char-fork/llama.cpp

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

Aaron Teomodel: fix all granite family parameter counts (#28643)d7e86430a

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
12.3 KiB320 linesraw
1#include "models.h"
2
3#include <sstream>
4
5void llama_model_granite::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    // Granite4 Vision uses array deepstack_mapping
13    ml.get_arr(LLM_KV_DEEPSTACK_MAPPING, hparams.deepstack_mapping_arr, false);
14
15    // Count the unique deepstack input indices
16    std::unordered_set<uint32_t> unique_deepstack_idxs;
17    for (const auto val : hparams.deepstack_mapping_arr) {
18        if (val >= 0) {
19            unique_deepstack_idxs.insert(val);
20        }
21    }
22    hparams.n_deepstack_layers = unique_deepstack_idxs.size();
23
24    // Ensure all values are valid (avoid overflow attacks)
25    for (const auto val : unique_deepstack_idxs) {
26        if (val > hparams.n_deepstack_layers) {
27            std::stringstream ss;
28            ss << "Invalid deepstack index: " << val << " > " << hparams.n_deepstack_layers;
29            throw std::runtime_error(ss.str());
30        }
31    }
32
33    // Granite uses rope_finetuned as a switch for rope, so default to true
34    bool rope_finetuned = true;
35    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
36    hparams.rope_finetuned = rope_finetuned; // needed for round trip save
37    std::fill(hparams.rope_pattern.begin(), hparams.rope_pattern.end(), rope_finetuned);
38
39    switch (hparams.n_layer()) {
40        case 32: type = LLM_TYPE_3B; break;
41        case 40: {
42            switch (hparams.n_embd) {
43                case 2048: type = LLM_TYPE_2B; break;
44                case 2560: type = LLM_TYPE_3B; break;
45                case 4096: type = LLM_TYPE_8B; break;
46                default: type = LLM_TYPE_UNKNOWN;
47            }
48            break;
49        }
50        case 64: type = LLM_TYPE_30B; break;
51        // Add additional layer/vocab/etc checks here for other model sizes
52        default: type = LLM_TYPE_UNKNOWN;
53    }
54
55    // For Granite MoE Shared
56    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
57}
58
59void llama_model_granite::load_arch_tensors(llama_model_loader &) {
60    LLAMA_LOAD_LOCALS;
61
62    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
63
64    // output
65    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
66    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
67
68    // if output is NULL, init from the input tok embed
69    if (output == NULL) {
70        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
71    }
72
73    for (int i = 0; i < n_layer; ++i) {
74        auto & layer = layers[i];
75
76        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
77
78        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
79        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
80
81        // optional bias tensors
82        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
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_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
106            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
107            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
108
109            // For Granite MoE Shared
110            if (hparams.n_ff_shexp > 0) {
111                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
112                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
113                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
114            }
115        }
116    }
117}
118
119std::unique_ptr<llm_graph_context> llama_model_granite::build_arch_graph(const llm_graph_params & params) const {
120    return std::make_unique<graph>(*this, params);
121}
122
123llama_model_granite::graph::graph(
124    const llama_model & model,
125    const llm_graph_params & params)
126    : llm_graph_context(params) {
127
128    const int64_t n_embd_head = hparams.n_embd_head_v();
129
130    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
131    GGML_ASSERT(n_embd_head == n_rot);
132
133    ggml_tensor * cur;
134    ggml_tensor * inpL;
135
136    inpL = build_inp_embd(model.tok_embd);
137
138    // inp_pos - built only if rope enabled
139    ggml_tensor * inp_pos = nullptr;
140    if (hparams.has_rope(0)) {
141        inp_pos = build_inp_pos();
142    }
143    auto * inp_attn = build_attn_inp_kv();
144
145    ggml_tensor * inp_out_ids = build_inp_out_ids();
146
147    for (int il = 0; il < n_layer; ++il) {
148
149        // Granite Vision 4.1 deepstack: inject the projector stream that
150        // targets decoder layer `il` before the decoder runs.
151        // NOTE: skip the first deepstack layer since that's inpL
152        const auto & deepstack_emb_idx = hparams.deepstack_mapping_arr[il];
153        if (il > 0 && deepstack_emb_idx >= 0) {
154            ggml_tensor * ds = ggml_view_2d(ctx0,
155                res->t_inp_embd, n_embd, n_tokens,
156                res->t_inp_embd->nb[1],
157                deepstack_emb_idx * n_embd * sizeof(float));
158            inpL = ggml_add(ctx0, inpL, ds);
159            cb(inpL, "deepstack_in", il);
160        }
161
162        ggml_tensor * inpSA = inpL;
163
164        // norm
165        cur = build_norm(inpL,
166                model.layers[il].attn_norm, NULL,
167                LLM_NORM_RMS, il);
