char-fork/llama.cpp

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

Sigbjørn Skjæretmodels : fix mimo2 swa pattern load (#28865)1aca1f9fc

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
16.6 KiB396 linesraw
1#include "models.h"
2
3void llama_model_mimo2::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
7
8    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
9    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,   hparams.n_swa);
10    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,         hparams.rope_freq_base_train_swa, false);
11
12    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);
13
14    float value_scale = 0.0f;
15    if (ml.get_key(LLM_KV_ATTENTION_VALUE_SCALE, value_scale, false) && value_scale != 1.0f) {
16        hparams.f_attn_value_scale = value_scale;
17    }
18
19    switch (hparams.n_layer()) {
20        case 48: type = LLM_TYPE_310B_A15B; break;
21        default: type = LLM_TYPE_UNKNOWN;
22    }
23}
24
25void llama_model_mimo2::load_arch_tensors(llama_model_loader & ml) {
26    LLAMA_LOAD_LOCALS;
27
28    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
29    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
30    int mtp_flags         = trunk_only ? TENSOR_NOT_REQUIRED : 0;
31
32    if (!ml.load_mtp) {
33        mtp_flags |= TENSOR_SKIP;
34    }
35
36    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
37
38    // output
39    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
40    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
41
42    for (int i = 0; i < n_layer_all; ++i) {
43        auto & layer = layers[i];
44        uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);
45        uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);
46        uint32_t n_head = hparams.n_head(i);
47
48        const bool is_nextn = i >= n_layer;
49        const int  flags    = is_nextn ? mtp_flags : 0;
50
51        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);
52        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_v * n_head, n_embd }, flags);
53
54        layer.attn_norm  = create_tensor(tn(LLM_TENSOR_ATTN_NORM,  "weight", i), {n_embd}, flags);
55        layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, TENSOR_NOT_REQUIRED | flags);
56
57        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
58
59        // non-MoE branch
60        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED | flags);
61        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, TENSOR_NOT_REQUIRED | flags);
62        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED | flags);
63
64        // MoE branch
65        int64_t n_ff_exp = hparams.n_ff_exp();
66        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED | flags);
67        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED | flags);
68        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, TENSOR_NOT_REQUIRED | flags);
69        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED | flags);
70        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);
71
72        if (is_nextn) {
73            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), {2 * n_embd, n_embd}, flags);
74            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), {n_embd}, flags);
75            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), {n_embd}, flags);
76            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED | flags);
77            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED | flags);
78            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED | flags);
79            layer.layer_out_norm         = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM,         "weight", i), {n_embd}, TENSOR_NOT_REQUIRED | flags);
80        }
81    }
82}
83
84std::unique_ptr<llm_graph_context> llama_model_mimo2::build_arch_graph(const llm_graph_params & params) const {
85    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
86        return std::make_unique<graph_mtp>(*this, params);
87    }
88    return std::make_unique<graph>(*this, params);
89}
90
91llama_model_mimo2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
92    ggml_tensor * cur;
93    ggml_tensor * inpL;
94
95    inpL = build_inp_embd(model.tok_embd);
96
97    ggml_tensor * inp_pos = build_inp_pos();
98    auto * inp_attn = build_attn_inp_kv_iswa();
99    ggml_tensor * inp_out_ids = build_inp_out_ids();
100
101    const float v_scale = hparams.f_attn_value_scale;
102    const bool emit_h_nextn = cparams.embeddings_nextn;
103    const bool crop_last_layer = inp_out_ids && (!emit_h_nextn || cparams.embeddings_nextn_masked);
104
105    for (int il = 0; il < n_layer; ++il) {
106        ggml_tensor * inpSA = inpL;
107
108        uint32_t n_head_l    = hparams.n_head(il);
109        uint32_t n_head_kv_l = hparams.n_head_kv(il);
110        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
111        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
112
113        cur = inpL;
114
115        // self_attention
116        {
117            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
118            cb(cur, "attn_norm", il);
119
120            ggml_tensor * Qcur;
121            ggml_tensor * Kcur;
122            ggml_tensor * Vcur;
123
124            if (model.layers[il].wqkv) {
125                // Fused qkv_proj - Q/K share head_dim_k, V uses head_dim_v
