From eaf2e1c15ad6f0a78b46731b49df053782197203 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Thu, 8 Jan 2026 16:46:41 +0200 Subject: [PATCH] Split mode "graph" for Ernie-4.5-MoE (#1121) * Ernie-4.5-MoE split mode graph * Cleanup --------- Co-authored-by: Iwan Kawrakow --- src/llama-build-context.cpp | 124 ++++++------------------------------ src/llama-load-tensors.cpp | 17 +++-- src/llama.cpp | 1 + 3 files changed, 28 insertions(+), 114 deletions(-) diff --git a/src/llama-build-context.cpp b/src/llama-build-context.cpp index 4615605c..7256bc7c 100644 --- a/src/llama-build-context.cpp +++ b/src/llama-build-context.cpp @@ -8178,122 +8178,42 @@ ggml_cgraph * llm_build_context::build_ernie4_5_moe() { GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0"); for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - // norm - // Pre-attention norm - cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - // self-attention - { - // Q, K, V projections - ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, 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 = llm_build_lora_mm(lctx, ctx0, 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); - } - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, 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); - } - - - // reshape for multi-head - 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); - - // apply RoPE - 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); - } + cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, nullptr, KQ_mask, nullptr, nullptr, + 1.0f/sqrtf(float(n_embd_head)), 0.0f, 0, il, true, false, true); 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); } - // residual connection for attention output - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - // feed-forward network bool is_moe_layer = static_cast(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0; if (!is_moe_layer) { - cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp, - model.layers[il].ffn_up, NULL, NULL, + // dense FFN + cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, cur, + model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + LLM_FFN_SILU, LLM_FFN_PAR, cb, il, gf, true); cb(cur, "ffn_out", il); - } - else { - // MoE branch - 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 * moe_out = llm_build_moe_ffn(ctx0, lctx, 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, + } else { + cur = llm_build_std_moe_ffn(ctx0, lctx, model.layers[il].ffn_norm, cur, + model.layers[il].ffn_gate_inp, model.layers[il].ffn_gate_inp_b, + model.layers[il].ffn_up_exps, model.layers[il].ffn_up_exps_b, + model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b, + model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b, model.layers[il].ffn_exp_probs_b, + model.layers[il].ffn_up_shexp, nullptr, // we don't have shared expert biases? + model.layers[il].ffn_gate_shexp, nullptr, + model.layers[il].ffn_down_shexp, nullptr, n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, + LLM_FFN_SILU, true, false, 0.0f, LLM_EXPERT_GATING_FUNC_SOFTMAX, - cb, il, gf); - cb(moe_out, "ffn_moe_out", il); - - // Shared expert (if present) - if (hparams.n_ff_shexp > 0) { - ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, nullptr, 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, cb, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - } - else { - cur = moe_out; - } - cb(cur, "ffn_out", il); + LLM_FFN_SILU, cb, il, gf, true); } - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - cur = lctx.cvec.apply_to(ctx0, cur, il); cb(cur, "l_out", il); @@ -8301,15 +8221,9 @@ ggml_cgraph * llm_build_context::build_ernie4_5_moe() { inpL = cur; } - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM_RMS, cb, -1); - - cb(cur, "result_norm", -1); - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - + cur = build_output(lctx, ctx0, inpL, model.output, model.output_norm, cb); cb(cur, "result_output", -1); + ggml_build_forward_expand(gf, cur); return gf; } diff --git a/src/llama-load-tensors.cpp b/src/llama-load-tensors.cpp index bc7e70e7..72f058c3 100644 --- a/src/llama-load-tensors.cpp +++ b/src/llama-load-tensors.cpp @@ -2490,10 +2490,9 @@ bool create_tensors_helper::create_ernie45_tensors(const LLM_TN & tn) { for (int i = 0; i < n_layer; ++i) { auto& layer = model.layers[i]; - ggml_context* ctx_layer = ctx_for_layer(i); ggml_context* ctx_split = ctx_for_layer_split(i); - layer.attn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0); + layer.attn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0); layer.wq = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0); layer.wk = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_gqa }, 0); @@ -2501,18 +2500,18 @@ bool create_tensors_helper::create_ernie45_tensors(const LLM_TN & tn) { layer.wo = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0); // optional bias tensors - layer.bq = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd }, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bk = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bv = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bo = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bq = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd }, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bk = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bv = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bo = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.ffn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0); + layer.ffn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0); if (model.arch == LLM_ARCH_ERNIE4_5_MOE && static_cast(i) >= hparams.n_layer_dense_lead) { // MoE layers int n_ff_exp = hparams.n_ff_exp; - layer.ffn_gate_inp = create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0); - layer.ffn_exp_probs_b = create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_gate_inp = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0); + layer.ffn_exp_probs_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, llama_model_loader::TENSOR_NOT_REQUIRED); layer.ffn_gate_exps = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, llama_model_loader::TENSOR_NOT_REQUIRED); layer.ffn_down_exps = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0); layer.ffn_up_exps = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0); diff --git a/src/llama.cpp b/src/llama.cpp index ce619ee2..5c28e224 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -1735,6 +1735,7 @@ static bool is_model_split_supported(const llama_model & model) { LLM_ARCH_QWEN3VL, LLM_ARCH_HUNYUAN_MOE, LLM_ARCH_OPENAI_MOE, + LLM_ARCH_ERNIE4_5_MOE, }; auto it = k_supported.find(model.arch); return it != k_supported.end();