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

llama : fix llm_type enum names

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2025-01-07 16:04:56 +02:00
parent fffa6b15c4
commit a48412f92b
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
5 changed files with 381 additions and 382 deletions
Download patch file Download diff file Expand all files Collapse all files

2
src/llama-mmap.cpp
View file

@ -35,7 +35,7 @@
// TODO: consider moving to llama-impl.h if needed in more places // TODO: consider moving to llama-impl.h if needed in more places
#if defined(_WIN32) #if defined(_WIN32)
std::string llama_format_win_err(DWORD err) { static std::string llama_format_win_err(DWORD err) {
LPSTR buf; LPSTR buf;
size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL); NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);

616
src/llama-model.cpp
View file

File diff suppressed because it is too large Load diff

131
src/llama-model.h
View file

@ -13,72 +13,71 @@
struct llama_model_loader; struct llama_model_loader;
// available models // available models
// TODO: this enum does not follow the enum naming convention
enum llm_type { enum llm_type {
MODEL_UNKNOWN, LLM_TYPE_UNKNOWN,
MODEL_14M, LLM_TYPE_14M,
MODEL_17M, LLM_TYPE_17M,
MODEL_22M, LLM_TYPE_22M,
MODEL_33M, LLM_TYPE_33M,
MODEL_60M, LLM_TYPE_60M,
MODEL_70M, LLM_TYPE_70M,
MODEL_80M, LLM_TYPE_80M,
MODEL_109M, LLM_TYPE_109M,
MODEL_137M, LLM_TYPE_137M,
MODEL_160M, LLM_TYPE_160M,
MODEL_220M, LLM_TYPE_220M,
MODEL_250M, LLM_TYPE_250M,
MODEL_270M, LLM_TYPE_270M,
MODEL_335M, LLM_TYPE_335M,
MODEL_410M, LLM_TYPE_410M,
MODEL_450M, LLM_TYPE_450M,
MODEL_770M, LLM_TYPE_770M,
MODEL_780M, LLM_TYPE_780M,
MODEL_0_5B, LLM_TYPE_0_5B,
MODEL_1B, LLM_TYPE_1B,
MODEL_1_3B, LLM_TYPE_1_3B,
MODEL_1_4B, LLM_TYPE_1_4B,
MODEL_1_5B, LLM_TYPE_1_5B,
MODEL_1_6B, LLM_TYPE_1_6B,
MODEL_2B, LLM_TYPE_2B,
MODEL_2_8B, LLM_TYPE_2_8B,
MODEL_3B, LLM_TYPE_3B,
MODEL_4B, LLM_TYPE_4B,
MODEL_6B, LLM_TYPE_6B,
MODEL_6_9B, LLM_TYPE_6_9B,
MODEL_7B, LLM_TYPE_7B,
MODEL_8B, LLM_TYPE_8B,
MODEL_9B, LLM_TYPE_9B,
MODEL_11B, LLM_TYPE_11B,
MODEL_12B, LLM_TYPE_12B,
MODEL_13B, LLM_TYPE_13B,
MODEL_14B, LLM_TYPE_14B,
MODEL_15B, LLM_TYPE_15B,
MODEL_16B, LLM_TYPE_16B,
MODEL_20B, LLM_TYPE_20B,
MODEL_30B, LLM_TYPE_30B,
MODEL_32B, LLM_TYPE_32B,
MODEL_34B, LLM_TYPE_34B,
MODEL_35B, LLM_TYPE_35B,
MODEL_40B, LLM_TYPE_40B,
MODEL_65B, LLM_TYPE_65B,
MODEL_70B, LLM_TYPE_70B,
MODEL_236B, LLM_TYPE_236B,
MODEL_314B, LLM_TYPE_314B,
MODEL_671B, LLM_TYPE_671B,
MODEL_SMALL, LLM_TYPE_SMALL,
MODEL_MEDIUM, LLM_TYPE_MEDIUM,
MODEL_LARGE, LLM_TYPE_LARGE,
MODEL_XL, LLM_TYPE_XL,
MODEL_A1_7B, LLM_TYPE_A1_7B,
MODEL_A2_7B, LLM_TYPE_A2_7B,
MODEL_8x7B, LLM_TYPE_8x7B,
MODEL_8x22B, LLM_TYPE_8x22B,
MODEL_16x12B, LLM_TYPE_16x12B,
MODEL_16x3_8B, LLM_TYPE_16x3_8B,
MODEL_10B_128x3_66B, LLM_TYPE_10B_128x3_66B,
MODEL_57B_A14B, LLM_TYPE_57B_A14B,
MODEL_27B, LLM_TYPE_27B,
}; };
struct llama_layer_posnet { struct llama_layer_posnet {
@ -284,7 +283,7 @@ struct llama_layer {
}; };
struct llama_model { struct llama_model {
llm_type type = MODEL_UNKNOWN; llm_type type = LLM_TYPE_UNKNOWN;
llm_arch arch = LLM_ARCH_UNKNOWN; llm_arch arch = LLM_ARCH_UNKNOWN;
std::string name = "n/a"; std::string name = "n/a";

2
src/llama-quant.cpp
View file

@ -235,7 +235,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K; use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
if (qs.model.type == MODEL_70B) { if (qs.model.type == LLM_TYPE_70B) {
// In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
// 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
// nearly negligible increase in model size by quantizing this tensor with more bits: // nearly negligible increase in model size by quantizing this tensor with more bits:

12
src/llama.cpp
View file

@ -1809,7 +1809,7 @@ struct llm_build_context {
inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
// inp_pos - contains the positions // inp_pos - contains the positions
struct ggml_tensor * inp_pos = model.type == MODEL_7B ? build_inp_pos() : nullptr; struct ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
// KQ_mask (mask for 1 head, it will be broadcasted to all heads) // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@ -1834,7 +1834,7 @@ struct llm_build_context {
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
switch (model.type) { switch (model.type) {
case MODEL_7B: case LLM_TYPE_7B:
Qcur = ggml_rope_ext( Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@ -1846,7 +1846,7 @@ struct llm_build_context {
ext_factor, attn_factor, beta_fast, beta_slow ext_factor, attn_factor, beta_fast, beta_slow
); );
break; break;
case MODEL_13B: case LLM_TYPE_13B:
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens); Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens);
break; break;
@ -4889,9 +4889,9 @@ struct llm_build_context {
// ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
switch (model.type) { switch (model.type) {
case llm_type::MODEL_2B: case LLM_TYPE_2B:
case llm_type::MODEL_9B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break; case LLM_TYPE_9B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break;
case llm_type::MODEL_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break; case LLM_TYPE_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
default: GGML_ABORT("fatal error"); default: GGML_ABORT("fatal error");
}; };
cb(Qcur, "Qcur_scaled", il); cb(Qcur, "Qcur_scaled", il);