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 →
3.4 KiB107 linesraw
1#include "models.h"
2
3std::unique_ptr<llm_graph_context> llama_model_paddleocr::build_arch_graph(const llm_graph_params & params) const {
4    return std::make_unique<graph>(*this, params);
5}
6
7llama_model_paddleocr::graph::graph(const llama_model & model, const llm_graph_params & params) :
8    llm_graph_context(params) {
9
10    // NOTE: same with qwen2vl.cpp, but bias tensors are optional
11
12    const int64_t n_embd_head = hparams.n_embd_head_v();
13
14    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
15    GGML_ASSERT(n_embd_head == n_rot);
16
17    ggml_tensor * cur;
18    ggml_tensor * inpL;
19
20    inpL = build_inp_embd(model.tok_embd);
21
22    int sections[4];
23    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
24
25    // inp_pos - contains the positions
26    ggml_tensor * inp_pos = build_inp_pos();
27
28    auto * inp_attn = build_attn_inp_kv();
29
30    ggml_tensor * inp_out_ids = build_inp_out_ids();
31
32    for (int il = 0; il < n_layer; ++il) {
33        ggml_tensor * inpSA = inpL;
34
35        // norm
36        {
37            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
38            cb(cur, "attn_norm", il);
39        }
40        // self-attention
41        {
42            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
43                    n_embd_head, n_head, n_head_kv, il);
44
45            Qcur = ggml_rope_multi(
46                    ctx0, Qcur, inp_pos, nullptr,
47                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
48                    ext_factor, attn_factor, beta_fast, beta_slow
49                    );
50
51            Kcur = ggml_rope_multi(
52                    ctx0, Kcur, inp_pos, nullptr,
53                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
54                    ext_factor, attn_factor, beta_fast, beta_slow
55                    );
56
57            cb(Qcur, "Qcur", il);
58            cb(Kcur, "Kcur", il);
59            cb(Vcur, "Vcur", il);
60
61            cur = build_attn(inp_attn,
62                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
63                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
64        }
65        if (il == n_layer - 1) {
66            // skip computing output for unused tokens
67            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
68            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
69        }
70        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
71        cb(ffn_inp, "ffn_inp", il);
72
73        // feed-forward network
74        {
75            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
76            cb(cur, "ffn_norm", il);
77
78            cur = build_ffn(cur,
79                    model.layers[il].ffn_up, NULL, NULL,
80                    model.layers[il].ffn_gate, NULL, NULL,
81                    model.layers[il].ffn_down, NULL, NULL,
82                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
83            cb(cur, "ffn_out", il);
84        }
85        cur = ggml_add(ctx0, cur, ffn_inp);
86
87        cur = build_cvec(cur, il);
88        cb(cur, "l_out", il);
89
90        // input for next layer
91        inpL = cur;
92    }
93    cur = inpL;
94
95    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
96
97    cb(cur, "result_norm", -1);
98    res->t_embd = cur;
99
100    // lm_head
101    cur = build_lora_mm(model.output, cur, model.output_s);
102
103    cb(cur, "result_output", -1);
104    res->t_logits = cur;
105
106    ggml_build_forward_expand(gf, cur);
107}