From f20c04f01fcbc3fae3450242a85d982c6a57d01d Mon Sep 17 00:00:00 2001 From: Pierrick HYMBERT Date: Mon, 8 Apr 2024 17:45:35 +0200 Subject: [PATCH] llama: factorize moe graph implementation between grok, mixtral and dbrx --- llama.cpp | 234 +++++++++++++++--------------------------------------- 1 file changed, 63 insertions(+), 171 deletions(-) diff --git a/llama.cpp b/llama.cpp index 808a700ae..26dc24ebc 100644 --- a/llama.cpp +++ b/llama.cpp @@ -6457,62 +6457,7 @@ struct llm_build_context { LLM_NORM_RMS, cb, il); cb(cur, "ffn_norm", il); - ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts] - cb(logits, "ffn_moe_logits", il); - - ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts] - cb(probs, "ffn_moe_probs", il); - - // select experts - ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok] - cb(selected_experts->src[0], "ffn_moe_argsort", il); - - ggml_tensor * weights = ggml_get_rows(ctx0, - ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); - cb(weights, "ffn_moe_weights", il); - - weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok] - - ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); - cb(weights_sum, "ffn_moe_weights_sum", il); - - weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok] - cb(weights, "ffn_moe_weights_norm", il); - - // compute expert outputs - ggml_tensor * moe_out = nullptr; - - for (int i = 0; i < n_expert_used; ++i) { - ggml_tensor * cur_expert; - - ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur); - cb(cur_up, "ffn_moe_up", il); - - ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur); - cb(cur_gate, "ffn_moe_gate", il); - - cur_gate = ggml_silu(ctx0, cur_gate); - cb(cur_gate, "ffn_moe_silu", il); - - cur_expert = ggml_mul(ctx0, cur_up, cur_gate); - cb(cur_expert, "ffn_moe_gate_par", il); - - cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd] - cb(cur_expert, "ffn_moe_down", il); - - cur_expert = ggml_mul(ctx0, cur_expert, - ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0])); - cb(cur_expert, "ffn_moe_weighted", il); - - if (i == 0) { - moe_out = cur_expert; - } else { - moe_out = ggml_add(ctx0, moe_out, cur_expert); - cb(moe_out, "ffn_moe_out", il); - } - } - - cur = moe_out; + cur = build_moe(cur, n_tokens, il); } cur = ggml_add(ctx0, cur, ffn_inp); @@ -6544,6 +6489,65 @@ struct llm_build_context { return gf; } + ggml_tensor * build_moe(ggml_tensor * cur, int32_t n_tokens, int il) { + ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts] + cb(logits, "ffn_moe_logits", il); + + ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts] + cb(probs, "ffn_moe_probs", il); + + // select experts + ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok] + cb(selected_experts->src[0], "ffn_moe_argsort", il); + + ggml_tensor * weights = ggml_get_rows(ctx0, + ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); + cb(weights, "ffn_moe_weights", il); + + weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok] + + ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); + cb(weights_sum, "ffn_moe_weights_sum", il); + + weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok] + cb(weights, "ffn_moe_weights_norm", il); + + // compute expert outputs + ggml_tensor * moe_out = nullptr; + + for (int i = 0; i < n_expert_used; ++i) { + ggml_tensor * cur_expert; + + ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur); + cb(cur_up, "ffn_moe_up", il); + + ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur); + cb(cur_gate, "ffn_moe_gate", il); + + cur_gate = ggml_silu(ctx0, cur_gate); + cb(cur_gate, "ffn_moe_silu", il); + + cur_expert = ggml_mul(ctx0, cur_up, cur_gate); + cb(cur_expert, "ffn_moe_gate_par", il); + + cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd] + cb(cur_expert, "ffn_moe_down", il); + + cur_expert = ggml_mul(ctx0, cur_expert, + ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0])); + cb(cur_expert, "ffn_moe_weighted", il); + + if (i == 0) { + moe_out = cur_expert; + } else { + moe_out = ggml_add(ctx0, moe_out, cur_expert); + cb(moe_out, "ffn_moe_out", il); + } + } + + return