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 →
6.4 KiB160 linesraw
1#include "models.h"
2
3void llama_model_refact::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 32: type = LLM_TYPE_1B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10
11    // TODO: become GGUF KV parameter
12    hparams.f_max_alibi_bias = 8.0f;
13}
14
15void llama_model_refact::load_arch_tensors(llama_model_loader &) {
16    LLAMA_LOAD_LOCALS;
17
18    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
19
20    // output
21    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
22    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
23
24    // if output is NULL, init from the input tok embed
25    if (output == NULL) {
26        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
27    }
28
29    for (int i = 0; i < n_layer; ++i) {
30        auto & layer = layers[i];
31
32        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
33
34        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
35        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
36
37        // optional bias tensors
38        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
39
40        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
41
42        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
43            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
44            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
45        }
46        else {
47            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
48        }
49
50        if (n_expert == 0) {
51            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
52            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
53            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
54
55            // optional MLP bias
56            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
57            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
58            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
59        } else {
60            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
61            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
62            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
63            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
64
65            // For Granite MoE Shared
66            if (hparams.n_ff_shexp > 0) {
67                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
68                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
69                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
70            }
71        }
72    }
73}
74
75std::unique_ptr<llm_graph_context> llama_model_refact::build_arch_graph(const llm_graph_params & params) const {
76    return std::make_unique<graph>(*this, params);
77}
78
79llama_model_refact::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
80    const int64_t n_embd_head = hparams.n_embd_head_v();
81
82    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
83
84    ggml_tensor * cur;
85    ggml_tensor * inpL;
86
87    inpL = build_inp_embd(model.tok_embd);
88
89    auto * inp_attn = build_attn_inp_kv();
90
91    ggml_tensor * inp_out_ids = build_inp_out_ids();
92
93    for (int il = 0; il < n_layer; ++il) {
94        ggml_tensor * inpSA = inpL;
95
96        cur = build_norm(inpL,
97                model.layers[il].attn_norm, NULL,
98                LLM_NORM_RMS, il);
99        cb(cur, "attn_norm", il);
100
101        // self-attention
102        {
103            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
104                    n_embd_head, n_head, n_head_kv, il);
105
106            cb(Qcur, "Qcur", il);
107            cb(Kcur, "Kcur", il);
108            cb(Vcur, "Vcur", il);
109
110            cur = build_attn(inp_attn,
111                    model.layers[il].wo, NULL, model.layers[il].wo_s,
112                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
113        }
114        if (il == n_layer - 1 && inp_out_ids) {
115            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
116            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
117        }
118        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
119        cb(ffn_inp, "ffn_inp", il);
120
121        // feed-forward network
122        {
123            cur = build_norm(ffn_inp,
124                    model.layers[il].ffn_norm, NULL,
125                    LLM_NORM_RMS, il);
126            cb(cur, "ffn_norm", il);
127
128            cur = build_ffn(cur,
129                    model.layers[il].ffn_up,   NULL, NULL,
130                    model.layers[il].ffn_gate, NULL, NULL,
131                    model.layers[il].ffn_down, NULL, NULL,
132                    NULL,
133                    LLM_FFN_SILU, LLM_FFN_PAR, il);
134            cb(cur, "ffn_out", il);
135        }
136        cur = ggml_add(ctx0, cur, ffn_inp);
137
138        cur = build_cvec(cur, il);
139        cb(cur, "l_out", il);
140
141        // input for next layer
142        inpL = cur;
143    }
144    cur = inpL;
145
146    cur = build_norm(cur,
147            model.output_norm, NULL,
148            LLM_NORM_RMS, -1);
149
150    cb(cur, "result_norm", -1);
151    res->t_embd = cur;
152
153    // lm_head
154    cur = build_lora_mm(model.output, cur, model.output_s);
155
156    cb(cur, "result_output", -1);
157    res->t_logits = cur;
158
159    ggml_build_forward_expand(gf, cur);
160}