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 →
5.3 KiB142 linesraw
1#include "models.h"
2
3void llama_model_phi2::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
5
6    switch (hparams.n_layer()) {
7        case 24: type = LLM_TYPE_1B; break;
8        case 32: type = LLM_TYPE_3B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_phi2::load_arch_tensors(llama_model_loader &) {
14    LLAMA_LOAD_LOCALS;
15
16    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
17
18    // output
19    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
20    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
21    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
22    output_b      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "bias"),   {n_vocab}, 0);
23
24    for (int i = 0; i < n_layer; ++i) {
25        auto & layer = layers[i];
26
27        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
28        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
29
30        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
31
32        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
33        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, 0);
34
35        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
36        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
37
38        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
39        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
40    }
41}
42
43std::unique_ptr<llm_graph_context> llama_model_phi2::build_arch_graph(const llm_graph_params & params) const {
44    return std::make_unique<graph>(*this, params);
45}
46
47llama_model_phi2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
48    const int64_t n_embd_head = hparams.n_embd_head_v();
49
50    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
51
52    ggml_tensor * cur;
53    ggml_tensor * attn_norm_output;
54    ggml_tensor * ffn_output;
55    ggml_tensor * inpL;
56
57    inpL = build_inp_embd(model.tok_embd);
58
59    // inp_pos - contains the positions
60    ggml_tensor * inp_pos = build_inp_pos();
61
62    auto * inp_attn = build_attn_inp_kv();
63
64    ggml_tensor * inp_out_ids = build_inp_out_ids();
65
66    for (int il = 0; il < n_layer; ++il) {
67        attn_norm_output = build_norm(inpL,
68                model.layers[il].attn_norm,
69                model.layers[il].attn_norm_b,
70                LLM_NORM, il);
71        cb(attn_norm_output, "attn_norm", il);
72
73        // self-attention
74        {
75            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], attn_norm_output,
76                    n_embd_head, n_head, n_head_kv, il);
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            // with phi2, we scale the Q to avoid precision issues
94            // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
95            Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
96
97            cur = build_attn(inp_attn,
98                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
99                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
100        }
101        if (il == n_layer - 1 && inp_out_ids) {
102            cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);
103            inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);
104            attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);
105        }
106        // FF
107        {
108            ffn_output = build_ffn(attn_norm_output,
109                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
110                    NULL,                      NULL,                        NULL,
111                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
112                    NULL,
113                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
114            cb(ffn_output, "ffn_out", il);
115        }
116        cur = ggml_add(ctx0, cur, ffn_output);
117        cur = ggml_add(ctx0, cur, inpL);
118
119        cur = build_cvec(cur, il);
120        cb(cur, "l_out", il);
121
122        // input for next layer
123        inpL = cur;
124    }
125    cur = build_norm(inpL,
126            model.output_norm,
127            model.output_norm_b,
128            LLM_NORM, -1);
129
130    cb(cur, "result_norm", -1);
131    res->t_embd = cur;
132
133    cur = build_lora_mm(model.output, cur, model.output_s);
134    cb(cur, "result_output_no_bias", -1);
135
136    cur = ggml_add(ctx0, cur, model.output_b);
137
138    cb(cur, "result_output", -1);
139    res->t_logits = cur;
140
141    ggml_build_forward_expand(gf, cur);
142}