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 →
6.8 KiB170 linesraw
1#include "models.h"
2
3void llama_model_mpt::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);
5    ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV,      hparams.f_clamp_kqv, false);
6    ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias, false);
7
8    switch (hparams.n_layer()) {
9        case 32: type = LLM_TYPE_7B; break;
10        case 48: type = LLM_TYPE_30B; break;
11        default: type = LLM_TYPE_UNKNOWN;
12    }
13}
14
15void llama_model_mpt::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    pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, n_ctx_train}, TENSOR_NOT_REQUIRED);
20
21    // output
22    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
23    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, TENSOR_NOT_REQUIRED);
24
25    output        = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
26    if (!output) {
27        output    = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // needs to be on GPU
28    }
29
30    for (int i = 0; i < n_layer; ++i) {
31        auto & layer = layers[i];
32
33        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
34        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
35
36        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
37        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
38
39        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
40        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
41
42        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
43        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
44
45        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
46        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
47
48        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
49        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
50
51        // FIXME test-llama-archs crashes if q_norm is created
52        layer.attn_q_norm   = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
53        layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
54
55        layer.attn_k_norm   = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
56        layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
57
58        // AWQ ScaleActivation layer
59        layer.ffn_act = create_tensor(tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, TENSOR_NOT_REQUIRED);
60    }
61}
62
63std::unique_ptr<llm_graph_context> llama_model_mpt::build_arch_graph(const llm_graph_params & params) const {
64    return std::make_unique<graph>(*this, params);
65}
66
67llama_model_mpt::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
68    const int64_t n_embd_head = hparams.n_embd_head_v();
69
70    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
71
72    ggml_tensor * cur;
73    ggml_tensor * pos;
74    ggml_tensor * inpL;
75
76    inpL = build_inp_embd(model.tok_embd);
77
78    auto * inp_attn = build_attn_inp_kv();
79
80    if (model.pos_embd) {
81        // inp_pos - contains the positions
82        ggml_tensor * inp_pos = build_inp_pos();
83        pos                   = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
84        cb(pos, "pos_embd", -1);
85
86        inpL = ggml_add(ctx0, inpL, pos);
87        cb(inpL, "inpL", -1);
88    }
89
90    ggml_tensor * inp_out_ids = build_inp_out_ids();
91
92    for (int il = 0; il < n_layer; ++il) {
93        ggml_tensor * attn_norm;
94
95        attn_norm = build_norm(inpL, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, il);
96        cb(attn_norm, "attn_norm", il);
97
98        // self-attention
99        {
100            cur = attn_norm;
101
102            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
103                    n_embd_head, n_head, n_head_kv, il);
104
105            // Q/K Layernorm
106            if (model.layers[il].attn_q_norm) {
107                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);
108                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);
109
110                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);
111
112                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);
113
114                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
115                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
116            }
117
118            cb(Qcur, "Qcur", il);
119            cb(Kcur, "Kcur", il);
120            cb(Vcur, "Vcur", il);
121
122            cur = build_attn(inp_attn,
123                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
124                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
125        }
126
127        if (il == n_layer - 1 && inp_out_ids) {
128            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
129            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
130        }
131
132        // Add the input
133        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
134        cb(ffn_inp, "ffn_inp", il);
135
136        // feed forward
137        {
138            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, LLM_NORM, il);
139            cb(cur, "ffn_norm", il);
140            cur = build_ffn(cur,
141                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
142                NULL, NULL, NULL,
143                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
144                model.layers[il].ffn_act, LLM_FFN_GELU, LLM_FFN_SEQ, il);
145            cb(cur, "ffn_out", il);
146        }
147
148        cur = ggml_add(ctx0, cur, ffn_inp);
149
150        cur = build_cvec(cur, il);
151        cb(cur, "l_out", il);
152
153        // input for next layer
154        inpL = cur;
155    }
156
157    cur = inpL;
158
159    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
160
161    cb(cur, "result_norm", -1);
162    res->t_embd = cur;
163
164    cur = build_lora_mm(model.output, cur, model.output_s);
165
166    cb(cur, "result_output", -1);
167    res->t_logits = cur;
168
169    ggml_build_forward_expand(gf, cur);
170}