moe_out; + } + struct ggml_cgraph * build_baichuan() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -6991,63 +6995,7 @@ struct llm_build_context { LLM_NORM_RMS, cb, il); cb(cur, "ffn_norm", il); - ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts] - cb(logits, "ffn_moe_logits", il); - - ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts] - cb(probs, "ffn_moe_probs", il); - - // select experts - ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok] - cb(selected_experts->src[0], "ffn_moe_argsort", il); - - ggml_tensor * weights = ggml_get_rows(ctx0, - ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); - cb(weights, "ffn_moe_weights", il); - - weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok] - - ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); - cb(weights_sum, "ffn_moe_weights_sum", il); - - weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok] - cb(weights, "ffn_moe_weights_norm", il); - - // compute expert outputs - ggml_tensor * moe_out = nullptr; - - for (int i = 0; i < n_expert_used; ++i) { - ggml_tensor * cur_expert; - - ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur); - cb(cur_up, "ffn_moe_up", il); - - ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur); - cb(cur_gate, "ffn_moe_gate", il); - - //GeLU - cur_gate = ggml_gelu(ctx0, cur_gate); - cb(cur_gate, "ffn_moe_gelu", il); - - cur_expert = ggml_mul(ctx0, cur_up, cur_gate); - cb(cur_expert, "ffn_moe_gate_par", il); - - cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd] - cb(cur_expert, "ffn_moe_down", il); - - cur_expert = ggml_mul(ctx0, cur_expert, - ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0])); - cb(cur_expert, "ffn_moe_weighted", il); - - if (i == 0) { - moe_out = cur_expert; - } else { - moe_out = ggml_add(ctx0, moe_out, cur_expert); - cb(moe_out, "ffn_moe_out", il); - } - } - - cur = moe_out; + cur = build_moe(cur, n_tokens, il); // Grok // if layer_out_norm is present then apply it before adding the input @@ -7163,7 +7111,6 @@ struct llm_build_context { cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } if (il == n_layer - 1) { @@ -7179,64 +7126,9 @@ struct llm_build_context { // feed-forward network // MoE branch - { - ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts] - cb(logits, "ffn_moe_logits", il); - - ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts] - cb(probs, "ffn_moe_probs", il); - - // select experts - ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok] - cb(selected_experts->src[0], "ffn_moe_argsort", il); - - ggml_tensor * weights = ggml_get_rows(ctx0, - ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); - cb(weights, "ffn_moe_weights", il); - - weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok] - - ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); - cb(weights_sum, "ffn_moe_weights_sum", il); - - weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok] - cb(weights, "ffn_moe_weights_norm", il); - - // compute expert outputs - ggml_tensor * moe_out = nullptr; - for (int i = 0; i < n_expert_used; ++i) { - ggml_tensor * cur_expert; - - ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur); - cb(cur_up, "ffn_moe_up", il); - - ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur); - cb(cur_gate, "ffn_moe_gate", il); - - //GeLU - cur_gate = ggml_gelu(ctx0, cur_gate); - cb(cur_gate, "ffn_moe_gelu", il); - - cur_expert = ggml_mul(ctx0, cur_up, cur_gate); - cb(cur_expert, "ffn_moe_gate_par", il); - - cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd] - cb(cur_expert, "ffn_moe_down", il); - - cur_expert = ggml_mul(ctx0, cur_expert, - ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i * weights->nb[0])); - cb(cur_expert, "ffn_moe_weighted", il); - - if (i == 0) { - moe_out = cur_expert; - } else { - moe_out = ggml_add(ctx0, moe_out, cur_expert); - cb(moe_out, "ffn_moe_out", il); - } - } - cur = moe_out; - } + cur = build_moe(cur, n_tokens, il); + // DBRX norm2 cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].layer_out_norm, NULL, LLM_NORM, cb, il);