From 218dcc572728c385cd914f985c236fe22cb5532a Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Mon, 5 Jan 2026 14:31:36 +0200 Subject: [PATCH] Split mode graph for Qwen3 (#1106) Co-authored-by: Iwan Kawrakow --- src/llama-build-context.cpp | 87 +++++++++++++++++++------------------ src/llama-load-tensors.cpp | 13 +++--- src/llama.cpp | 1 + 3 files changed, 51 insertions(+), 50 deletions(-) diff --git a/src/llama-build-context.cpp b/src/llama-build-context.cpp index 30c0f018..d7862de5 100644 --- a/src/llama-build-context.cpp +++ b/src/llama-build-context.cpp @@ -3904,64 +3904,71 @@ ggml_cgraph * llm_build_context::build_qwen3() { // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - auto rope_cache = cparams.rope_cache && (rope_type == LLAMA_ROPE_TYPE_NEOX || rope_type == LLAMA_ROPE_TYPE_NORM) ? - ggml_rope_cache(ctx0, inp_pos, nullptr, n_embd_head, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow) : nullptr; + ggml_tensor * rope_cache = nullptr; + if (model.split_mode != LLAMA_SPLIT_MODE_GRAPH && cparams.rope_cache && + (rope_type == LLAMA_ROPE_TYPE_NEOX || rope_type == LLAMA_ROPE_TYPE_NORM)) { + rope_cache = ggml_rope_cache(ctx0, inp_pos, nullptr, n_embd_head, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + } for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * inpSA = inpL; - // norm - cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); + if (!rope_cache) { + 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); + } else { - // self-attention - { - auto [Qcur, Kcur, Vcur] = llm_build_mul_mat_qkv(gf, cur, - model.layers[il].wqkv, nullptr, - model.layers[il].wqk, nullptr, - model.layers[il].wq, nullptr, - model.layers[il].wk, nullptr, - model.layers[il].wv, nullptr, - model.layers[il].attn_q_norm, model.layers[il].attn_k_norm, 0, il); + // norm + cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); - if (rope_cache) { - Qcur = ggml_rope_fast(ctx0, Qcur, rope_cache); - Kcur = ggml_rope_fast(ctx0, Kcur, rope_cache); - } else { - 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); + // self-attention + { + auto [Qcur, Kcur, Vcur] = llm_build_mul_mat_qkv(gf, cur, + model.layers[il].wqkv, nullptr, + model.layers[il].wqk, nullptr, + model.layers[il].wq, nullptr, + model.layers[il].wk, nullptr, + model.layers[il].wv, nullptr, + model.layers[il].attn_q_norm, model.layers[il].attn_k_norm, 0, il); + + if (rope_cache) { + Qcur = ggml_rope_fast(ctx0, Qcur, rope_cache); + Kcur = ggml_rope_fast(ctx0, Kcur, rope_cache); + } else { + 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); + + 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); + + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "attn_with_inp", il); } - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - - 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); } if (il == n_layer - 1) { // skip computing output for unused tokens struct 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); } - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - // feed-forward network - cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp, + 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); - cur = ggml_add(ctx0, cur, ffn_inp); cur = lctx.cvec.apply_to(ctx0, cur, il); cb(cur, "l_out", il); @@ -3969,13 +3976,7 @@ ggml_cgraph * llm_build_context::build_qwen3() { 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); diff --git a/src/llama-load-tensors.cpp b/src/llama-load-tensors.cpp index 707c8811..c4fb3518 100644 --- a/src/llama-load-tensors.cpp +++ b/src/llama-load-tensors.cpp @@ -1116,8 +1116,8 @@ bool create_tensors_helper::create_qwen3_tensors(const LLM_TN & tn) { // output { - model.output_norm = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); - model.output = create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + model.output_norm = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); // if output is NULL, init from the input tok embed if (model.output == NULL) { model.output = create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); @@ -1125,21 +1125,20 @@ bool create_tensors_helper::create_qwen3_tensors(const LLM_TN & tn) { } for (int i = 0; i < n_layer; ++i) { - ggml_context * ctx_layer = ctx_for_layer(i); ggml_context * ctx_split = ctx_for_layer_split(i); auto & layer = model.layers[i]; - layer.attn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); use_mmap_buffer &= !merge_qkv(tn, i, 0); layer.wo = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}); - layer.attn_k_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}); - layer.attn_q_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}); + layer.attn_k_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}); + layer.attn_q_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}); - layer.ffn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); create_std_ffn(i, tn, layer, n_ff, n_embd, ctx_split); } return use_mmap_buffer; diff --git a/src/llama.cpp b/src/llama.cpp index 49b697c7..2e706055 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -1731,6 +1731,7 @@ static bool is_model_split_supported(const llama_model & model) { LLM_ARCH_MISTRAL3, LLM_ARCH_COHERE2, LLM_ARCH_MIMO2, + LLM_ARCH_QWEN3, }; auto it = k_supported.find(model.arch); return it != k_supported.end();