_scaling_type = LLAMA_ROPE_SCALING_UNSPECIFIED;
int32_t rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
// // sampling parameters

10
common/train.cpp
View file

@ -31,7 +31,7 @@ struct train_state * init_train_state() {
state->opt = new struct ggml_opt_context;
state->opt->ctx = NULL;
state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;
state->opt->loss_after = 0.0f;
@ -556,7 +556,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
std::string opt_type;
GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
opt->params.type = GGML_OPT_ADAM;
opt->params.type = GGML_OPT_TYPE_ADAM;
GGUF_GET_KEY(fctx, opt->adam.fx_best, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
GGUF_GET_KEY(fctx, opt->adam.fx_prev, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
@ -568,7 +568,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
copy_tensor_by_name(opt->adam.v, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
} else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
opt->params.type = GGML_OPT_LBFGS;
opt->params.type = GGML_OPT_TYPE_LBFGS;
GGUF_GET_KEY(fctx, opt->params.lbfgs.m, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
GGUF_GET_KEY(fctx, opt->lbfgs.fx_best, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
@ -603,7 +603,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
switch (opt->params.type) {
case GGML_OPT_ADAM:
case GGML_OPT_TYPE_ADAM:
{
gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, opt->adam.fx_best);
@ -622,7 +622,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
gguf_add_tensor(fctx, opt->adam.pf);
}
} break;
case GGML_OPT_LBFGS:
case GGML_OPT_TYPE_LBFGS:
{
gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);