| File: | root/firefox-clang/third_party/llama.cpp/src/models/minimax-m2.cpp |
| Warning: | line 65, column 9 Value stored to 'cur' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | #include "models.h" |
| 2 | |
| 3 | void llama_model_minimax_m2::load_arch_hparams(llama_model_loader & ml) { |
| 4 | ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); |
| 5 | ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); |
| 6 | ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false); |
| 7 | |
| 8 | switch (hparams.n_layer()) { |
| 9 | case 62: type = LLM_TYPE_230B_A10B; break; |
| 10 | default: type = LLM_TYPE_UNKNOWN; |
| 11 | } |
| 12 | } |
| 13 | |
| 14 | void llama_model_minimax_m2::load_arch_tensors(llama_model_loader &) { |
| 15 | LLAMA_LOAD_LOCALSconst int n_layer = hparams.n_layer(); (void)(n_layer); const int n_layer_all = hparams.n_layer_all; (void)(n_layer_all); const int n_layer_nextn = hparams.n_layer_nextn; (void)(n_layer_nextn ); const int64_t n_head = hparams.n_head(); (void)(n_head); const int64_t n_head_kv = hparams.n_head_kv(); (void)(n_head_kv); const int64_t n_embd = hparams.n_embd; (void)(n_embd); const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); (void)(n_embd_k_gqa); const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); (void)( n_embd_v_gqa); const int64_t n_embd_head_k = hparams.n_embd_head_k (); (void)(n_embd_head_k); const int64_t n_embd_head_v = hparams .n_embd_head_v(); (void)(n_embd_head_v); const int64_t n_ff = hparams.n_ff(); (void)(n_ff); const int64_t n_embd_gqa = n_embd_v_gqa ; (void)(n_embd_gqa); const int64_t n_vocab = vocab.n_tokens( ); (void)(n_vocab); const int64_t n_token_types = vocab.n_token_types (); (void)(n_token_types); const int64_t n_rot = hparams.n_rot (); (void)(n_rot); const int64_t n_expert = hparams.n_expert; (void)(n_expert); const int64_t n_expert_used = hparams.n_expert_used ; (void)(n_expert_used); const int64_t n_ctx_train = hparams. n_ctx_train; (void)(n_ctx_train);; |
| 16 | |
| 17 | tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); |
| 18 | |
| 19 | // output |
| 20 | output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); |
| 21 | output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); |
| 22 | |
| 23 | for (int i = 0; i < n_layer; ++i) { |
| 24 | auto & layer = layers[i]; |
| 25 | |
| 26 | create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0); |
| 27 | layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0); |
| 28 | |
| 29 | layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); |
| 30 | layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k * n_head}, 0); |
| 31 | layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_k_gqa}, 0); |
| 32 | |
| 33 | layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); |
| 34 | |
| 35 | layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); |
| 36 | layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); |
| 37 | layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff, n_embd, n_expert}, 0); |
| 38 | layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); |
| 39 | layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0); |
| 40 | } |
| 41 | } |
| 42 | |
| 43 | std::unique_ptr<llm_graph_context> llama_model_minimax_m2::build_arch_graph(const llm_graph_params & params) const { |
| 44 | return std::make_unique<graph>(*this, params); |
| 45 | } |
| 46 | |
| 47 | llama_model_minimax_m2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { |
| 48 | const int64_t n_embd_head = hparams.n_embd_head_v(); |
| 49 | |
| 50 | GGML_ASSERT(n_embd_head == hparams.n_embd_head_k())if (!(n_embd_head == hparams.n_embd_head_k())) ggml_abort("/root/firefox-clang/third_party/llama.cpp/src/models/minimax-m2.cpp" , 50, "GGML_ASSERT(%s) failed", "n_embd_head == hparams.n_embd_head_k()" ); |
| 51 | // GGML_ASSERT(n_embd_head == n_rot); this is wrong in case of minimax, head_dim = 128, n_rot = 64 |
