Skip to content

Commit cff16cc

Browse files
ngxsonngxson
committed
wip
1 parent 51886a4 commit cff16cc

File tree

2 files changed

+46
-48
lines changed

2 files changed

+46
-48
lines changed

‎src/llama-graph.cpp‎

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -705,6 +705,13 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
705705
ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
706706
cb(weights, "ffn_moe_weights", il);
707707

708+
if (arch == LLM_ARCH_HUNYUAN_MOE) {
709+
weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_expert_used, n_tokens]
710+
weights = ggml_div(ctx0, weights, ggml_sum_rows(ctx0, weights)); // [1, n_tokens]
711+
weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens); // [1, n_expert_used, n_tokens]
712+
cb(weights, "ffn_moe_weights_scaled", il);
713+
}
714+
708715
if (norm_w) {
709716
weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
710717

‎src/llama-model.cpp‎

Lines changed: 39 additions & 48 deletions
Original file line numberDiff line numberDiff line change
@@ -1507,10 +1507,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
15071507
} break;
15081508
case LLM_ARCH_HUNYUAN_MOE:
15091509
{
1510-
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
1511-
1512-
hparams.n_ff_exp = hparams.n_ff(0);
1513-
hparams.n_ff_shexp = hparams.n_ff_exp;
1510+
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
1511+
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
1512+
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp);
15141513

15151514
switch (hparams.n_layer) {
15161515
case 32: type = LLM_TYPE_A13B; break;
@@ -14377,29 +14376,25 @@ struct llm_build_hunyuan_moe : public llm_graph_context {
1437714376
ext_factor, attn_factor, beta_fast, beta_slow
1437814377
);
1437914378

14379+
cb(Qcur, "Qcur", il);
14380+
cb(Kcur, "Kcur", il);
14381+
cb(Vcur, "Vcur", il);
14382+
1438014383
Kcur = ggml_rope_ext(
1438114384
ctx0, Kcur, inp_pos, rope_factors,
1438214385
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
1438314386
ext_factor, attn_factor, beta_fast, beta_slow
1438414387
);
1438514388

14386-
if (model.layers[il].attn_k_norm) {
14387-
Kcur = build_norm(Kcur,
14388-
model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b,
14389-
LLM_NORM_RMS, il);
14390-
cb(Kcur, "Kcur_norm", il);
14391-
}
14392-
14393-
if (model.layers[il].attn_q_norm) {
14394-
Qcur = build_norm(Qcur,
14395-
model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b,
14396-
LLM_NORM_RMS, il);
14397-
cb(Qcur, "Qcur_norm", il);
14398-
}
14389+
Kcur = build_norm(Kcur,
14390+
model.layers[il].attn_k_norm, nullptr,
14391+
LLM_NORM_RMS, il);
14392+
cb(Kcur, "Kcur_norm", il);
1439914393

14400-
cb(Qcur, "Qcur", il);
14401-
cb(Kcur, "Kcur", il);
14402-
cb(Vcur, "Vcur", il);
14394+
Qcur = build_norm(Qcur,
14395+
model.layers[il].attn_q_norm, nullptr,
14396+
LLM_NORM_RMS, il);
14397+
cb(Qcur, "Qcur_norm", il);
1440314398

1440414399
cur = build_attn(inp_attn, gf,
1440514400
model.layers[il].wo, model.layers[il].bo,
@@ -14415,42 +14410,38 @@ struct llm_build_hunyuan_moe : public llm_graph_context {
1441514410
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
1441614411
cb(ffn_inp, "ffn_inp", il);
1441714412

14418-
ffn_inp = build_norm(ffn_inp,
14413+
cur = build_norm(ffn_inp,
1441914414
model.layers[il].ffn_norm, NULL,
1442014415
LLM_NORM_RMS, il);
1442114416
cb(cur, "ffn_norm", il);
1442214417

1442314418
// feed-forward network (non-MoE)
14424-
ggml_tensor * cur_mlp = nullptr;
14425-
{
14426-
cur_mlp = build_ffn(ffn_inp,
14427-
model.layers[il].ffn_up_shexp, NULL, NULL,
14428-
model.layers[il].ffn_gate_shexp, NULL, NULL,
14429-
model.layers[il].ffn_down_shexp, NULL, NULL,
14430-
NULL,
14431-
LLM_FFN_SILU, LLM_FFN_PAR, il);
14432-
cb(cur_mlp, "ffn_out", il);
14433-
}
14419+
ggml_tensor * cur_mlp = build_ffn(cur,
14420+
model.layers[il].ffn_up_shexp, NULL, NULL,
14421+
model.layers[il].ffn_gate_shexp, NULL, NULL,
14422+
model.layers[il].ffn_down_shexp, NULL, NULL,
14423+
NULL,
14424+
LLM_FFN_SILU, LLM_FFN_PAR, il);
14425+
cb(cur_mlp, "ffn_mlp", il);
1443414426

1443514427
// MoE branch
14436-
ggml_tensor * cur_moe = nullptr;
14437-
{
14438-
cur_moe = build_moe_ffn(ffn_inp,
14439-
model.layers[il].ffn_gate_inp,
14440-
model.layers[il].ffn_up_exps,
14441-
model.layers[il].ffn_gate_exps,
14442-
model.layers[il].ffn_down_exps,
14443-
nullptr,
14444-
n_expert, n_expert_used,
14445-
LLM_FFN_SILU, true,
14446-
false, 0.0,
14447-
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
14448-
il);
14449-
cb(cur_moe, "ffn_moe_out", il);
14450-
}
14428+
ggml_tensor * cur_moe = build_moe_ffn(cur,
14429+
model.layers[il].ffn_gate_inp,
14430+
model.layers[il].ffn_up_exps,
14431+
model.layers[il].ffn_gate_exps,
14432+
model.layers[il].ffn_down_exps,
14433+
nullptr,
14434+
n_expert, n_expert_used,
14435+
LLM_FFN_SILU, false,
14436+
false, 0.0,
14437+
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
14438+
il);
14439+
cb(cur_moe, "ffn_moe_out", il);
1445114440

14452-
cur = ggml_add(ctx0, ggml_add(ctx0, cur_moe, cur_mlp), ffn_inp);
14453-
cb(cur, "ffn_out", il);
14441+
ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp);
14442+
cb(ffn_out, "ffn_out", il);
14443+
14444+
cur = ggml_add(ctx0, ffn_out, ffn_inp);
1445414445

1445514446
cur = build_cvec(cur, il);
1445614447
cb(cur, "l_out", il);

0 commit comments

Comments
 (0)