From fb168ac5f7b085ad3649be5fcb1e602e55cdbff6 Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 17 Apr 2024 18:12:03 +0200 Subject: [PATCH] fix merge ggml-ci --- llama.cpp | 166 +++++++++++++++++++++++++++++++++++++++++++++++++----- 1 file changed, 153 insertions(+), 13 deletions(-) diff --git a/llama.cpp b/llama.cpp index 4de763405..8c7b7d3f6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -6103,6 +6103,7 @@ static struct ggml_tensor * llm_build_moe_ffn( int64_t n_expert, int64_t n_expert_used, llm_ffn_op_type type_op, + bool norm_w, const llm_build_cb & cb, int il) { int64_t n_embd = cur->ne[0]; @@ -6125,11 +6126,13 @@ static struct ggml_tensor * llm_build_moe_ffn( weights = ggml_reshape_2d(ctx, weights, n_expert_used, n_tokens); - ggml_tensor * weights_sum = ggml_sum_rows(ctx, weights); // [1, n_tokens] - cb(weights_sum, "ffn_moe_weights_sum", il); + if (norm_w) { + ggml_tensor * weights_sum = ggml_sum_rows(ctx, weights); // [1, n_tokens] + cb(weights_sum, "ffn_moe_weights_sum", il); - weights = ggml_div(ctx, weights, weights_sum); // [n_expert_used, n_tokens] - cb(weights, "ffn_moe_weights_norm", il); + weights = ggml_div(ctx, weights, weights_sum); // [n_expert_used, n_tokens] + cb(weights, "ffn_moe_weights_norm", il); + } cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens); ggml_tensor * up = ggml_mul_mat_id(ctx, up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] @@ -6180,6 +6183,7 @@ static struct ggml_tensor * llm_build_moe_ffn( return moe_out; } + // if max_alibi_bias > 0 then apply ALiBi static struct ggml_tensor * llm_build_kqv( struct ggml_context * ctx, @@ -6729,7 +6733,8 @@ struct llm_build_context { model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps, n_expert, n_expert_used, - LLM_FFN_SILU, cb, il); + LLM_FFN_SILU, true, + cb, il); cb(cur, "ffn_moe_out", il); } @@ -7215,7 +7220,8 @@ struct llm_build_context { model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps, n_expert, n_expert_used, - LLM_FFN_GELU, cb, il); + LLM_FFN_GELU, true, + cb, il); cb(cur, "ffn_moe_out", il); // Grok @@ -7243,6 +7249,138 @@ struct llm_build_context { 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 = ggml_mul_mat(ctx0, model.output, cur); + + // Grok + // multiply logits by output_multiplier_scale of 0.5773502691896257 + + cur = ggml_scale(ctx0, cur, 0.5773502691896257f); + + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_dbrx() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + 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, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = nullptr; + struct ggml_tensor * Kcur = nullptr; + struct ggml_tensor * Vcur = nullptr; + + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", 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))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + 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) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + 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 + // MoE branch + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].attn_out_norm, NULL, + LLM_NORM, cb, il); + cb(cur, "attn_out_norm", il); + + cur = llm_build_moe_ffn(ctx0, 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, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + cb, il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM, cb, -1); @@ -8397,12 +8535,6 @@ struct llm_build_context { Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); cb(Vcur, "Vcur", il); - // these nodes are added to the graph together so that they are not reordered - // by doing so, the number of splits in the graph is reduced - ggml_build_forward_expand(gf, Qcur); - ggml_build_forward_expand(gf, Kcur); - ggml_build_forward_expand(gf, Vcur); - Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, @@ -8553,7 +8685,15 @@ struct llm_build_context { LLM_NORM_RMS, cb, il); cb(cur, "ffn_norm", il); - ggml_tensor * moe_out = build_moe_ffn(cur, n_tokens, LLM_FFN_SILU, false, il); + ggml_tensor * moe_out = + llm_build_moe_ffn(ctx0, 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, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + cb, il); // FFN shared expert {