126                ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
127                cb(qkv, "wqkv", il);
128
129                const size_t row_k    = ggml_row_size(qkv->type, n_embd_head_k);
130                const size_t row_v    = ggml_row_size(qkv->type, n_embd_head_v);
131                const size_t row_full = qkv->nb[1];
132                const size_t k_off    = row_k * n_head_l;
133                const size_t v_off    = k_off + row_k * n_head_kv_l;
134
135                Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_l,    n_tokens, row_k, row_full, 0);
136                Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv_l, n_tokens, row_k, row_full, k_off);
137                Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv_l, n_tokens, row_v, row_full, v_off);
138            } else {
139                // Split path
140                Qcur = build_lora_mm(model.layers[il].wq, cur);
141                cb(Qcur, "Qcur", il);
142
143                Kcur = build_lora_mm(model.layers[il].wk, cur);
144                cb(Kcur, "Kcur", il);
145
146                Vcur = build_lora_mm(model.layers[il].wv, cur);
147                cb(Vcur, "Vcur", il);
148
149                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l,    n_tokens);
150                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
151                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
152            }
153
154            Qcur = ggml_rope_ext(
155                ctx0, Qcur, inp_pos, nullptr,
156                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
157                ext_factor, attn_factor, beta_fast, beta_slow
158                );
159
160            Kcur = ggml_rope_ext(
161                ctx0, Kcur, inp_pos, nullptr,
162                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
163                ext_factor, attn_factor, beta_fast, beta_slow
164                );
165
166            cb(Qcur, "Qcur", il);
167            cb(Kcur, "Kcur", il);
168            cb(Vcur, "Vcur", il);
169
170            ggml_tensor * sinks = model.layers[il].attn_sinks;
171
172            cur = build_attn(inp_attn,
173                    model.layers[il].wo, NULL, model.layers[il].wo_s,
174                    Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), il);
175            cb(cur, "attn_out", il);
176
177            if (v_scale) {
178                cur = ggml_scale(ctx0, cur, v_scale);
179                cb(cur, "attn_out_scaled", il);
180            }
181        }
182
183        if (il == n_layer - 1 && crop_last_layer) {
184            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
185            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
186        }
187
188        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
189        cb(ffn_inp, "ffn_inp", il);
190
191        cur = build_norm(ffn_inp,
192                model.layers[il].ffn_norm, NULL,
193                LLM_NORM_RMS, il);
194        cb(cur, "ffn_norm", il);
195
196        // feed-forward network
197        if (model.layers[il].ffn_gate_inp == nullptr) {
198            // dense branch
199            cur = build_ffn(cur,
200                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
201                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
202                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
203                    NULL,
204                    LLM_FFN_SILU, LLM_FFN_PAR, il);
205            cb(cur, "ffn_out", il);
206        } else {
207            // MoE branch
208            cur = build_moe_ffn(cur,
209                    model.layers[il].ffn_gate_inp,
210                    model.layers[il].ffn_up_exps,
211                    model.layers[il].ffn_gate_exps,
212                    model.layers[il].ffn_down_exps,
213                    model.layers[il].ffn_exp_probs_b,
214                    n_expert, n_expert_used,
215                    LLM_FFN_SILU, true,
216                    hparams.expert_weights_scale,
217                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
218                    il);
219            cb(cur, "ffn_moe_out", il);
220        }
221
222        cur = ggml_add(ctx0, cur, ffn_inp);
223
224        cur = build_cvec(cur, il);
225        cb(cur, "l_out", il);
226
227        // input for next layer
228        inpL = cur;
229    }
230
231    cur = inpL;
232
233    if (emit_h_nextn) {
234        cb(cur, "h_nextn", -1);
235        res->t_h_nextn = cur;
236
237        if (!cparams.embeddings_nextn_masked && inp_out_ids) {
238            cur = ggml_get_rows(ctx0, cur, inp_out_ids);
239        }
240    }
241
242    cur = build_norm(cur,
243            model.output_norm, NULL,
244            LLM_NORM_RMS, -1);
245
246    cb(cur, "result_norm", -1);
247    res->t_embd = cur;
248
249    // lm_head
250    cur = build_lora_mm(model.output, cur, model.output_s);
251
252    cb(cur, "result_output", -1);
253    res->t_logits = cur;
254
255    ggml_build_forward_expand(gf, cur);
256}
257
258// Mirrors MiMo's appended NextN block: normalize and fuse token and hidden inputs, run the decoder block,
259// expose its pre-head-norm state to the next draft step, then apply the shared output norm and LM head.
260// Converted checkpoints may store that shared norm as layer_out_norm, so it remains in the fallback chain.
