char-fork/llama.cpp

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

Xuan-Son Nguyenmtmd: support pocket-tts (#26871)6e62ba538

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
5.1 KiB146 linesraw
1#include "models.h"
2
3// backbone of the pocket-tts CALM pipeline: the "text" side of a flow language model.
4// it has no lm_head, the audio latents are produced by the flow net inside the mmproj
5
6void llama_model_pockettts::load_arch_hparams(llama_model_loader & ml) {
7    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
8
9    switch (hparams.n_layer()) {
10        case 6:  type = LLM_TYPE_109M; break;
11        case 24: type = LLM_TYPE_335M; break;
12        default: type = LLM_TYPE_UNKNOWN;
13    }
14}
15
16void llama_model_pockettts::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_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
22    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
23    // no output head, the logits are unused; reuse the embedding table so a sampler can still run
24    output        = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
25
26    for (int i = 0; i < n_layer; ++i) {
27        auto & layer = layers[i];
28
29        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
30        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
31
32        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, TENSOR_NOT_REQUIRED);
33
34        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
35
36        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
37        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
38
39        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
40        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
41    }
42}
43
44std::unique_ptr<llm_graph_context> llama_model_pockettts::build_arch_graph(const llm_graph_params & params) const {
45    return std::make_unique<graph>(*this, params);
46}
47
48llama_model_pockettts::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
49    const int64_t n_embd_head = hparams.n_embd_head_v();
50
51    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
52    GGML_ASSERT(n_embd_head == n_rot);
53
54    ggml_tensor * cur;
55    ggml_tensor * inpL;
56
57    inpL = build_inp_embd(model.tok_embd);
58
59    ggml_tensor * inp_pos = build_inp_pos();
60
61    auto * inp_attn = build_attn_inp_kv();
62
63    ggml_tensor * inp_out_ids = build_inp_out_ids();
64
65    for (int il = 0; il < n_layer; ++il) {
66        cur = build_norm(inpL,
67                model.layers[il].attn_norm,
68                model.layers[il].attn_norm_b,
69                LLM_NORM, il);
70        cb(cur, "attn_norm", il);
71
72        // self-attention
73        {
74            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
75                    n_embd_head, n_head, n_head_kv, il);
76
77            Qcur = ggml_rope_ext(
78                    ctx0, Qcur, inp_pos, nullptr,
79                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
80                    ext_factor, attn_factor, beta_fast, beta_slow
81                    );
82
83            Kcur = ggml_rope_ext(
84                    ctx0, Kcur, inp_pos, nullptr,
85                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
86                    ext_factor, attn_factor, beta_fast, beta_slow
87                    );
88
89            cb(Qcur, "Qcur", il);
90            cb(Kcur, "Kcur", il);
91            cb(Vcur, "Vcur", il);
92
93            cur = build_attn(inp_attn,
94                    model.layers[il].wo, NULL, model.layers[il].wo_s,
95                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
96        }
97
98        if (il == n_layer - 1 && inp_out_ids) {
99            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
100            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
101        }
102
103        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
104        cb(ffn_inp, "ffn_inp", il);
105
106        // FF
107        {
108            cur = build_norm(ffn_inp,
109                    model.layers[il].ffn_norm,
110                    model.layers[il].ffn_norm_b,
111                    LLM_NORM, il);
112            cb(cur, "ffn_norm", il);
113
114            cur = build_ffn(cur,
115                    model.layers[il].ffn_up,   NULL, NULL,
116                    NULL,                      NULL, NULL,
117                    model.layers[il].ffn_down, NULL, NULL,
118                    NULL,
119                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
120            cb(cur, "ffn_out", il);
121        }
122
123        cur = ggml_add(ctx0, cur, ffn_inp);
124
125        cur = build_cvec(cur, il);
126        cb(cur, "l_out", il);
127
128        // input for next layer
129        inpL = cur;
130    }
131
132    cur = build_norm(inpL,
133            model.output_norm,
134            model.output_norm_b,
135            LLM_NORM, -1);
136
137    cb(cur, "result_norm", -1);
138    res->t_embd = cur;
139
140    cur = build_lora_mm(model.output, cur, model.output_s);
141
142    cb(cur, "result_output", -1);
143    res->t_logits = cur;
144
145    ggml_build_forward_expand(gf, cur);
146}