ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-14 12:58:10 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-05-19 13:39:12 +03:00
parent 05501c218d
commit 6b6cf19c65
8 changed files with 256 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files

240
examples/talk-llama/llama-model.cpp
View File

@ -1389,6 +1389,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
// Add additional layer/vocab/etc checks here for other model sizes
default: type = LLM_TYPE_UNKNOWN;
}
// For Granite MoE Shared
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
} break;
case LLM_ARCH_CHAMELEON:
{
@ -1772,6 +1775,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
// For Granite MoE Shared
if (hparams.n_ff_shexp > 0) {
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
}
}
}
} break;
@ -4385,10 +4395,13 @@ void llama_model::print_info() const {
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
}
if (arch == LLM_ARCH_MINICPM || arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE) {
if (arch == LLM_ARCH_MINICPM ||
arch == LLM_ARCH_GRANITE ||
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);
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
}
if (arch == LLM_ARCH_BAILINGMOE) {
@ -4598,11 +4611,6 @@ struct llm_build_llama : public llm_graph_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);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
@ -4674,11 +4682,6 @@ struct llm_build_llama : public llm_graph_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);
@ -4701,11 +4704,6 @@ struct llm_build_llama : public llm_graph_context {
// lm_head
cur = build_lora_mm(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);
res->t_logits = cur;
@ -4816,11 +4814,6 @@ struct llm_build_deci : public llm_graph_context {
continue;
}
// For Granite architecture
if (hparams.f_residual_scale) {
cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
}
// modified to support attention-free layer of Llama-3_1-Nemotron-51B
ggml_tensor * ffn_inp = cur;
if (n_head > 0) {
@ -4844,11 +4837,6 @@ struct llm_build_deci : public llm_graph_context {
cb(cur, "ffn_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);
@ -4871,11 +4859,6 @@ struct llm_build_deci : public llm_graph_context {
// lm_head
cur = build_lora_mm(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);
res->t_logits = cur;
@ -12214,6 +12197,194 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
}
};
struct llm_build_granite : public llm_graph_context {
llm_build_granite(
const llama_model & model,
const llm_graph_params & params,
ggml_cgraph * gf,
const bool use_rope = true)
: llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
// inp_pos - built only if rope enabled
ggml_tensor * inp_pos = nullptr;
if (use_rope) {
inp_pos = build_inp_pos();
}
auto * inp_attn = build_attn_inp_kv_unified();
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
// norm
cur = build_norm(inpL,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and (optionally) RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
if (use_rope) {
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
}
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}
if (il == n_layer - 1) {
// skip computing output for unused tokens
ggml_tensor * inp_out_ids = build_inp_out_ids();
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
// For Granite architectures - scale residual
cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// feed-forward network (non-MoE)
if (model.layers[il].ffn_gate_inp == nullptr) {
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
cur = build_ffn(cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
} else {
// MoE branch
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
ggml_tensor * moe_out = build_moe_ffn(cur,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
cb(moe_out, "ffn_moe_out", il);
// For Granite MoE Shared
if (hparams.n_ff_shexp > 0) {
ggml_tensor * ffn_shexp = build_ffn(cur,
model.layers[il].ffn_up_shexp, NULL, NULL,
model.layers[il].ffn_gate_shexp, NULL, NULL,
model.layers[il].ffn_down_shexp, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(ffn_shexp, "ffn_shexp", il);
cur = ggml_add(ctx0, moe_out, ffn_shexp);
cb(cur, "ffn_out", il);
} else {
cur = moe_out;
}
}
// For Granite architectures - scale residual
cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", il);
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = build_norm(cur,
model.output_norm, NULL,
LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// lm_head
cur = build_lora_mm(model.output, cur);
// For Granite architectures - scale logits
cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
};
// ref: https://github.com/facebookresearch/chameleon
// based on the original build_llama() function, changes:
// * qk-norm
@ -12921,8 +13092,6 @@ llm_graph_result_ptr llama_model::build_graph(
case LLM_ARCH_LLAMA:
case LLM_ARCH_LLAMA4:
case LLM_ARCH_MINICPM:
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
{
llm = std::make_unique<llm_build_llama>(*this, params, gf);
} break;
@ -13153,6 +13322,11 @@ llm_graph_result_ptr llama_model::build_graph(
{
llm = std::make_unique<llm_build_arwkv7>(*this, params, gf);
} break;
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
{
llm = std::make_unique<llm_build_granite>(*this, params, gf);
} break;
case LLM_ARCH_CHAMELEON:
{
llm = std::make_unique<llm_build_chameleon>(*this, params, gf);