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 →
10.6 KiB252 linesraw
1#include "models.h"
2
3void llama_model_minicpm3::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_Q_LORA_RANK,       hparams.n_lora_q);
6    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,      hparams.n_lora_kv);
7
8    switch (hparams.n_layer()) {
9        case 62: type = LLM_TYPE_4B; break;
10        default: type = LLM_TYPE_UNKNOWN;
11    }
12}
13
14void llama_model_minicpm3::load_arch_tensors(llama_model_loader &) {
15    LLAMA_LOAD_LOCALS;
16
17    const int64_t n_embd_head_qk_rope = hparams.n_rot();
18    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
19
20    const int64_t q_lora_rank  = hparams.n_lora_q;
21    const int64_t kv_lora_rank = hparams.n_lora_kv;
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
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        layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
38
39        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
40
41        layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
42        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}, 0);
43
44        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);
45        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);
46        layer.wo        = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd}, 0);
47
48        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
49
50        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
51        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
52        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
53
54        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_embd_head_qk_rope/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
55        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head_qk_rope/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
56    }
57}
58
59std::unique_ptr<llm_graph_context> llama_model_minicpm3::build_arch_graph(const llm_graph_params & params) const {
60    return std::make_unique<graph>(*this, params);
61}
62
63llama_model_minicpm3::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
64    //TODO: if the model varies, these parameters need to be read from the model
65    const int64_t n_embd_base = 256;
66    const float scale_embd  = 12.0f;
67    const float scale_depth = 1.4f;
68    const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k()));
69
70    const uint32_t n_embd_head_qk_rope = hparams.n_rot();
71    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
72
73    const uint32_t kv_lora_rank = hparams.n_lora_kv;
74
75    ggml_tensor * cur;
76    ggml_tensor * inpL;
77
78    inpL = build_inp_embd(model.tok_embd);
79
80    // scale the input embeddings
81    inpL = ggml_scale(ctx0, inpL, scale_embd);
82    cb(inpL, "inp_scaled", -1);
83
84    // inp_pos - contains the positions
85    ggml_tensor * inp_pos = build_inp_pos();
86
87    auto * inp_attn = build_attn_inp_kv();
88
89    ggml_tensor * inp_out_ids = build_inp_out_ids();
90
91    for (int il = 0; il < n_layer; ++il) {
92        ggml_tensor * inpSA = inpL;
93
94        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
95
96        // norm
97        cur = build_norm(inpL,
98                model.layers[il].attn_norm, NULL,
99                LLM_NORM_RMS, il);
100        cb(cur, "attn_norm", il);
101
102        // self_attention
103        {
104            ggml_tensor * q = NULL;
105            // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
106            q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
107            cb(q, "q", il);
108
109            q = build_norm(q,
110                    model.layers[il].attn_q_a_norm, NULL,
111                    LLM_NORM_RMS, il);
112            cb(q, "q", il);
113
114            // {q_lora_rank, n_head * hparams.n_embd_head_k()} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k(), n_tokens}
115            q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
116            cb(q, "q", il);
117
118            // {n_embd_head_k, n_head, n_tokens}, RoPE is applied to the trailing dims only
119            q = ggml_reshape_3d(ctx0, q, hparams.n_embd_head_k(), n_head, n_tokens);
120            cb(q, "q", il);
121
122            // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
123            ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
124            cb(kv_pe_compresseed, "kv_pe_compresseed", il);
125
126            // split into {kv_lora_rank, n_tokens}
127            ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
128                    kv_pe_compresseed->nb[1],
129                    0);
130            cb(kv_compressed, "kv_compressed", il);
131
132            // and {n_embd_head_qk_rope, n_tokens}
133            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
