Skip to content

Commit 65afc73

Browse files
gabe-l-hartgabe-l-hart
committed
feat(llama.cpp): Add config parsing for Granite multiplier params
Branch: GraniteLM Signed-off-by: Gabe Goodhart <[email protected]>
1 parent 0e31ab8 commit 65afc73

File tree

1 file changed

+38
-8
lines changed

1 file changed

+38
-8
lines changed

‎src/llama.cpp

Lines changed: 38 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -213,6 +213,7 @@ enum llm_arch {
213213
LLM_ARCH_NEMOTRON,
214214
LLM_ARCH_EXAONE,
215215
LLM_ARCH_RWKV6,
216+
LLM_ARCH_GRANITE,
216217
LLM_ARCH_UNKNOWN,
217218
};
218219

@@ -261,6 +262,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
261262
{ LLM_ARCH_NEMOTRON, "nemotron" },
262263
{ LLM_ARCH_EXAONE, "exaone" },
263264
{ LLM_ARCH_RWKV6, "rwkv6" },
265+
{ LLM_ARCH_GRANITE, "granite" },
264266
{ LLM_ARCH_UNKNOWN, "(unknown)" },
265267
};
266268

@@ -300,6 +302,8 @@ enum llm_kv {
300302
LLM_KV_RESCALE_EVERY_N_LAYERS,
301303
LLM_KV_TIME_MIX_EXTRA_DIM,
302304
LLM_KV_TIME_DECAY_EXTRA_DIM,
305+
LLM_KV_RESIDUAL_MULTIPLIER,
306+
LLM_KV_EMBEDDING_MULTIPLIER,
303307

304308
LLM_KV_ATTENTION_HEAD_COUNT,
305309
LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -314,6 +318,7 @@ enum llm_kv {
314318
LLM_KV_ATTENTION_KV_LORA_RANK,
315319
LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
316320
LLM_KV_ATTENTION_SLIDING_WINDOW,
321+
LLM_KV_ATTENTION_MULTIPLIER,
317322

318323
LLM_KV_ROPE_DIMENSION_COUNT,
319324
LLM_KV_ROPE_FREQ_BASE,
@@ -404,6 +409,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
404409
{ LLM_KV_RESCALE_EVERY_N_LAYERS, "%s.rescale_every_n_layers" },
405410
{ LLM_KV_TIME_MIX_EXTRA_DIM, "%s.time_mix_extra_dim" },
406411
{ LLM_KV_TIME_DECAY_EXTRA_DIM, "%s.time_decay_extra_dim" },
412+
{ LLM_KV_RESIDUAL_MULTIPLIER, "%s.residual_multiplier" },
413+
{ LLM_KV_EMBEDDING_MULTIPLIER, "%s.embedding_multiplier" },
407414

408415
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
409416
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
@@ -418,6 +425,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
418425
{ LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" },
419426
{ LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
420427
{ LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
428+
{ LLM_KV_ATTENTION_MULTIPLIER, "%s.attention.multiplier" },
421429

422430
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
423431
{ LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
@@ -2321,6 +2329,11 @@ struct llama_hparams {
23212329
float f_max_alibi_bias = 0.0f;
23222330
float f_logit_scale = 0.0f;
23232331

2332+
// For Granite architecture
2333+
float f_residual_multiplier = 0.0f;
2334+
float f_embedding_multiplier = 0.0f;
2335+
float f_attention_multiplier = 0.0f;
2336+
23242337
bool causal_attn = true;
23252338
bool use_alibi = false;
23262339
bool attn_soft_cap = false;
@@ -2376,13 +2389,16 @@ struct llama_hparams {
23762389

23772390
const float EPSILON = 1e-9f;
23782391

2379-
if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true;
2380-
if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true;
2381-
if (!is_float_close(this->rope_attn_factor, other.rope_attn_factor, EPSILON)) return true;
2382-
if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true;
2383-
if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
2384-
if (!is_float_close(this->expert_weights_scale, other.expert_weights_scale, EPSILON)) return true;
2385-
if (!is_float_close(this->rope_yarn_log_mul, other.rope_yarn_log_mul, EPSILON)) return true;
2392+
if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true;
2393+
if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true;
2394+
if (!is_float_close(this->rope_attn_factor, other.rope_attn_factor, EPSILON)) return true;
2395+
if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true;
2396+
if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
2397+
if (!is_float_close(this->expert_weights_scale, other.expert_weights_scale, EPSILON)) return true;
2398+
if (!is_float_close(this->rope_yarn_log_mul, other.rope_yarn_log_mul, EPSILON)) return true;
2399+
if (!is_float_close(this->f_residual_multiplier, other.f_residual_multiplier, EPSILON)) return true;
2400+
if (!is_float_close(this->f_embedding_multiplier, other.f_embedding_multiplier, EPSILON)) return true;
2401+
if (!is_float_close(this->f_attention_multiplier, other.f_attention_multiplier, EPSILON)) return true;
23862402

23872403
return false;
23882404
}
@@ -5351,6 +5367,7 @@ static void llm_load_hparams(
53515367
// arch-specific KVs
53525368
switch (model.arch) {
53535369
case LLM_ARCH_LLAMA:
5370+
case LLM_ARCH_GRANITE:
53545371
{
53555372
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
53565373

@@ -5367,13 +5384,20 @@ static void llm_load_hparams(
53675384
// granite uses a vocab with len 49152
53685385
case 32: model.type = hparams.n_vocab == 49152 ? e_model::MODEL_3B : (hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B); break;
53695386
case 36: model.type = e_model::MODEL_8B; break; // granite
5370-
case 40: model.type = e_model::MODEL_13B; break;
5387+
case 40: model.type = hparams.n_vocab == 49152 ? e_model::MODEL_3B : e_model::MODEL_13B; break;
53715388
case 48: model.type = e_model::MODEL_34B; break;
53725389
case 60: model.type = e_model::MODEL_30B; break;
53735390
case 80: model.type = hparams.n_head() == hparams.n_head_kv() ? e_model::MODEL_65B : e_model::MODEL_70B; break;
53745391
default: model.type = e_model::MODEL_UNKNOWN;
53755392
}
53765393
}
5394+
// Extra multipliers for Granite architecture
5395+
if (model.arch == LLM_ARCH_GRANITE) {
5396+
ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
5397+
ml.get_key(LLM_KV_RESIDUAL_MULTIPLIER, hparams.f_residual_multiplier);
5398+
ml.get_key(LLM_KV_EMBEDDING_MULTIPLIER, hparams.f_embedding_multiplier);
5399+
ml.get_key(LLM_KV_ATTENTION_MULTIPLIER, hparams.f_attention_multiplier);
5400+
}
53775401
} break;
53785402
case LLM_ARCH_MINICPM:
53795403
{
@@ -6632,6 +6656,12 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
66326656
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
66336657
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
66346658
}
6659+
6660+
if (model.arch == LLM_ARCH_GRANITE) {
6661+
LLAMA_LOG_INFO("%s: f_embedding_multiplier = %f\n", __func__, hparams.f_embedding_multiplier);
6662+
LLAMA_LOG_INFO("%s: f_residual_multiplier = %f\n", __func__, hparams.f_residual_multiplier);
6663+
LLAMA_LOG_INFO("%s: f_attention_multiplier = %f\n", __func__, hparams.f_attention_multiplier);
6664+
}
66356665
}
66366666

66376667
// Returns false if cancelled by progress_callback

0 commit comments

Comments
 (0)