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 →
7.0 KiB190 linesraw
1#include "models.h"
2
3void llama_model_qwen3vlmoe::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false);
5    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
6    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
7    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
8
9    switch (hparams.n_layer()) {
10        case 48: type = LLM_TYPE_30B_A3B; break;
11        case 94: type = LLM_TYPE_235B_A22B; break;
12        default: type = LLM_TYPE_UNKNOWN;
13    }
14}
15
16void llama_model_qwen3vlmoe::load_arch_tensors(llama_model_loader &) {
17    LLAMA_LOAD_LOCALS;
18
19    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
20
21    // output
22    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
23    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
24    // if output is NULL, init from the input tok embed
25    if (output == NULL) {
26        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
27    }
28
29    for (int i = 0; i < n_layer; ++i) {
30        auto & layer = layers[i];
31
32        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
33
34        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
35        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
36
37        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
38        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
39
40        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
41
42        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
43
44        if (n_expert == 0) {
45            throw std::runtime_error("n_expert must be > 0 for QWEN3MOE");
46        }
47        if (n_expert_used == 0) {
48            throw std::runtime_error("n_expert_used must be > 0 for QWEN3MOE");
49        }
50
51        // MoE branch
52        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
53
54        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
55        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
56        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
57    }
58}
59
60std::unique_ptr<llm_graph_context> llama_model_qwen3vlmoe::build_arch_graph(const llm_graph_params & params) const {
61    return std::make_unique<graph>(*this, params);
62}
63
64llama_model_qwen3vlmoe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
65    const size_t n_deepstack_layers = hparams.n_deepstack_layers;
66
67    const int64_t n_embd      = hparams.n_embd;
68    const int64_t n_embd_head = hparams.n_embd_head_v();
69
70    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
71    GGML_ASSERT(n_embd_head == n_rot);
72
73    ggml_tensor * cur;
74    ggml_tensor * inpL;
75
76    inpL = build_inp_embd(model.tok_embd);
77
78    int sections[4];
79    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
80
81    // inp_pos - contains the positions
82    ggml_tensor * inp_pos = build_inp_pos();
83
84    auto * inp_attn = build_attn_inp_kv();
85
86    ggml_tensor * inp_out_ids = build_inp_out_ids();
87
88    for (int il = 0; il < n_layer; ++il) {
89        ggml_tensor * inpSA = inpL;
90
91        // norm
92        cur = build_norm(inpL,
93                model.layers[il].attn_norm, NULL,
94                LLM_NORM_RMS, il);
95        cb(cur, "attn_norm", il);
96
97        // self_attention
98        {
99            // compute Q and K and RoPE them
100            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
101                    n_embd_head, n_head, n_head_kv, il);
102
103            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
104            cb(Qcur, "Qcur_normed", il);
105
106            Qcur = ggml_rope_multi(
107                    ctx0, Qcur, inp_pos, nullptr,
108                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
109                    ext_factor, attn_factor, beta_fast, beta_slow
110                    );
111
112            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
113            cb(Kcur, "Kcur_normed", il);
114
115            Kcur = ggml_rope_multi(
116                    ctx0, Kcur, inp_pos, nullptr,
117                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
118                    ext_factor, attn_factor, beta_fast, beta_slow
119                    );
120
121            cb(Qcur, "Qcur", il);
122            cb(Kcur, "Kcur", il);
123            cb(Vcur, "Vcur", il);
124
125            cur = build_attn(inp_attn,
126                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
127                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
128        }
129
130        if (il == n_layer - 1 && inp_out_ids) {
131            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
132            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
133        }
134
135        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
136        cb(ffn_inp, "ffn_inp", il);
137
138        // MoE branch
139        cur = build_norm(ffn_inp,
140                model.layers[il].ffn_norm, NULL,
141                LLM_NORM_RMS, il);
142        cb(cur, "ffn_norm", il);
143
144        ggml_tensor * moe_out =
145            build_moe_ffn(cur,
146                    model.layers[il].ffn_gate_inp,
147                    model.layers[il].ffn_up_exps,
148                    model.layers[il].ffn_gate_exps,
149                    model.layers[il].ffn_down_exps,
150                    nullptr,
151                    n_expert, n_expert_used,
152                    LLM_FFN_SILU, true,
153                    hparams.expert_weights_scale,
154                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
155                    il);
156        cb(moe_out, "ffn_moe_out", il);
157        cur = moe_out;
158
159        cur = ggml_add(ctx0, cur, ffn_inp);
160
161        cur = build_cvec(cur, il);
162        cb(cur, "l_out", il);
163
164        if (il < (int) n_deepstack_layers) {
165            ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
166            cur = ggml_add(ctx0, cur, ds);
167            cb(cur, "deepstack_out", il);
168        }
169
170        // input for next layer
171        inpL = cur;
172    }
173
174    cur = inpL;
175
176    cur = build_norm(cur,
177            model.output_norm, NULL,
178            LLM_NORM_RMS, -1);
179
180    cb(cur, "result_norm", -1);
181    res->t_embd = cur;
182
183    // lm_head
184    cur = build_lora_mm(model.output, cur, model.output_s);
185
186    cb(cur, "result_output", -1);
187    res->t_logits = cur;
188
189    ggml_build_forward_expand(gf, cur);
190}