nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Use standard graph helpers for MiniMax-M3

This commit is contained in:
Jun Yamog 2026-06-15 01:57:09 +00:00
parent c538210e6d
commit c08d194edd
3 changed files with 33 additions and 87 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

111
src/graphs/build_minimaxm3.cpp
View File

@ -7,117 +7,54 @@ ggml_cgraph* llm_build_context::build_minimaxm3() {
const int64_t n_embd_head = hparams.n_embd_head_v(0);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k(0));
constexpr float swiglu_alpha = 1.702f;
constexpr float swiglu_limit = 7.0f;
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
ggml_tensor * inp_pos = build_inp_pos();
ggml_tensor * inp_out_ids = build_inp_out_ids();
ggml_tensor * inp_out_ids = n_tokens > 1 ? build_inp_out_ids() : nullptr;
ggml_tensor * KQ_mask = build_inp_KQ_mask();
for (int il = 0; il < n_layer; ++il) {
GGML_ASSERT(model.split_mode != LLAMA_SPLIT_MODE_GRAPH && model.split_mode != LLAMA_SPLIT_MODE_ATTN);
ggml_tensor * inpSA = inpL;
cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
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);
Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, cb, il);
cb(Qcur, "Qcur_normed", il);
Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, cb, il);
cb(Kcur, "Kcur_normed", il);
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,
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, nullptr,
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 = llm_build_kv(ctx0, lctx, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask,
n_tokens, kv_head, n_kv,
1.0f / sqrtf(float(n_embd_head)), cb, il);
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
cur = llm_build_norm(ctx0, ffn_inp, hparams, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il);
ggml_tensor * ffn_inp = build_std_attention(gf, model.layers[il].attn_norm, inpL,
inp_pos, il == n_layer - 1 ? inp_out_ids : nullptr, nullptr,
KQ_mask, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), 0.0f, 0,
il, true, false, true);
if ((uint32_t) il < hparams.n_layer_dense_lead) {
ggml_tensor * gate = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_gate, cur);
cb(gate, "ffn_gate", il);
ggml_tensor * up = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_up, cur);
cb(up, "ffn_up", il);
gate = ggml_swiglu_oai(ctx0, gate, up, swiglu_alpha, swiglu_limit);
cb(gate, "ffn_gate_par", il);
cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_down, gate);
cb(cur, "ffn_down", il);
cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
model.layers[il].ffn_up, nullptr, nullptr,
model.layers[il].ffn_gate, nullptr, nullptr,
model.layers[il].ffn_down, nullptr, nullptr,
nullptr,
LLM_FFN_SWIGLU_OAI, LLM_FFN_PAR, cb, il, gf, true);
} else {
ggml_tensor * moe_out = llm_build_moe_ffn(ctx0, lctx, cur,
cur = llm_build_std_moe_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
model.layers[il].ffn_gate_inp,
nullptr,
model.layers[il].ffn_up_exps,
nullptr,
model.layers[il].ffn_gate_exps,
nullptr,
model.layers[il].ffn_down_exps,
nullptr,
model.layers[il].ffn_exp_probs_b,
model.layers[il].ffn_up_shexp,
nullptr,
model.layers[il].ffn_gate_shexp,
nullptr,
model.layers[il].ffn_down_shexp,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SWIGLU_OAI_MOE,
hparams.expert_weights_norm,
hparams.expert_weights_scale != 0.0f, hparams.expert_weights_scale,
(llm_expert_gating_func_type) hparams.expert_gating_func,
cb, il, gf);
cb(moe_out, "ffn_moe_out", il);
ggml_tensor * gate = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_gate_shexp, cur);
cb(gate, "ffn_shexp_gate", il);
ggml_tensor * up = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_up_shexp, cur);
cb(up, "ffn_shexp_up", il);
gate = ggml_swiglu_oai(ctx0, gate, up, swiglu_alpha, swiglu_limit);
cb(gate, "ffn_shexp_gate_par", il);
ggml_tensor * ffn_shexp = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_down_shexp, gate);
cb(ffn_shexp, "ffn_shexp", il);
cur = ggml_add(ctx0, moe_out, ffn_shexp);
cb(cur, "ffn_out", il);
LLM_FFN_SWIGLU_OAI,
cb, il, gf, true);
}
cur = ggml_add(ctx0, cur, ffn_inp);
cur = lctx.cvec.apply_to(ctx0, cur, il);
cb(cur, "l_out", il);

8
src/llama-build-context.cpp
View File

@ -962,6 +962,14 @@ ggml_tensor * llm_build_context::llm_build_ffn(
cur = ggml_swiglu(ctx, cur);
cb(cur, "ffn_swiglu", il);
} break;
case LLM_FFN_SWIGLU_OAI:
{
constexpr float alpha = 1.702f;
constexpr float limit = 7.0f;
cur = ggml_swiglu_oai(ctx, cur, tmp, alpha, limit);
cb(cur, "ffn_swiglu_oai", il);
type_gate = LLM_FFN_SEQ;
} break;
default:
GGML_ABORT("fatal error");
}

1
src/llama-build-context.h
View File

@ -24,6 +24,7 @@ enum llm_ffn_op_type {
LLM_FFN_RELU,
LLM_FFN_RELU_SQR,
LLM_FFN_SWIGLU,
LLM_FFN_SWIGLU_OAI,
LLM_FFN_SWIGLU_OAI_MOE,
};