char-fork/llama.cpp

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

Yaniss Amazouzmodel: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (#25444)c61b98b87

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
12.3 KiB319 linesraw
1#include "models.h"
2
3#include <sstream>
4
5void llama_model_granite_swa::load_arch_hparams(llama_model_loader & ml) {
6    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
7    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
8    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, false);
9    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, false);
10    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, false);
11
12    // MoE expert configuration
13    ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert, false);
14    ml.get_key_or_arr(LLM_KV_EXPERT_USED_COUNT,    hparams.n_expert_used_arr, hparams.n_layer_all, false);
15
16     // iSWA configuration
17    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);
18    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
19    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
20
21    // Granite4 Vision uses array deepstack_mapping
22    ml.get_arr(LLM_KV_DEEPSTACK_MAPPING, hparams.deepstack_mapping_arr, false);
23
24    // Count the unique deepstack input indices
25    std::unordered_set<uint32_t> unique_deepstack_idxs;
26    for (const auto val : hparams.deepstack_mapping_arr) {
27        if (val >= 0) {
28            unique_deepstack_idxs.insert(val);
29        }
30    }
31    hparams.n_deepstack_layers = unique_deepstack_idxs.size();
32
33    // Ensure all values are valid (avoid overflow attacks)
34    for (const auto val : unique_deepstack_idxs) {
35        if (val > hparams.n_deepstack_layers) {
36            std::stringstream ss;
37            ss << "Invalid deepstack index: " << val << " > " << hparams.n_deepstack_layers;
38            throw std::runtime_error(ss.str());
39        }
40    }
41
42    // Per-layer RoPE pattern (optional)
43    ml.get_arr(LLM_KV_ATTENTION_ROPE_PATTERN, hparams.rope_pattern, false);
44
45    switch (hparams.n_layer()) {
46        case 32: type = LLM_TYPE_3B; break;
47        case 40: type = LLM_TYPE_3B; break;
48        // Add additional layer/vocab/etc checks here for other model sizes
49        default: type = LLM_TYPE_UNKNOWN;
50    }
51
52    // For Granite MoE Shared
53    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
54}
55
56void llama_model_granite_swa::load_arch_tensors(llama_model_loader &) {
57    LLAMA_LOAD_LOCALS;
58
59    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
60
61    // output
62    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
63    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
64
65    // if output is NULL, init from the input tok embed
66    if (output == NULL) {
67        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
68    }
69
70    for (int i = 0; i < n_layer; ++i) {
71        auto & layer = layers[i];
72
73        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
74
75        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
76        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
77
78        // optional bias tensors
79        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
80
81        // Per-layer attention sinks for iSWA
82        layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, 0);
83
84        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
85
86        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
87            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
88            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
89        }
90        else {
91            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
92        }
93
94        if (n_expert == 0) {
95            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
96            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
97            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
98
99            // optional MLP bias
100            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
101            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
102            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
103        } else {
104            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
105            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
106            create_tensor_gate_up_exps(layer, i, n_embd, n_ff, n_expert, 0);
107
108            // For Granite MoE Shared - gate+up kept fused in ffn_up_shexp (see LLM_FFN_SWIGLU below)
109            if (hparams.n_ff_shexp > 0) {
110                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, 2*hparams.n_ff_shexp}, 0);
111                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
112            }
113        }
114    }
115}
116
117std::unique_ptr<llm_graph_context> llama_model_granite_swa::build_arch_graph(const llm_graph_params & params) const {
118    return std::make_unique<graph>(*this, params);
119}
120
121llama_model_granite_swa::graph::graph(
122    const llama_model & model,
123    const llm_graph_params & params)
124    : llm_graph_context(params) {
125
126    const int64_t n_embd_head = hparams.n_embd_head_v();
127
128    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
129    GGML_ASSERT(n_embd_head == n_rot);
130
131    ggml_tensor * cur;
132    ggml_tensor * inpL;
133
134    inpL = build_inp_embd(model.tok_embd);
135
136    // inp_pos - built only if rope enabled
137    ggml_tensor * inp_pos = build_inp_pos();
138    auto * inp_attn = build_attn_inp_kv_iswa();
139
140    ggml_tensor * inp_out_ids = build_inp_out_ids();
141
142    for (int il = 0; il < n_layer; ++il) {
143
144        // Granite Vision 4.1 deepstack: inject the projector stream that
145        // targets decoder layer `il` before the decoder runs.
146        // NOTE: skip the first deepstack layer since that's inpL
147        const auto & deepstack_emb_idx = hparams.deepstack_mapping_arr[il];
148        if (il > 0 && deepstack_emb_idx >= 0) {
149            ggml_tensor * ds = ggml_view_2d(ctx0,