168        cb(cur, "attn_norm", il);
169
170        // self-attention
171        cur = build_attention_layer(
172            cur, inp_pos, inp_attn,
173            model, n_embd_head, il);
174
175        if (il == n_layer - 1 && inp_out_ids) {
176            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
177            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
178        }
179        // ffn
180        cur = build_layer_ffn(cur, inpSA, model, il);
181
182        // input for next layer
183        inpL = cur;
184    }
185    cur = inpL;
186
187    cur = build_norm(cur,
188            model.output_norm, NULL,
189            LLM_NORM_RMS, -1);
190
191    cb(cur, "result_norm", -1);
192    res->t_embd = cur;
193
194    // lm_head
195    cur = build_lora_mm(model.output, cur, model.output_s);
196
197    // For Granite architectures - scale logits
198    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
199    cb(cur, "result_output", -1);
200    res->t_logits = cur;
201
202    ggml_build_forward_expand(gf, cur);
203}
204
205ggml_tensor * llama_model_granite::graph::build_attention_layer(
206          ggml_tensor             * cur,
207          ggml_tensor             * inp_pos,
208          llm_graph_input_attn_kv * inp_attn,
209    const llama_model             & model,
210    const int64_t                 n_embd_head,
211    const int                     il) {
212
213    auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
214            n_embd_head, hparams.n_head(il), hparams.n_head_kv(il), il);
215
216    if (hparams.has_rope(il)) {
217        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
218        Qcur = ggml_rope_ext(
219                ctx0, Qcur, inp_pos, rope_factors,
220                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
221                ext_factor, attn_factor, beta_fast, beta_slow
222                );
223
224        Kcur = ggml_rope_ext(
225                ctx0, Kcur, inp_pos, rope_factors,
226                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
227                ext_factor, attn_factor, beta_fast, beta_slow
228                );
229    }
230
231    cb(Qcur, "Qcur", il);
232    cb(Kcur, "Kcur", il);
233    cb(Vcur, "Vcur", il);
234
235    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
236    cur = build_attn(inp_attn,
237            model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
238            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
239            cb(cur, "attn_out", il);
240    return cur;
241}
242
243ggml_tensor * llama_model_granite::graph::build_layer_ffn(
244          ggml_tensor       * cur,
245          ggml_tensor       * inpSA,
246    const llama_model       & model,
247    const int                 il) {
248
249    // For Granite architectures - scale residual
250    if (hparams.f_residual_scale) {
251        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
252    }
253    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
254    cb(ffn_inp, "ffn_inp", il);
255
256    // feed-forward network (non-MoE)
257    if (model.layers[il].ffn_gate_inp == nullptr) {
258
259        cur = build_norm(ffn_inp,
260                model.layers[il].ffn_norm, NULL,
261                LLM_NORM_RMS, il);
262                cb(cur, "ffn_norm", il);
263
264        cur = build_ffn(cur,
265                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
266                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
267                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
268                NULL,
269                LLM_FFN_SILU, LLM_FFN_PAR, il);
270                cb(cur, "ffn_out", il);
271
272    } else {
273        // MoE branch
274        cur = build_norm(ffn_inp,
275                model.layers[il].ffn_norm, NULL,
276                LLM_NORM_RMS, il);
277                cb(cur, "ffn_norm", il);
278
279        ggml_tensor * moe_out = build_moe_ffn(cur,
280                model.layers[il].ffn_gate_inp,
281                model.layers[il].ffn_up_exps,
282                model.layers[il].ffn_gate_exps,
283                model.layers[il].ffn_down_exps,
284                nullptr,
285                n_expert, n_expert_used,
286                LLM_FFN_SILU, true,
287                hparams.expert_weights_scale,
288                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
289                il);
290        cb(moe_out, "ffn_moe_out", il);
291
292        // For Granite MoE Shared
293        if (hparams.n_ff_shexp > 0) {
294            ggml_tensor * ffn_shexp = build_ffn(cur,
295                model.layers[il].ffn_up_shexp,   NULL, NULL,
296                model.layers[il].ffn_gate_shexp, NULL, NULL,
297                model.layers[il].ffn_down_shexp, NULL, NULL,
298                NULL,
299                LLM_FFN_SILU, LLM_FFN_PAR, il);
300            cb(ffn_shexp, "ffn_shexp", il);
301
302            cur = ggml_add(ctx0, moe_out, ffn_shexp);
303            cb(cur, "ffn_out", il);
304        } else {
305            cur = moe_out;
306        }
307    }
308
309    // For Granite architectures - scale residual
310    if (hparams.f_residual_scale) {
311        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
312    }
313    cur = ggml_add(ctx0, cur, ffn_inp);
314    cb(cur, "ffn_out", il);
315
316    cur = build_cvec(cur, il);
317    cb(cur, "l_out", il);
318
319    return cur;
320}