diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 083366fd..7e162c55 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -92,36 +92,28 @@ void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) { } void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) { - if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) { - //GGML_ASSERT(out_ids && "every model that can must skip unused outputs"); + GGML_ASSERT(out_ids); - if (!out_ids) { - LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__); - } else { - const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_tokens = ubatch->n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); - int32_t * data = (int32_t *) out_ids->data; + GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); + int32_t * data = (int32_t *) out_ids->data; - if (n_outputs == n_tokens) { - for (int i = 0; i < n_tokens; ++i) { - data[i] = i; - } - } else if (ubatch->output) { - int32_t n_outputs = 0; - for (int i = 0; i < n_tokens; ++i) { - if (ubatch->output[i]) { - data[n_outputs++] = i; - } - } - // the graph needs to have been passed the correct number of outputs - GGML_ASSERT(n_outputs == n_outputs); - } else if (n_outputs == 1) { - // only keep last output - data[0] = n_tokens - 1; - } else { - GGML_ASSERT(n_outputs == 0); - } + if (n_outputs == n_tokens) { + for (int i = 0; i < n_tokens; ++i) { + data[i] = i; + } + + return; + } + + GGML_ASSERT(ubatch->output); + + int n_outputs = 0; + + for (int i = 0; i < n_tokens; ++i) { + if (ubatch->output[i]) { + data[n_outputs++] = i; } } } @@ -874,6 +866,14 @@ ggml_tensor * llm_graph_context::build_inp_attn_scale() const { } ggml_tensor * llm_graph_context::build_inp_out_ids() const { + // note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls, + // but this would make the graph topology depend on the number of output tokens, which can interere with + // features that require constant topology such as pipline parallelism + // ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471 + //if (n_outputs < n_tokens) { + // return nullptr; + //} + auto inp = std::make_unique(hparams, cparams, n_outputs); auto & cur = inp->out_ids; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index a5853f8b..e2c82017 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -4707,6 +4707,8 @@ struct llm_build_llama : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -4769,9 +4771,7 @@ struct llm_build_llama : public llm_graph_context { 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(); + 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); } @@ -4867,6 +4867,8 @@ struct llm_build_llama_iswa : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -4943,9 +4945,7 @@ struct llm_build_llama_iswa : public llm_graph_context { 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(); + 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); } @@ -5045,6 +5045,9 @@ struct llm_build_deci : public llm_graph_context { 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; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; const int64_t n_head_kv = hparams.n_head_kv(il); @@ -5118,9 +5121,7 @@ struct llm_build_deci : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -5199,6 +5200,8 @@ struct llm_build_baichuan : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5250,9 +5253,7 @@ struct llm_build_baichuan : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -5321,6 +5322,8 @@ struct llm_build_xverse : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5365,9 +5368,7 @@ struct llm_build_xverse : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -5435,6 +5436,8 @@ struct llm_build_falcon : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * attn_norm; @@ -5490,9 +5493,7 @@ struct llm_build_falcon : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids); @@ -5561,6 +5562,8 @@ struct llm_build_grok : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5620,9 +5623,7 @@ struct llm_build_grok : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -5721,6 +5722,8 @@ struct llm_build_dbrx : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5771,9 +5774,7 @@ struct llm_build_dbrx : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -5853,6 +5854,8 @@ struct llm_build_starcoder : public llm_graph_context { inpL = ggml_add(ctx0, inpL, pos); cb(inpL, "inpL", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -5885,9 +5888,7 @@ struct llm_build_starcoder : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -5952,6 +5953,8 @@ struct llm_build_refact : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5984,9 +5987,7 @@ struct llm_build_refact : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -6072,78 +6073,79 @@ struct llm_build_bert : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); - // iterate layers + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * cur = inpL; - ggml_tensor * Qcur; - ggml_tensor * Kcur; - ggml_tensor * Vcur; + { + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; - // self-attention - if (model.layers[il].wqkv) { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); + // self-attention + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); - if (model.layers[il].bqkv) { - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); } - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - } else { - Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); - Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); - Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + } + + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, 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); + + // RoPE + if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) { + 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 = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); } - if (model.layers[il].attn_q_norm) { - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, il); - } - - if (model.layers[il].attn_k_norm) { - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, 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); - - // RoPE - if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) { - 