ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-09 13:30:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add support for Granite and GraniteMoE models (#102)

Browse Source 
* Add Granite and GranoteMoE models

* Granite: avoid NaNs on CUDA by scaling Q before K*Q multiplication

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-10-22 17:28:14 +02:00
committed by GitHub
parent 462c6cd7b1
commit b61cf7d0d7
1 changed files with 117 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

118
src/llama.cpp
View File

@@ -212,6 +212,8 @@ enum llm_arch {
LLM_ARCH_T5,
LLM_ARCH_T5ENCODER,
LLM_ARCH_JAIS,
LLM_ARCH_GRANITE = 46,
LLM_ARCH_GRANITE_MOE,
LLM_ARCH_UNKNOWN,
};
@@ -257,6 +259,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_T5, "t5" },
{ LLM_ARCH_T5ENCODER, "t5encoder" },
{ LLM_ARCH_JAIS, "jais" },
{ LLM_ARCH_GRANITE, "granite" },
{ LLM_ARCH_GRANITE_MOE, "granitemoe" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
@@ -293,6 +297,12 @@ enum llm_kv {
LLM_KV_DECODER_START_TOKEN_ID,
LLM_KV_ATTN_LOGIT_SOFTCAPPING,
LLM_KV_FINAL_LOGIT_SOFTCAPPING,
LLM_KV_SWIN_NORM,
LLM_KV_RESCALE_EVERY_N_LAYERS,
LLM_KV_TIME_MIX_EXTRA_DIM,
LLM_KV_TIME_DECAY_EXTRA_DIM,
LLM_KV_RESIDUAL_SCALE,
LLM_KV_EMBEDDING_SCALE,
LLM_KV_ATTENTION_HEAD_COUNT,
LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -307,6 +317,7 @@ enum llm_kv {
LLM_KV_ATTENTION_KV_LORA_RANK,
LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
LLM_KV_ATTENTION_SLIDING_WINDOW,
LLM_KV_ATTENTION_SCALE,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_FREQ_BASE,
@@ -391,6 +402,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" },
{ LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" },
{ LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" },
{ LLM_KV_RESIDUAL_SCALE, "%s.residual_scale" },
{ LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
@@ -405,6 +418,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" },
{ LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
{ LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
@@ -1298,6 +1312,42 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
},
},
{
LLM_ARCH_GRANITE,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_GRANITE_MOE,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
},
},
{
LLM_ARCH_UNKNOWN,
{
@@ -2203,6 +2253,11 @@ struct llama_hparams {
float f_max_alibi_bias = 0.0f;
float f_logit_scale = 0.0f;
// Additional scale factors (Granite/Granite MoE)
float f_residual_scale = 0.0f;
float f_embedding_scale = 0.0f;
float f_attention_scale = 0.0f;
bool causal_attn = true;
bool use_alibi = false;
bool attn_soft_cap = false;
@@ -2259,6 +2314,9 @@ struct llama_hparams {
if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
if (!is_float_close(this->expert_weights_scale, other.expert_weights_scale, EPSILON)) return true;
if (!is_float_close(this->rope_yarn_log_mul, other.rope_yarn_log_mul, EPSILON)) return true;
if (!is_float_close(this->f_residual_scale, other.f_residual_scale, EPSILON)) return true;
if (!is_float_close(this->f_embedding_scale, other.f_embedding_scale, EPSILON)) return true;
if (!is_float_close(this->f_attention_scale, other.f_attention_scale, EPSILON)) return true;
return false;
}
@@ -5283,6 +5341,22 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale);
ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale);
ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale);
switch (hparams.n_layer) {
case 32: model.type = e_model::MODEL_3B; break;
case 40: model.type = e_model::MODEL_3B; break;
// Add additional layer/vocab/etc checks here for other model sizes
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
default: (void)0;
}
@@ -5970,6 +6044,13 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
}
if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale);
LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
}
}
// Returns false if cancelled by progress_callback
@@ -6138,6 +6219,8 @@ static bool llm_load_tensors(
case LLM_ARCH_LLAMA:
case LLM_ARCH_REFACT:
case LLM_ARCH_MINICPM:
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@@ -7927,6 +8010,11 @@ static struct ggml_tensor * llm_build_inp_embd(
ggml_set_input(lctx.inp_embd);
}
// For Granite architecture
if (hparams.f_embedding_scale != 0.0f) {
inpL = ggml_scale(ctx, inpL, hparams.f_embedding_scale);
}
cb(inpL, "inp_embd", -1);
return inpL;
@@ -8358,12 +8446,15 @@ static struct ggml_tensor * llm_build_kqv(
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
}
//ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);
} else {
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
cb(kq, "kq", il);
//ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
// for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
// ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
@@ -8917,6 +9008,8 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
//const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : 1.f;
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
@@ -8933,6 +9026,9 @@ struct llm_build_context {
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
if (hparams.f_attention_scale != 0) {
Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
}
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
@@ -8969,7 +9065,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, lctx, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
}
if (il == n_layer - 1) {
@@ -8980,6 +9076,11 @@ struct llm_build_context {
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
// For Granite architecture
if (hparams.f_residual_scale) {
cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
}
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
@@ -9016,6 +9117,11 @@ struct llm_build_context {
cb(cur, "ffn_moe_out", il);
}
// For Granite architecture
if (hparams.f_residual_scale) {
cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
}
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", il);
@@ -9035,6 +9141,12 @@ struct llm_build_context {
// lm_head
cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
// For Granite architecture
if (hparams.f_logit_scale) {
cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
}
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
@@ -14032,6 +14144,8 @@ static struct ggml_cgraph * llama_build_graph(
switch (model.arch) {
case LLM_ARCH_LLAMA:
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
{
result = llm.build_llama();
} break;
@@ -17470,6 +17584,8 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_ARCTIC:
case LLM_ARCH_DEEPSEEK2:
case LLM_ARCH_CHATGLM:
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2