char-fork/llama.cpp

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

Georgi Gerganovmetal : single-source fusion table + fusion debug rework (#28164)a2878d30d

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
7.1 KiB197 linesraw
1#include "models.h"
2
3void llama_model_qwen3vl::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(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
7
8    switch (hparams.n_layer()) {
9        case 28: type = LLM_TYPE_1_7B; break;
10        case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break;
11        case 64: type = LLM_TYPE_32B; break;
12        default: type = LLM_TYPE_UNKNOWN;
13    }
14}
15
16void llama_model_qwen3vl::load_arch_tensors(llama_model_loader &) {
17    LLAMA_LOAD_LOCALS;
18
19    int64_t n_vocab_out = n_vocab;
20    if (arch == LLM_ARCH_QWEN3TTS) {
21        // [TAG_LLAMA_N_VOCAB_OUT]
22        n_vocab_out = 3072;
23    }
24
25    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
26
27    // output
28    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
29    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab_out}, TENSOR_NOT_REQUIRED);
30    // if output is NULL, init from the input tok embed
31    if (output == NULL) {
32        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
33    }
34
35    // output rerank head
36    cls_out = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
37
38    for (int i = 0; i < n_layer; ++i) {
39        auto & layer = layers[i];
40
41        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
42
43        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
44        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
45
46        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
47        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
48
49        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
50        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
51        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
52        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
53    }
54}
55
56std::unique_ptr<llm_graph_context> llama_model_qwen3vl::build_arch_graph(const llm_graph_params & params) const {
57    return std::make_unique<graph>(*this, params);
58}
59
60llama_model_qwen3vl::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
61    const size_t n_deepstack_layers = hparams.n_deepstack_layers;
62
63    const int64_t n_embd      = hparams.n_embd;
64    const int64_t n_embd_head = hparams.n_embd_head_v();
65
66    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
67    GGML_ASSERT(n_embd_head == n_rot);
68
69    ggml_tensor * cur;
70    ggml_tensor * inpL;
71
72    inpL = build_inp_embd(model.tok_embd);
73
74    int sections[4];
75    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
76
77    // inp_pos - contains the positions
78    ggml_tensor * inp_pos = build_inp_pos();
79
80    auto * inp_attn = build_attn_inp_kv();
81
82    ggml_tensor * inp_out_ids = build_inp_out_ids();
83
84    for (int il = 0; il < n_layer; ++il) {
85        ggml_tensor * inpSA = inpL;
86
87        // norm
88        cur = build_norm(inpL,
89                model.layers[il].attn_norm, NULL,
90                LLM_NORM_RMS, il);
91        cb(cur, "attn_norm", il);
92
93        // self-attention
94        {
95            // compute Q and K and RoPE them
96            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
97                    n_embd_head, n_head, n_head_kv, il);
98
99            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
100            cb(Qcur, "Qcur_normed", il);
101
102            Qcur = ggml_rope_multi(
103                    ctx0, Qcur, inp_pos, nullptr,
104                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
105                    ext_factor, attn_factor, beta_fast, beta_slow
106                    );
107
108            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
109            cb(Kcur, "Kcur_normed", il);
110
111            Kcur = ggml_rope_multi(
112                    ctx0, Kcur, inp_pos, nullptr,
113                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
114                    ext_factor, attn_factor, beta_fast, beta_slow
115                    );
116
117            cb(Qcur, "Qcur", il);
118            cb(Kcur, "Kcur", il);
119            cb(Vcur, "Vcur", il);
120
121            cur = build_attn(inp_attn,
122                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
123                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
124        }
125
126        if (il == n_layer - 1 && inp_out_ids) {
127            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
128            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
129        }
130
131        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
132        cb(ffn_inp, "ffn_inp", il);
133
134        // feed-forward network
135        cur = build_norm(ffn_inp,
136                model.layers[il].ffn_norm, NULL,
137                LLM_NORM_RMS, il);
138        cb(cur, "ffn_norm", il);
139
140        cur = build_ffn(cur,
141                model.layers[il].ffn_up,   NULL, NULL,
142                model.layers[il].ffn_gate, NULL, NULL,
143                model.layers[il].ffn_down, NULL, NULL,
144                NULL,
145                LLM_FFN_SILU, LLM_FFN_PAR, il);
146        cb(cur, "ffn_out", il);
147
148        cur = ggml_add(ctx0, cur, ffn_inp);
149
150        cur = build_cvec(cur, il);
151        cb(cur, "l_out", il);
152
153        if (il < (int) n_deepstack_layers) {
154            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));
155            cur = ggml_add(ctx0, cur, ds);
156            cb(cur, "deepstack_out", il);
157        }
158
159        // input for next layer
160        inpL = cur;
161    }
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    int64_t n_vocab_in  = model.tok_embd->ne[1];
176    int64_t n_vocab_out = model.output->ne[1];
177    if (n_vocab_in > n_vocab_out) {
178        // case: Qwen3TTS model with codec_head as output
179        GGML_ASSERT(model.output_norm);
180        int64_t pad = n_vocab_in - n_vocab_out;
181
182        // using this trick to get a scalar -inf tensor to pad the output
183        ggml_tensor * neg_inf = ggml_scale_bias(ctx0,
184                ggml_view_1d(ctx0, model.output_norm, 1, 0),
185                0.0f, -INFINITY);
186        neg_inf = ggml_repeat_4d(ctx0, neg_inf, pad, cur->ne[1], 1, 1);
187        cur = ggml_concat(ctx0, neg_inf, cur, 0); // [padded .. n_vocab_out, n_stream]
188
189    } else if (n_vocab_in < n_vocab_out) {
190        GGML_ABORT("invalid case");
191    }
192
193    cb(cur, "result_output", -1);
194    res->t_logits = cur;
195
196    ggml_build_forward_expand(gf, cur);
197}