134                    kv_pe_compresseed->nb[1],
135                    kv_pe_compresseed->nb[1],
136                    ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
137            cb(k_pe, "k_pe", il);
138
139            kv_compressed = build_norm(kv_compressed,
140                    model.layers[il].attn_kv_a_norm, NULL,
141                    LLM_NORM_RMS, il);
142            cb(kv_compressed, "kv_compressed", il);
143
144            // {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}
145            ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
146            cb(kv, "kv", il);
147
148            // split into {n_head * n_embd_head_qk_nope, n_tokens}
149            ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
150                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v()),
151                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v())),
152                    0);
153            cb(k_nope, "k_nope", il);
154
155            // and {n_head * n_embd_head_v, n_tokens}
156            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v(), n_head, n_tokens,
157                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())),
158                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())*n_head),
159                    ggml_row_size(kv->type, (n_embd_head_qk_nope)));
160            cb(v_states, "v_states", il);
161
162            v_states = ggml_cont(ctx0, v_states);
163            cb(v_states, "v_states", il);
164
165            q = ggml_rope_ext(
166                    ctx0, q, inp_pos, rope_factors,
167                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
168                    ext_factor, attn_factor, beta_fast, beta_slow
169                    );
170            q = ggml_rope_set_offset(q, n_embd_head_qk_nope);
171            cb(q, "q_rope", il);
172
173            // shared RoPE key
174            k_pe = ggml_rope_ext(
175                    ctx0, k_pe, inp_pos, rope_factors,
176                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
177                    ext_factor, attn_factor, beta_fast, beta_slow
178                    );
179            cb(k_pe, "k_pe", il);
180
181            ggml_tensor * q_states = q;
182            cb(q_states, "q_states", il);
183
184            ggml_tensor * k_states = ggml_concat(ctx0, k_nope,
185                    ggml_repeat_4d(ctx0, k_pe, n_embd_head_qk_rope, n_head, n_tokens, 1), 0);
186            cb(k_states, "k_states", il);
187
188            cur = build_attn(inp_attn,
189                    model.layers[il].wo, NULL, model.layers[il].wo_s,
190                    q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
191        }
192        if (il == n_layer - 1 && inp_out_ids) {
193            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
194            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
195        }
196        // scale_res - scale the hidden states for residual connection
197        const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct?
198        cur = ggml_scale(ctx0, cur, scale_res);
199        cb(cur, "hidden_scaled", il);
200
201        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
202        cb(ffn_inp, "ffn_inp", il);
203
204        // feed-forward network
205        {
206            cur = build_norm(ffn_inp,
207                    model.layers[il].ffn_norm, NULL,
208                    LLM_NORM_RMS, il);
209            cb(cur, "ffn_norm", il);
210
211            cur = build_ffn(cur,
212                    model.layers[il].ffn_up,   NULL, NULL,
213                    model.layers[il].ffn_gate, NULL, NULL,
214                    model.layers[il].ffn_down, NULL, NULL,
215                    NULL,
216                    LLM_FFN_SILU, LLM_FFN_PAR, il);
217            cb(cur, "ffn_out", il);
218        }
219        // scale the hidden states for residual connection
220        cur = ggml_scale(ctx0, cur, scale_res);
221        cb(cur, "hidden_scaled_ffn", il);
222
223        cur = ggml_add(ctx0, cur, ffn_inp);
224
225        cur = build_cvec(cur, il);
226        cb(cur, "l_out", il);
227
228        // input for next layer
229        inpL = cur;
230    }
231    cur = inpL;
232
233    cur = build_norm(cur,
234            model.output_norm, NULL,
235            LLM_NORM_RMS, -1);
236
237    cb(cur, "result_norm", -1);
238    res->t_embd = cur;
239
240    // lm_head scaling
241    const float scale_lmhead = float(n_embd_base)/float(n_embd);
242    cur = ggml_scale(ctx0, cur, scale_lmhead);
243    cb(cur, "lmhead_scaling", -1);
244
245    // lm_head
246    cur = build_lora_mm(model.output, cur, model.output_s);
247
248    cb(cur, "result_output", -1);
249    res->t_logits = cur;
250
251    ggml_build_forward_expand(gf, cur);
252}