char-fork/llama.cpp

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

ynankanimodel : NvFP4 quantized LM head support (#23046)42928bc14

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
4.9 KiB143 linesraw
1#include "models.h"
2
3void llama_model_qwen2vl::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
5}
6// fall through
7
8void llama_model_qwen2vl::load_arch_tensors(llama_model_loader &) {
9    LLAMA_LOAD_LOCALS;
10
11    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
12
13    // output
14    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
15    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
16    output_b    = create_tensor(tn(LLM_TENSOR_OUTPUT,      "bias"),   {n_vocab}, TENSOR_NOT_REQUIRED);
17    // if output is NULL, init from the input tok embed
18    if (output == NULL) {
19        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
20    }
21
22    for (int i = 0; i < n_layer; ++i) {
23        auto & layer = layers[i];
24
25        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
26
27        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
28        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
29
30        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
31
32        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
33        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
34        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
35    }
36}
37
38std::unique_ptr<llm_graph_context> llama_model_qwen2vl::build_arch_graph(const llm_graph_params & params) const {
39    return std::make_unique<graph>(*this, params);
40}
41
42llama_model_qwen2vl::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
43    const int64_t n_embd_head = hparams.n_embd_head_v();
44
45    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
46    GGML_ASSERT(n_embd_head == n_rot);
47
48    ggml_tensor * cur;
49    ggml_tensor * inpL;
50
51    inpL = build_inp_embd(model.tok_embd);
52
53    // inp_pos - contains the positions
54    ggml_tensor * inp_pos = build_inp_pos();
55
56    auto * inp_attn = build_attn_inp_kv();
57
58    int sections[4];
59    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
60
61    ggml_tensor * inp_out_ids = build_inp_out_ids();
62
63    for (int il = 0; il < n_layer; ++il) {
64        ggml_tensor * inpSA = inpL;
65
66        // norm
67        cur = build_norm(inpL,
68                model.layers[il].attn_norm, NULL,
69                LLM_NORM_RMS, il);
70        cb(cur, "attn_norm", il);
71
72        // self-attention
73        {
74            // compute Q and K and RoPE them
75            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
76                    n_embd_head, n_head, n_head_kv, il);
77
78            Qcur = ggml_rope_multi(
79                    ctx0, Qcur, inp_pos, nullptr,
80                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
81                    ext_factor, attn_factor, beta_fast, beta_slow
82                    );
83
84            Kcur = ggml_rope_multi(
85                    ctx0, Kcur, inp_pos, nullptr,
86                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
87                    ext_factor, attn_factor, beta_fast, beta_slow
88                    );
89
90            cb(Qcur, "Qcur", il);
91            cb(Kcur, "Kcur", il);
92            cb(Vcur, "Vcur", il);
93
94            cur = build_attn(inp_attn,
95                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
96                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
97        }
98        if (il == n_layer - 1 && inp_out_ids) {
99            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
100            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
101        }
102        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
103        cb(ffn_inp, "ffn_inp", il);
104
105        // feed-forward network
106        cur = build_norm(ffn_inp,
107                model.layers[il].ffn_norm, NULL,
108                LLM_NORM_RMS, il);
109        cb(cur, "ffn_norm", il);
110
111        cur = build_ffn(cur,
112                model.layers[il].ffn_up,   NULL, NULL,
113                model.layers[il].ffn_gate, NULL, NULL,
114                model.layers[il].ffn_down, NULL, NULL,
115                NULL,
116                LLM_FFN_SILU, LLM_FFN_PAR, il);
117        cb(cur, "ffn_out", il);
118
119        cur = ggml_add(ctx0, cur, ffn_inp);
120
121        cur = build_cvec(cur, il);
122        cb(cur, "l_out", il);
123
124        // input for next layer
125        inpL = cur;
126    }
127    cur = inpL;
128
129    cur = build_norm(cur,
130            model.output_norm, NULL,
131            LLM_NORM_RMS, -1);
132
133    cb(cur, "result_norm", -1);
134    res->t_embd = cur;
135
136    // lm_head
137    cur = build_lora_mm(model.output, cur, model.output_s);
138
139    cb(cur, "result_output", -1);
140    res->t_logits = cur;
141
142    ggml_build_forward_expand(gf, cur);
143}