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 = build_attn(inp_attn, gf, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "kqv_out", il); - - if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6240,56 +6242,57 @@ struct llm_build_neo_bert : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); - // iterate layers + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * cur = inpL; - ggml_tensor * Qcur; - ggml_tensor * Kcur; - ggml_tensor * Vcur; - // pre-norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - // self-attention - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); + { + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + // self-attention + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", 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 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - // 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 - ); + 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); - 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 - ); + // 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 + ); - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); + 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 + ); - cur = build_attn(inp_attn, gf, - model.layers[il].wo, nullptr, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "kqv_out", il); + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); - if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = build_attn(inp_attn, gf, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6354,6 +6357,8 @@ struct llm_build_bloom : public llm_graph_context { LLM_NORM, -1); cb(inpL, "inp_norm", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -6386,9 +6391,7 @@ struct llm_build_bloom : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6465,6 +6468,8 @@ struct llm_build_mpt : public llm_graph_context { cb(inpL, "inpL", -1); } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * attn_norm; @@ -6527,9 +6532,7 @@ struct llm_build_mpt : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6598,6 +6601,8 @@ struct llm_build_stablelm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -6673,9 +6678,7 @@ struct llm_build_stablelm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); @@ -6750,6 +6753,8 @@ struct llm_build_qwen : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -6796,9 +6801,7 @@ struct llm_build_qwen : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -6867,6 +6870,8 @@ struct llm_build_qwen2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -6916,9 +6921,7 @@ struct llm_build_qwen2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -6988,6 +6991,8 @@ struct llm_build_qwen2vl : public llm_graph_context { int sections[4]; std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7037,9 +7042,7 @@ struct llm_build_qwen2vl : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -7106,6 +7109,8 @@ struct llm_build_qwen2moe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7164,9 +7169,7 @@ struct llm_build_qwen2moe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -7265,6 +7268,8 @@ struct llm_build_qwen3 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7317,9 +7322,7 @@ struct llm_build_qwen3 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -7386,6 +7389,8 @@ struct llm_build_qwen3moe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7438,9 +7443,7 @@ struct llm_build_qwen3moe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -7516,6 +7519,8 @@ struct llm_build_phi2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { attn_norm_output = build_norm(inpL, model.layers[il].attn_norm, @@ -7578,9 +7583,7 @@ struct llm_build_phi2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); @@ -7652,6 +7655,8 @@ struct llm_build_phi3 : public llm_graph_context { inp_attn = build_attn_inp_kv_unified(); } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { auto * residual = inpL; @@ -7715,9 +7720,7 @@ struct llm_build_phi3 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor* inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); residual = ggml_get_rows(ctx0, residual, inp_out_ids); } @@ -7803,15 +7806,16 @@ struct llm_build_plamo : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_norm", il); - ggml_tensor * attention_norm = cur; + ggml_tensor * sa_inp = cur; // self-attention { @@ -7849,18 +7853,17 @@ struct llm_build_plamo : public llm_graph_context { model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - ggml_tensor * sa_out = cur; - cur = attention_norm; - - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); - sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids); + sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } + ggml_tensor * sa_out = cur; + + cur = sa_inp; + // feed-forward network { cur = build_ffn(cur, @@ -7925,6 +7928,8 @@ struct llm_build_gpt2 : public llm_graph_context { inpL = ggml_add(ctx0, inpL, pos); cb(inpL, "inpL", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -7957,9 +7962,7 @@ struct llm_build_gpt2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8029,6 +8032,8 @@ struct llm_build_codeshell : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -8073,9 +8078,7 @@ struct llm_build_codeshell : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8129,128 +8132,128 @@ struct llm_build_codeshell : public llm_graph_context { struct llm_build_orion : public llm_graph_context { llm_build_orion(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; + 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_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); - ggml_tensor * cur; - ggml_tensor * inpL; + ggml_tensor * cur; + ggml_tensor * inpL; - inpL = build_inp_embd(model.tok_embd); + inpL = build_inp_embd(model.tok_embd); - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); - auto * inp_attn = build_attn_inp_kv_unified(); + auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; + ggml_tensor * inp_out_ids = build_inp_out_ids(); - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; - // self-attention - { - // compute Q and K and 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); - // } + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", 