char-fork/llama.cpp

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

Georgi Gerganovhparams : refactor `hparams.n_layer` (#24060)7acb4e8cd

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
5.4 KiB154 linesraw
1#include "models.h"
2
3void llama_model_qwen2::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    switch (hparams.n_layer()) {
7        case 24: type = hparams.n_embd == 1024 ? LLM_TYPE_0_5B : LLM_TYPE_1B; break;
8        case 28: type = hparams.n_embd == 1536 ? LLM_TYPE_1_5B : LLM_TYPE_7B; break;
9        case 32: type = LLM_TYPE_7B; break;
10        case 36: type = LLM_TYPE_3B; break;
11        case 40: type = hparams.n_head() == 20 ? LLM_TYPE_4B : LLM_TYPE_13B; break;
12        case 48: type = LLM_TYPE_14B; break;
13        case 64: type = LLM_TYPE_32B; break;
14        case 80: type = LLM_TYPE_70B; break;
15        default: type = LLM_TYPE_UNKNOWN;
16    }
17}
18
19void llama_model_qwen2::load_arch_tensors(llama_model_loader &) {
20    LLAMA_LOAD_LOCALS;
21
22    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
23
24    // output
25    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
26    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
27    output_b    = create_tensor(tn(LLM_TENSOR_OUTPUT,      "bias"),   {n_vocab}, TENSOR_NOT_REQUIRED);
28    // if output is NULL, init from the input tok embed
29    if (output == NULL) {
30        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
31    }
32
33    for (int i = 0; i < n_layer; ++i) {
34        auto & layer = layers[i];
35
36        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
37
38        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
39        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
40
41        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
42
43        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
44        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
45        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
46    }
47}
48
49std::unique_ptr<llm_graph_context> llama_model_qwen2::build_arch_graph(const llm_graph_params & params) const {
50    return std::make_unique<graph>(*this, params);
51}
52
53llama_model_qwen2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
54    const int64_t n_embd_head = hparams.n_embd_head_v();
55
56    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
57    GGML_ASSERT(n_embd_head == n_rot);
58
59    ggml_tensor * cur;
60    ggml_tensor * inpL;
61
62    inpL = build_inp_embd(model.tok_embd);
63
64    // inp_pos - contains the positions
65    ggml_tensor * inp_pos = build_inp_pos();
66
67    auto * inp_attn = build_attn_inp_kv();
68
69    ggml_tensor * inp_out_ids = build_inp_out_ids();
70
71    for (int il = 0; il < n_layer; ++il) {
72        ggml_tensor * inpSA = inpL;
73
74        // norm
75        cur = build_norm(inpL,
76                model.layers[il].attn_norm, NULL,
77                LLM_NORM_RMS, il);
78        cb(cur, "attn_norm", il);
79
80        // self-attention
81        {
82            // compute Q and K and RoPE them
83            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
84                    n_embd_head, n_head, n_head_kv, il);
85
86            Qcur = ggml_rope_ext(
87                    ctx0, Qcur, inp_pos, nullptr,
88                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
89                    ext_factor, attn_factor, beta_fast, beta_slow
90                    );
91
92            Kcur = ggml_rope_ext(
93                    ctx0, Kcur, inp_pos, nullptr,
94                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
95                    ext_factor, attn_factor, beta_fast, beta_slow
96                    );
97
98            cb(Qcur, "Qcur", il);
99            cb(Kcur, "Kcur", il);
100            cb(Vcur, "Vcur", il);
101
102            cur = build_attn(inp_attn,
103                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
104                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
105        }
106        if (il == n_layer - 1 && inp_out_ids) {
107            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
108            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
109        }
110        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
111        cb(ffn_inp, "ffn_inp", il);
112
113        // feed-forward network
114        cur = build_norm(ffn_inp,
115                model.layers[il].ffn_norm, NULL,
116                LLM_NORM_RMS, il);
117        cb(cur, "ffn_norm", il);
118
119        cur = build_ffn(cur,
120                model.layers[il].ffn_up,   NULL, NULL,
121                model.layers[il].ffn_gate, NULL, NULL,
122                model.layers[il].ffn_down, NULL, NULL,
123                NULL,
124                LLM_FFN_SILU, LLM_FFN_PAR, il);
125        cb(cur, "ffn_out", il);
126
127        cur = ggml_add(ctx0, cur, ffn_inp);
128
129        cur = build_cvec(cur, il);
130        cb(cur, "l_out", il);
131
132        // input for next layer
133        inpL = cur;
134    }
135    cur = inpL;
136
137    cur = build_norm(cur,
138            model.output_norm, NULL,
139            LLM_NORM_RMS, -1);
140
141    cb(cur, "result_norm", -1);
142    res->t_embd = cur;
143
144    // lm_head
145    cur = build_lora_mm(model.output, cur, model.output_s);
146
147    if (model.output_b != nullptr) {
148        cur = ggml_add(ctx0, cur, model.output_b);
149    }
150    cb(cur, "result_output", -1);
151    res->t_logits = cur;
152
153    ggml_build_forward_expand(gf, cur);
154}