char-fork/llama.cpp

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

Xuan-Son Nguyenmodel: use ggml_rope_set_offset() (#27382)873e5d8e3

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
8.4 KiB206 linesraw
1#include "models.h"
2
3void llama_model_plm::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
6
7    switch (hparams.n_layer()) {
8        case 32: type = LLM_TYPE_1_8B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_plm::load_arch_tensors(llama_model_loader &) {
14    LLAMA_LOAD_LOCALS;
15
16    const int64_t n_embd_head_qk_rope = hparams.n_rot();
17    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
18    const int64_t kv_lora_rank = hparams.n_lora_kv;
19
20    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
21
22    // output
23    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
24    // output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
25    output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
26
27    for (int i = 0; i < n_layer; ++i) {
28        auto & layer = layers[i];
29
30        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
31
32        layer.wq        = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
33        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0);
34        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
35        layer.wkv_b     = create_tensor(tn(LLM_TENSOR_ATTN_KV_B,     "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0);
36        layer.wo        = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd}, 0);
37
38        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
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_plm::build_arch_graph(const llm_graph_params & params) const {
45    return std::make_unique<graph>(*this, params);
46}
47
48llama_model_plm::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
49    const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k()));
50
51    const uint32_t n_embd_head_qk_rope = hparams.n_rot();
52    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
53
54    const uint32_t kv_lora_rank = hparams.n_lora_kv;
55
56    ggml_tensor * cur;
57    ggml_tensor * inpL;
58
59    // {n_embd, n_tokens}
60    inpL = build_inp_embd(model.tok_embd);
61
62    // inp_pos - contains the positions
63    ggml_tensor * inp_pos = build_inp_pos();
64
65    auto * inp_attn = build_attn_inp_kv();
66
67    ggml_tensor * inp_out_ids = build_inp_out_ids();
68
69    for (int il = 0; il < n_layer; ++il) {
70        ggml_tensor * inpSA = inpL;
71
72        // norm
73        cur = build_norm(inpL,
74                model.layers[il].attn_norm, NULL,
75                LLM_NORM_RMS, il);
76        cb(cur, "attn_norm", il);
77
78        // self_attention
79        {
80            ggml_tensor * q = NULL;
81            q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
82            cb(q, "q", il);
83
84            // {n_embd_head_k, n_head, n_tokens}, RoPE is applied to the trailing dims only
85            q = ggml_reshape_3d(ctx0, q, hparams.n_embd_head_k(), n_head, n_tokens);
86            cb(q, "q", il);
87
88            // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
89            ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
90            cb(kv_pe_compresseed, "kv_pe_compresseed", il);
91
92            // split into {kv_lora_rank, n_tokens}
93            ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
94                    kv_pe_compresseed->nb[1],
95                    0);
96            cb(kv_compressed, "kv_compressed", il);
97
98            // and {n_embd_head_qk_rope, n_tokens}
99            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
100                    kv_pe_compresseed->nb[1],
101                    kv_pe_compresseed->nb[1],
102                    ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
103            cb(k_pe, "k_pe", il);
104
105            kv_compressed = build_norm(kv_compressed,
106                    model.layers[il].attn_kv_a_norm, NULL,
107                    LLM_NORM_RMS, il);
108            cb(kv_compressed, "kv_compressed", il);
109
110            // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
111            ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
112            cb(kv, "kv", il);
113
114            // split into {n_head * n_embd_head_qk_nope, n_tokens}
115            ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
116                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v()),
117                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v())),
118                    0);
119            cb(k_nope, "k_nope", il);
120
121            // and {n_head * n_embd_head_v, n_tokens}
122            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v(), n_head, n_tokens,
123                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())),
124                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())*n_head),
125                    ggml_row_size(kv->type, (n_embd_head_qk_nope)));
126            cb(v_states, "v_states", il);
127
128            v_states = ggml_cont(ctx0, v_states);
129            cb(v_states, "v_states", il);
130
131            v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v() * n_head, n_tokens,
132                    ggml_row_size(kv->type, hparams.n_embd_head_v() * n_head),
133                    0);
134            cb(v_states, "v_states", il);
135
136            q = ggml_rope_ext(
137                    ctx0, q, inp_pos, nullptr,
138                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
139                    ext_factor, attn_factor, beta_fast, beta_slow
140                    );
141            q = ggml_rope_set_offset(q, n_embd_head_qk_nope);
142            cb(q, "q_rope", il);
143
144            // shared RoPE key
145            k_pe = ggml_rope_ext(
146                    ctx0, k_pe, inp_pos, nullptr,
147                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
148                    ext_factor, attn_factor, beta_fast, beta_slow
149                    );
150            cb(k_pe, "k_pe", il);
151
152            ggml_tensor * q_states = q;
153            cb(q_states, "q_states", il);
154
155            ggml_tensor * k_states = ggml_concat(ctx0, k_nope,
156                    ggml_repeat_4d(ctx0, k_pe, n_embd_head_qk_rope, n_head, n_tokens, 1), 0);
157            cb(k_states, "k_states", il);
158
159            cur = build_attn(inp_attn,
160                    model.layers[il].wo, NULL, model.layers[il].wo_s,
161                    q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
162        }
163        if (il == n_layer - 1 && inp_out_ids) {
164            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
165            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
166        }
167        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
168        cb(ffn_inp, "ffn_inp", il);
169
170        cur = build_norm(ffn_inp,
171                model.layers[il].ffn_norm, NULL,
172                LLM_NORM_RMS, il);
173        cb(cur, "ffn_norm", il);
174
175        cur = build_ffn(cur,
176                model.layers[il].ffn_up,   NULL, NULL,
177                NULL, NULL, NULL,
178                model.layers[il].ffn_down, NULL, NULL,
179                NULL,
180                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
181        cb(cur, "ffn_out", il);
182
183        cur = ggml_add(ctx0, cur, ffn_inp);
184
185        cur = build_cvec(cur, il);
186        cb(cur, "l_out", il);
187
188        // input for next layer
189        inpL = cur;
190    }
191    cur = inpL;
192
193    cur = build_norm(cur,
194            model.output_norm, NULL,
195            LLM_NORM_RMS, -1);
196
197    cb(cur, "result_norm", -1);
198    res->t_embd = cur;
199
200    cur = build_lora_mm(model.output, cur, model.output_s);
201
202    cb(cur, "result_output", -1);
203    res->t_logits = cur;
204
205    ggml_build_forward_expand(gf, cur);
206}