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); - // } + // self-attention + { + // compute Q and K and 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 * 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); - // } + 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); + // } - 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); + 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_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 - ); + 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); - 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 - ); + 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 + ); - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); + 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 + ); - cur = build_attn(inp_attn, gf, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), 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); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(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, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; } - 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); - } + cur = inpL; - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); + cb(cur, "result_norm", -1); + res->t_embd = cur; - cur = build_ffn(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, il); - cb(cur, "ffn_out", il); + // lm_head + cur = build_lora_mm(model.output, cur); - cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "result_output", -1); + res->t_logits = cur; - 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, model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); + ggml_build_forward_expand(gf, cur); } }; @@ -8271,6 +8274,8 @@ struct llm_build_internlm2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -8329,9 +8334,7 @@ struct llm_build_internlm2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -8407,6 +8410,8 @@ struct llm_build_minicpm3 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -8526,15 +8531,13 @@ struct llm_build_minicpm3 : public llm_graph_context { q_states, k_states, v_states, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } // scale_res - scale the hidden states for residual connection - const float scale_res = scale_depth/sqrtf(float(n_layer)); + const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct? cur = ggml_scale(ctx0, cur, scale_res); cb(cur, "hidden_scaled", il); @@ -8611,6 +8614,8 @@ struct llm_build_gemma : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -8656,9 +8661,7 @@ struct llm_build_gemma : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8727,6 +8730,8 @@ struct llm_build_gemma2_iswa : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -8771,18 +8776,16 @@ struct llm_build_gemma2_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", 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); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); cb(sa_out, "sa_out", il); @@ -8861,6 +8864,8 @@ struct llm_build_gemma3_iswa : public llm_graph_context { // TODO: is causal == true correct? might need some changes auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const float freq_base_l = model.get_rope_freq_base (cparams, il); const float freq_scale_l = model.get_rope_freq_scale(cparams, il); @@ -8913,18 +8918,16 @@ struct llm_build_gemma3_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", 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); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); cb(sa_out, "sa_out", il); @@ -8995,6 +8998,8 @@ struct llm_build_starcoder2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9053,9 +9058,7 @@ struct llm_build_starcoder2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -9118,6 +9121,8 @@ struct llm_build_mamba : public llm_graph_context { auto * rs_inp = build_rs_inp(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -9127,9 +9132,7 @@ struct llm_build_mamba : public llm_graph_context { cur = build_mamba_layer(rs_inp, gf, cur, ubatch, il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -9311,13 +9314,15 @@ struct llm_build_command_r : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); cb(cur, "attn_norm", il); + ggml_tensor * ffn_inp = cur; // self-attention @@ -9385,9 +9390,7 @@ struct llm_build_command_r : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); @@ -9458,6 +9461,8 @@ struct llm_build_cohere2_iswa : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const bool is_swa = hparams.is_swa(il); @@ -9520,9 +9525,7 @@ struct llm_build_cohere2_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); @@ -9593,6 +9596,8 @@ struct llm_build_olmo : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9651,9 +9656,7 @@ struct llm_build_olmo : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -9721,6 +9724,8 @@ struct llm_build_olmo2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9771,18 +9776,16 @@ struct llm_build_olmo2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), 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); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", 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); - } - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); @@ -9850,6 +9853,8 @@ struct llm_build_olmoe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9904,9 +9909,7 @@ struct llm_build_olmoe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -9976,6 +9979,8 @@ struct llm_build_openelm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const int64_t n_head = hparams.n_head(il); const int64_t n_head_kv = hparams.n_head_kv(il); @@ -10037,11 +10042,9 @@ struct llm_build_openelm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { residual = ggml_get_rows(ctx0, residual, inp_out_ids); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); } ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); @@ -10107,6 +10110,8 @@ struct llm_build_gptneox : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -10151,9 +10156,7 @@ struct llm_build_gptneox : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -10255,6 +10258,8 @@ struct llm_build_arctic : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10301,9 +10306,7 @@ struct llm_build_arctic : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -10395,6 +10398,8 @@ struct llm_build_deepseek : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10456,14 +10461,11 @@ struct llm_build_deepseek : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); @@ -10571,6 +10573,8 @@ struct llm_build_deepseek2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10720,9 +10724,7 @@ struct llm_build_deepseek2 : public llm_graph_context { } } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -10818,6 +10820,8 @@ struct llm_build_bitnet : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10900,9 +10904,7 @@ struct llm_build_bitnet : public llm_graph_context { cb(cur, "attn_o_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -10977,6 +10979,8 @@ struct llm_build_t5_enc : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11010,9 +11014,7 @@ struct llm_build_t5_enc : public llm_graph_context { cb(cur, "kqv_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -11083,6 +11085,8 @@ struct llm_build_t5_dec : public llm_graph_context { auto * inp_attn_self = build_attn_inp_kv_unified(); auto * inp_attn_cross = build_attn_inp_cross(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11174,11 +11178,8 @@ struct llm_build_t5_dec : public llm_graph_context { //cb(cur, "kqv_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); } @@ -11248,6 +11249,8 @@ struct llm_build_jais : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -11280,9 +11283,7 @@ struct llm_build_jais : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -11346,6 +11347,8 @@ struct llm_build_chatglm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11412,9 +11415,7 @@ struct llm_build_chatglm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -11479,6 +11480,8 @@ struct llm_build_glm4 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11545,9 +11548,7 @@ struct llm_build_glm4 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -11630,6 +11631,8 @@ struct llm_build_nemotron : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11689,9 +11692,7 @@ struct llm_build_nemotron : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -11759,6 +11760,8 @@ struct llm_build_exaone : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11820,9 +11823,7 @@ struct llm_build_exaone : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -12098,6 +12099,8 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12139,13 +12142,16 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base { ); ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); - ffn_norm = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens), inp_out_ids); - x_prev = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens), inp_out_ids); - cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); + x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); + x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); } cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6); @@ -12193,6 +12199,8 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12217,11 +12225,12 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); cb(ffn_inp, "ffn_inp", 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, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); } // feed-forward network @@ -12447,6 +12456,8 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12488,12 +12499,14 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base { ); ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); - ffn_norm = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens), inp_out_ids); - x_prev = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens), inp_out_ids); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); + x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); + x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); } cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7); @@ -12538,6 +12551,8 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12562,11 +12577,12 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base { ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); cb(ffn_inp, "ffn_inp", 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, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); } // feed-forward network @@ -12635,6 +12651,9 @@ struct llm_build_granite : public llm_graph_context { 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; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -12697,9 +12716,7 @@ struct llm_build_granite : public llm_graph_context { 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(); + 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); } @@ -12818,6 +12835,8 @@ struct llm_build_chameleon : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -12894,21 +12913,19 @@ struct llm_build_chameleon : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, nullptr, Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - - if (hparams.swin_norm) { - cur = build_norm(cur, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - } } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); @@ -13149,6 +13166,8 @@ struct llm_build_plm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13252,9 +13271,7 @@ struct llm_build_plm : public llm_graph_context { q_states, k_states, v_states, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -13314,6 +13331,8 @@ struct llm_build_bailingmoe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13375,9 +13394,7 @@ struct llm_build_bailingmoe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -13463,6 +13480,8 @@ struct llm_build_dots1 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13515,9 +13534,7 @@ struct llm_build_dots1 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + 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); } @@ -13615,6 +13632,8 @@ struct llm_build_arcee : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13677,9 +13696,7 @@ struct llm_build_arcee : public llm_graph_context { 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(); + 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); }