| 52 | |
| 53 | ggml_tensor * cur; |
| 54 | ggml_tensor * inpL; |
| 55 | |
| 56 | inpL = build_inp_embd(model.tok_embd); |
| 57 | |
| 58 | ggml_tensor * inp_pos = build_inp_pos(); |
| 59 | auto inp_attn = build_attn_inp_kv(); |
| 60 | ggml_tensor * inp_out_ids = build_inp_out_ids(); |
| 61 | |
| 62 | for (int il = 0; il < n_layer; ++il) { |
| 63 | ggml_tensor * inpSA = inpL; |
| 64 | |
| 65 | cur = inpL; |
Value stored to 'cur' is never read | |
| 66 | |
| 67 | // self_attention |
| 68 | { |
| 69 | cur = build_norm(inpL, model.layers[il].attn_norm, NULL__null, LLM_NORM_RMS, il); |
| 70 | cb(cur, "attn_norm", il); |
| 71 | |
| 72 | // compute Q and K and RoPE them |
| 73 | ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); |
| 74 | cb(Qcur, "Qcur", il); |
| 75 | |
| 76 | ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); |
| 77 | cb(Kcur, "Kcur", il); |
| 78 | |
| 79 | ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); |
| 80 | cb(Vcur, "Vcur", il); |
| 81 | |
| 82 | Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL__null, |
| 83 | LLM_NORM_RMS, il); |
| 84 | cb(Qcur, "Qcur_normed", il); |
| 85 | |
| 86 | Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL__null, |
| 87 | LLM_NORM_RMS, il); |
| 88 | cb(Kcur, "Kcur_normed", il); |
| 89 | |
| 90 | Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); |
| 91 | Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); |
| 92 | Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); |
| 93 | |
| 94 | Qcur = ggml_rope_ext( |
| 95 | ctx0, Qcur, inp_pos, nullptr, |
| 96 | n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, |
| 97 | ext_factor, attn_factor, beta_fast, beta_slow |
| 98 | ); |
| 99 | |
| 100 | Kcur = ggml_rope_ext( |
| 101 | ctx0, Kcur, inp_pos, nullptr, |
| 102 | n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, |
| 103 | ext_factor, attn_factor, beta_fast, beta_slow |
| 104 | ); |
| 105 | |
| 106 | cb(Qcur, "Qcur", il); |
| 107 | cb(Kcur, "Kcur", il); |
| 108 | cb(Vcur, "Vcur", il); |
| 109 | |
| 110 | cur = build_attn(inp_attn, |
| 111 | model.layers[il].wo, NULL__null, model.layers[il].wo_s, |
| 112 | Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); |
| 113 | } |
| 114 | |
| 115 | if (il == n_layer - 1 && inp_out_ids) { |
| 116 | cur = ggml_get_rows(ctx0, cur, inp_out_ids); |
| 117 | inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); |
| 118 | } |
| 119 | |
| 120 | ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); |
| 121 | cb(ffn_inp, "ffn_inp", il); |
| 122 | |
| 123 | // MoE branch |
| 124 | cur = build_norm(ffn_inp, |
| 125 | model.layers[il].ffn_norm, NULL__null, |
| 126 | LLM_NORM_RMS, il); |
| 127 | cb(cur, "ffn_norm", il); |
| 128 | |
| 129 | cur = build_moe_ffn(cur, |
| 130 | model.layers[il].ffn_gate_inp, |
| 131 | model.layers[il].ffn_up_exps, |
| 132 | model.layers[il].ffn_gate_exps, |
| 133 | model.layers[il].ffn_down_exps, |
| 134 | model.layers[il].ffn_exp_probs_b, |
| 135 | n_expert, n_expert_used, |
| 136 | LLM_FFN_SILU, true, |
| 137 | hparams.expert_weights_scale, |
| 138 | (llama_expert_gating_func_type) hparams.expert_gating_func, |
| 139 | il); |
| 140 | cb(cur, "ffn_moe_out", il); |
| 141 | |
| 142 | cur = ggml_add(ctx0, cur, ffn_inp); |
| 143 | |
| 144 | cur = build_cvec(cur, il); |
| 145 | cb(cur, "l_out", il); |
| 146 | |
| 147 | // input for next layer |
| 148 | inpL = cur; |
| 149 | } |
| 150 | |
| 151 | cur = inpL; |
| 152 | |
| 153 | cur = build_norm(cur, |
| 154 | model.output_norm, NULL__null, |
| 155 | LLM_NORM_RMS, -1); |
| 156 | |
| 157 | cb(cur, "result_norm", -1); |
| 158 | res->t_embd = cur; |
| 159 | |
| 160 | // lm_head |
| 161 | cur = build_lora_mm(model.output, cur, model.output_s); |
| 162 | |
| 163 | cb(cur, "result_output", -1); |
| 164 | res->t_logits = cur; |
| 165 | |
| 166 | ggml_build_forward_expand(gf, cur); |
| 167 | } |