261llama_model_mimo2::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
262    : llm_graph_context(params) {
263    GGML_ASSERT(hparams.n_layer_nextn > 0 && "MIMO2 MTP requires n_layer_nextn > 0");
264
265    const int il = hparams.n_layer() + cparams.nextn_layer_offset;
266    GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&
267                cparams.nextn_layer_offset < (int) hparams.n_layer_nextn &&
268                "nextn_layer_offset out of range [0, n_layer_nextn)");
269
270    const auto & layer = model.layers[il];
271    GGML_ASSERT(layer.nextn.eh_proj && "MIMO2 MTP block missing nextn.eh_proj");
272    GGML_ASSERT(layer.nextn.enorm   && "MIMO2 MTP block missing nextn.enorm");
273    GGML_ASSERT(layer.nextn.hnorm   && "MIMO2 MTP block missing nextn.hnorm");
274    GGML_ASSERT(layer.wqkv          && "MIMO2 MTP requires fused attn_qkv");
275
276    const uint32_t n_head_l    = hparams.n_head(il);
277    const uint32_t n_head_kv_l = hparams.n_head_kv(il);
278
279    const float freq_base_l  = model.get_rope_freq_base(cparams, il);
280    const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
281    const float v_scale      = hparams.f_attn_value_scale;
282
283    auto inp = std::make_unique<llm_graph_input_embd>(hparams.n_embd);
284
285    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
286    ggml_set_input(inp->tokens);
287
288    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
289    ggml_set_input(inp->embd);
290    ggml_set_name(inp->embd, "mtp_h_input");
291
292    ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
293    ggml_tensor * h_input    = inp->embd;
294    ggml_tensor * tok_embd   = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
295    cb(tok_embd, "mtp_tok_embd", il);
296
297    res->add_input(std::move(inp));
298
299    ggml_tensor * inp_pos     = build_inp_pos();
300    ggml_tensor * inp_out_ids = build_inp_out_ids();
301    auto        * inp_attn    = build_attn_inp_kv_iswa();
302
303    ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
304    cb(h_norm, "mtp_hnorm", il);
305
306    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
307    cb(e_norm, "mtp_enorm", il);
308
309    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);
310    cb(concat, "mtp_concat", il);
311
312    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
313    cb(cur, "mtp_eh_proj", il);
314
315    ggml_tensor * inpSA = cur;
316
317    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
318    cb(cur, "mtp_attn_norm", il);
319
320    ggml_tensor * qkv = build_lora_mm(layer.wqkv, cur, layer.wqkv_s);
321    cb(qkv, "mtp_wqkv", il);
322
323    const size_t row_k    = ggml_row_size(qkv->type, n_embd_head_k);
324    const size_t row_v    = ggml_row_size(qkv->type, n_embd_head_v);
325    const size_t row_full = qkv->nb[1];
326    const size_t k_off    = row_k * n_head_l;
327    const size_t v_off    = k_off + row_k * n_head_kv_l;
328
329    ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_l,    n_tokens, row_k, row_full, 0);
330    ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv_l, n_tokens, row_k, row_full, k_off);
331    ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv_l, n_tokens, row_v, row_full, v_off);
332
333    Qcur = ggml_rope_ext(
334        ctx0, Qcur, inp_pos, nullptr,
335        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
336        ext_factor, attn_factor, beta_fast, beta_slow);
337
338    Kcur = ggml_rope_ext(
339        ctx0, Kcur, inp_pos, nullptr,
340        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
341        ext_factor, attn_factor, beta_fast, beta_slow);
342
343    cb(Qcur, "mtp_Qcur", il);
344    cb(Kcur, "mtp_Kcur", il);
345    cb(Vcur, "mtp_Vcur", il);
346
347    cur = build_attn(inp_attn,
348            layer.wo, nullptr, layer.wo_s,
349            Qcur, Kcur, Vcur, nullptr, layer.attn_sinks, nullptr,
350            1.0f / sqrtf(float(n_embd_head_k)), il);
351    cb(cur, "mtp_attn_out", il);
352
353    if (v_scale) {
354        cur = ggml_scale(ctx0, cur, v_scale);
355        cb(cur, "mtp_attn_out_scaled", il);
356    }
357
358    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
359    cb(ffn_inp, "mtp_ffn_inp", il);
360
361    cur = build_norm(ffn_inp, layer.ffn_norm, nullptr, LLM_NORM_RMS, il);
362    cb(cur, "mtp_ffn_norm", il);
363
364    GGML_ASSERT(layer.ffn_gate && layer.ffn_down && layer.ffn_up && "MIMO2 MTP requires dense FFN tensors");
365    cur = build_ffn(cur,
366            layer.ffn_up,   layer.ffn_up_b,   nullptr,
367            layer.ffn_gate, layer.ffn_gate_b, nullptr,
368            layer.ffn_down, layer.ffn_down_b, nullptr,
369            nullptr,
370            LLM_FFN_SILU, LLM_FFN_PAR, il);
371    cb(cur, "mtp_ffn_out", il);
372
373    cur = ggml_add(ctx0, cur, ffn_inp);
374    cb(cur, "mtp_post_ffn", il);
375
376    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
377
378    cb(cur, "h_nextn", -1);
379    res->t_h_nextn = cur;
380
381    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
382            ? layer.nextn.shared_head_norm
383            : (layer.layer_out_norm ? layer.layer_out_norm : model.output_norm);
384    GGML_ASSERT(head_norm_w && "MIMO2 MTP missing head norm fallback");
385    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
386    cb(cur, "mtp_shared_head_norm", -1);
387
388    ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;
389    ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s;
390    GGML_ASSERT(head_w && "MIMO2 MTP missing LM head fallback");
391    cur = build_lora_mm(head_w, cur, head_s);
392    cb(cur, "result_output", -1);
393
394    res->t_logits = cur;
395    ggml_build_forward_expand(gf, cur);
396}