150                res->t_inp_embd, n_embd, n_tokens,
151                res->t_inp_embd->nb[1],
152                deepstack_emb_idx * n_embd * sizeof(float));
153            inpL = ggml_add(ctx0, inpL, ds);
154            cb(inpL, "deepstack_in", il);
155        }
156
157        ggml_tensor * inpSA = inpL;
158
159        // norm
160        cur = build_norm(inpL,
161                model.layers[il].attn_norm, NULL,
162                LLM_NORM_RMS, il);
163        cb(cur, "attn_norm", il);
164
165        // self-attention
166        cur = build_attention_layer(
167            cur, inp_pos, inp_attn,
168            model, n_embd_head, il);
169
170        if (il == n_layer - 1 && inp_out_ids) {
171            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
172            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
173        }
174        // ffn
175        cur = build_layer_ffn(cur, inpSA, model, il);
176
177        // input for next layer
178        inpL = cur;
179    }
180    cur = inpL;
181
182    cur = build_norm(cur,
183            model.output_norm, NULL,
184            LLM_NORM_RMS, -1);
185
186    cb(cur, "result_norm", -1);
187    res->t_embd = cur;
188
189    // lm_head
190    cur = build_lora_mm(model.output, cur, model.output_s);
191
192    // For Granite architectures - scale logits
193    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
194    cb(cur, "result_output", -1);
195    res->t_logits = cur;
196
197    ggml_build_forward_expand(gf, cur);
198}
199
200ggml_tensor * llama_model_granite_swa::graph::build_attention_layer(
201          ggml_tensor                  * cur,
202          ggml_tensor                  * inp_pos,
203          llm_graph_input_attn_kv_iswa * inp_attn,
204    const llama_model                  & model,
205    const int64_t                        n_embd_head,
206    const int                            il) {
207
208    auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
209            n_embd_head, hparams.n_head(il), hparams.n_head_kv(il), il);
210
211    const bool use_rope = hparams.has_rope(il);
212    if (use_rope) {
213        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
214        Qcur = ggml_rope_ext(
215                ctx0, Qcur, inp_pos, rope_factors,
216                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
217                ext_factor, attn_factor, beta_fast, beta_slow
218                );
219
220        Kcur = ggml_rope_ext(
221                ctx0, Kcur, inp_pos, rope_factors,
222                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
223                ext_factor, attn_factor, beta_fast, beta_slow
224                );
225    }
226
227    cb(Qcur, "Qcur", il);
228    cb(Kcur, "Kcur", il);
229    cb(Vcur, "Vcur", il);
230
231    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
232
233    // Pass layer.attn_sinks to build_attn for sink-based attention modulation
234    cur = build_attn(inp_attn,
235            model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
236            Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, kq_scale, il);
237    cb(cur, "attn_out", il);
238    return cur;
239}
240
241ggml_tensor * llama_model_granite_swa::graph::build_layer_ffn(
242          ggml_tensor * cur,
243          ggml_tensor * inpSA,
244    const llama_model & model,
245    const int           il) {
246
247    // For Granite architectures - scale residual
248    if (hparams.f_residual_scale) {
249        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
250    }
251    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
252    cb(ffn_inp, "ffn_inp", il);
253
254    // feed-forward network (non-MoE)
255    if (model.layers[il].ffn_gate_inp == nullptr) {
256
257        cur = build_norm(ffn_inp,
258                model.layers[il].ffn_norm, NULL,
259                LLM_NORM_RMS, il);
260        cb(cur, "ffn_norm", il);
261
262        cur = build_ffn(cur,
263                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
264                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
265                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
266                NULL,
267                LLM_FFN_SILU, LLM_FFN_PAR, il);
268        cb(cur, "ffn_out", il);
269
270    } else {
271        // MoE branch
272        cur = build_norm(ffn_inp,
273                model.layers[il].ffn_norm, NULL,
274                LLM_NORM_RMS, il);
275        cb(cur, "ffn_norm", il);
276
277        ggml_tensor * moe_out = build_moe_ffn(cur,
278                model.layers[il].ffn_gate_inp,
279                model.layers[il].ffn_up_exps,
280                model.layers[il].ffn_gate_exps,
281                model.layers[il].ffn_down_exps,
282                nullptr,
283                n_expert, n_expert_used,
284                LLM_FFN_SILU, true,
285                hparams.expert_weights_scale,
286                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
287                il,
288                nullptr, model.layers[il].ffn_gate_up_exps);
289        cb(moe_out, "ffn_moe_out", il);
290
291        // For Granite MoE Shared - gate+up kept fused in ffn_up_shexp
292        if (hparams.n_ff_shexp > 0) {
293            ggml_tensor * ffn_shexp = build_ffn(cur,
294                model.layers[il].ffn_up_shexp,   NULL, NULL,
295                NULL,                            NULL, NULL,
296                model.layers[il].ffn_down_shexp, NULL, NULL,
297                NULL,
298                LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
299            cb(ffn_shexp, "ffn_shexp", il);
300
301            cur = ggml_add(ctx0, moe_out, ffn_shexp);
302            cb(cur, "ffn_out", il);
303        } else {
304            cur = moe_out;
305        }
306    }
307
308    // For Granite architectures - scale residual
309    if (hparams.f_residual_scale) {
310        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
311    }
312    cur = ggml_add(ctx0, cur, ffn_inp);
313    cb(cur, "ffn_out", il);
314
315    cur = build_cvec(cur, il);
316    cb(cur, "l_out", il);
317
318    return cur;
319}