| 1 | #include "models.h" |
| 2 | |
| 3 | template<bool iswa> |
| 4 | llm_build_phi3<iswa>::llm_build_phi3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { |
| 5 | const int64_t n_embd_head = hparams.n_embd_head_v; |
| 6 | const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); |
| 7 | |
| 8 | GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); |
| 9 | |
| 10 | ggml_tensor * cur; |
| 11 | ggml_tensor * inpL; |
| 12 | |
| 13 | inpL = build_inp_embd(tok_embd: model.tok_embd); |
| 14 | |
| 15 | // inp_pos - contains the positions |
| 16 | ggml_tensor * inp_pos = build_inp_pos(); |
| 17 | |
| 18 | using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>; |
| 19 | inp_attn_type * inp_attn = nullptr; |
| 20 | |
| 21 | if constexpr (iswa) { |
| 22 | inp_attn = build_attn_inp_kv_iswa(); |
| 23 | } else { |
| 24 | inp_attn = build_attn_inp_kv(); |
| 25 | } |
| 26 | ggml_tensor * inp_out_ids = build_inp_out_ids(); |
| 27 | |
| 28 | for (int il = 0; il < n_layer; ++il) { |
| 29 | auto * residual = inpL; |
| 30 | |
| 31 | // self-attention |
| 32 | { |
| 33 | // rope freq factors for 128k context |
| 34 | ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); |
| 35 | |
| 36 | ggml_tensor* attn_norm_output = build_norm(cur: inpL, |
| 37 | mw: model.layers[il].attn_norm, |
| 38 | mb: model.layers[il].attn_norm_b, |
| 39 | type: LLM_NORM_RMS, il); |
| 40 | cb(cur: attn_norm_output, name: "attn_norm" , il); |
| 41 | |
| 42 | ggml_tensor * Qcur = nullptr; |
| 43 | ggml_tensor * Kcur = nullptr; |
| 44 | ggml_tensor * Vcur = nullptr; |
| 45 | |
| 46 | if (model.layers[il].wqkv) { |
| 47 | cur = build_lora_mm(w: model.layers[il].wqkv, cur: attn_norm_output); |
| 48 | cb(cur, name: "wqkv" , il); |
| 49 | |
| 50 | Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 0 * sizeof(float) * (n_embd)); |
| 51 | Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd)); |
| 52 | Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)); |
| 53 | } |
| 54 | else { |
| 55 | Qcur = ggml_add(ctx0, build_lora_mm(w: model.layers[il].wq, cur: attn_norm_output), model.layers[il].bq); |
| 56 | Kcur = ggml_add(ctx0, build_lora_mm(w: model.layers[il].wk, cur: attn_norm_output), model.layers[il].bk); |
| 57 | Vcur = ggml_add(ctx0, build_lora_mm(w: model.layers[il].wv, cur: attn_norm_output), model.layers[il].bv); |
| 58 | |
| 59 | Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); |
| 60 | Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); |
| 61 | Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); |
| 62 | } |
| 63 | Qcur = ggml_rope_ext( |
| 64 | ctx0, Qcur, inp_pos, rope_factors, |
| 65 | n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, |
| 66 | ext_factor, attn_factor, beta_fast, beta_slow |
| 67 | ); |
| 68 | |
| 69 | Kcur = ggml_rope_ext( |
| 70 | ctx0, Kcur, inp_pos, rope_factors, |
| 71 | n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, |
| 72 | ext_factor, attn_factor, beta_fast, beta_slow |
| 73 | ); |
| 74 | |
| 75 | cb(cur: Qcur, name: "Qcur" , il); |
| 76 | cb(cur: Kcur, name: "Kcur" , il); |
| 77 | cb(cur: Vcur, name: "Vcur" , il); |
| 78 | |
| 79 | Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(x: float(n_embd_head))); |
| 80 | cb(cur: Qcur, name: "Qcur" , il); |
| 81 | |
| 82 | cur = build_attn(inp_attn, |
| 83 | model.layers[il].wo, model.layers[il].bo, |
| 84 | Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); |
| 85 | } |
| 86 | if (il == n_layer - 1 && inp_out_ids) { |
| 87 | cur = ggml_get_rows(ctx0, cur, inp_out_ids); |
| 88 | residual = ggml_get_rows(ctx0, residual, inp_out_ids); |
| 89 | } |
| 90 | cur = ggml_add(ctx0, cur, residual); |
| 91 | residual = cur; |
| 92 | |
| 93 | cur = build_norm(cur, |
| 94 | mw: model.layers[il].ffn_norm, mb: model.layers[il].ffn_norm_b, |
| 95 | type: LLM_NORM_RMS, il); |
| 96 | cb(cur, name: "ffn_norm" , il); |
| 97 | |
| 98 | // feed-forward network |
| 99 | if (model.layers[il].ffn_gate_inp == nullptr) { |
| 100 | cur = build_ffn(cur, |
| 101 | up: model.layers[il].ffn_up, NULL, NULL, |
| 102 | NULL, NULL, NULL, |
| 103 | down: model.layers[il].ffn_down, NULL, NULL, |
| 104 | NULL, |
| 105 | type_op: LLM_FFN_SWIGLU, type_gate: LLM_FFN_SEQ, il); |
| 106 | cb(cur, name: "ffn_out" , il); |
| 107 | } else { |
| 108 | // MoE branch |
| 109 | cur = build_moe_ffn(cur, |
| 110 | model.layers[il].ffn_gate_inp, |
| 111 | model.layers[il].ffn_up_exps, |
| 112 | model.layers[il].ffn_gate_exps, |
| 113 | model.layers[il].ffn_down_exps, |
| 114 | nullptr, |
| 115 | n_expert, n_expert_used, |
| 116 | LLM_FFN_SILU, true, |
| 117 | false, 0.0, |
| 118 | LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, |
| 119 | il); |
| 120 | cb(cur, name: "ffn_moe_out" , il); |
| 121 | } |
| 122 | cur = ggml_add(ctx0, residual, cur); |
| 123 | |
| 124 | cur = build_cvec(cur, il); |
| 125 | cb(cur, name: "l_out" , il); |
| 126 | |
| 127 | // input for next layer |
| 128 | inpL = cur; |
| 129 | } |
| 130 | cur = build_norm(cur: inpL, |
| 131 | mw: model.output_norm, |
| 132 | mb: model.output_norm_b, |
| 133 | type: LLM_NORM_RMS, il: -1); |
| 134 | |
| 135 | cb(cur, name: "result_norm" , il: -1); |
| 136 | res->t_embd = cur; |
| 137 | |
| 138 | cur = build_lora_mm(w: model.output, cur); |
| 139 | |
| 140 | if (model.output_b != nullptr) { |
| 141 | cb(cur, name: "result_output_no_bias" , il: -1); |
| 142 | cur = ggml_add(ctx0, cur, model.output_b); |
| 143 | } |
| 144 | cb(cur, name: "result_output" , il: -1); |
| 145 | res->t_logits = cur; |
| 146 | |
| 147 | ggml_build_forward_expand(gf, cur); |
| 148 | } |
| 149 | |
| 150 | // Explicit template instantiations |
| 151 | template struct llm_build_phi3<false>; |
| 152 | template struct llm_build_phi3<true>; |
| 153 | |