char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Yaniss Amazouzmodel: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (#25444)c61b98b87

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.5 KiB179 linesraw
1#include "models.h"
2
3void llama_model_qwen3moe::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
5    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
6
7    switch (hparams.n_layer()) {
8        case 48: type = LLM_TYPE_30B_A3B; break;
9        case 94: type = LLM_TYPE_235B_A22B; break;
10        default: type = LLM_TYPE_UNKNOWN;
11    }
12}
13
14void llama_model_qwen3moe::load_arch_tensors(llama_model_loader &) {
15    LLAMA_LOAD_LOCALS;
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}, TENSOR_NOT_REQUIRED);
22    // if output is NULL, init from the input tok embed
23    if (output == NULL) {
24        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
25    }
26
27    for (int i = 0; i < n_layer; ++i) {
28        auto & layer = layers[i];
29
30        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
31
32        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
33        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
34
35        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
36        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
37
38        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
39
40        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
41
42        if (n_expert == 0) {
43            throw std::runtime_error("n_expert must be > 0 for QWEN3MOE");
44        }
45        if (n_expert_used == 0) {
46            throw std::runtime_error("n_expert_used must be > 0 for QWEN3MOE");
47        }
48
49        // MoE branch
50        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
51
52        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
53        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
54        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
55    }
56}
57
58std::unique_ptr<llm_graph_context> llama_model_qwen3moe::build_arch_graph(const llm_graph_params & params) const {
59    return std::make_unique<graph>(*this, params);
60}
61
62llama_model_qwen3moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
63    const int64_t n_embd_head = hparams.n_embd_head_v();
64
65    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
66    GGML_ASSERT(n_embd_head == n_rot);
67
68    ggml_tensor * cur;
69    ggml_tensor * inpL;
70
71    inpL = build_inp_embd(model.tok_embd);
72
73    // inp_pos - contains the positions
74    ggml_tensor * inp_pos = build_inp_pos();
75
76    auto * inp_attn = build_attn_inp_kv();
77
78    ggml_tensor * inp_out_ids = build_inp_out_ids();
79
80    for (int il = 0; il < n_layer; ++il) {
81        res->t_layer_inp[il] = inpL;
82
83        ggml_tensor * inpSA = inpL;
84
85        // norm
86        cur = build_norm(inpL,
87                model.layers[il].attn_norm, NULL,
88                LLM_NORM_RMS, il);
89        cb(cur, "attn_norm", il);
90
91        // self_attention
92        {
93            // compute Q and K and RoPE them
94            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
95                    n_embd_head, n_head, n_head_kv, il);
96
97            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
98            cb(Qcur, "Qcur_normed", il);
99
100            Qcur = ggml_rope_ext(
101                    ctx0, Qcur, 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            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
107            cb(Kcur, "Kcur_normed", il);
108
109            Kcur = ggml_rope_ext(
110                    ctx0, Kcur, inp_pos, nullptr,
111                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
112                    ext_factor, attn_factor, beta_fast, beta_slow
113                    );
114
115            cb(Qcur, "Qcur", il);
116            cb(Kcur, "Kcur", il);
117            cb(Vcur, "Vcur", il);
118
119            cur = build_attn(inp_attn,
120                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
121                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
122        }
123        if (il == n_layer - 1 && inp_out_ids) {
124            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
125            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
126        }
127        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
128        cb(ffn_inp, "ffn_inp", il);
129
130        // MoE branch
131        cur = build_norm(ffn_inp,
132                model.layers[il].ffn_norm, NULL,
133                LLM_NORM_RMS, il);
134        cb(cur, "ffn_norm", il);
135
136        ggml_tensor * moe_out =
137            build_moe_ffn(cur,
138                    model.layers[il].ffn_gate_inp,
139                    model.layers[il].ffn_up_exps,
140                    model.layers[il].ffn_gate_exps,
141                    model.layers[il].ffn_down_exps,
142                    nullptr,
143                    n_expert, n_expert_used,
144                    LLM_FFN_SILU, true,
145                    hparams.expert_weights_scale,
146                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
147                    il,
148                    nullptr, nullptr,
149                    model.layers[il].ffn_up_exps_s,
150                    model.layers[il].ffn_gate_exps_s,
151                    model.layers[il].ffn_down_exps_s);
152        cb(moe_out, "ffn_moe_out", il);
153        cur = moe_out;
154
155        cur = ggml_add(ctx0, cur, ffn_inp);
156
157        cur = build_cvec(cur, il);
158        cb(cur, "l_out", il);
159
160        // input for next layer
161        inpL = cur;
162    }
163    cur = inpL;
164
165    cur = build_norm(cur,
166            model.output_norm, NULL,
167            LLM_NORM_RMS, -1);
168
169    cb(cur, "result_norm", -1);
170    res->t_embd = cur;
171
172    // lm_head
173    cur = build_lora_mm(model.output, cur, model.output_s);
174
175    cb(cur, "result_output", -1);
176    res->t_logits = cur;
177
178    ggml_build_forward_expand(gf, cur);
179}