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 →
13.3 KiB302 linesraw
1#include "models.h"
2
3void llama_model_granite_hybrid::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_LOGIT_SCALE,                 hparams.f_logit_scale, /* required */ false);
6    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, /* required */ false);
7    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, /* required */ false);
8    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, /* required */ false);
9
10    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
11    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
12    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
13    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
14    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
15
16    // Granite uses rope_finetuned as a switch for rope, so default to true
17    bool rope_finetuned = true;
18    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
19    hparams.rope_finetuned = rope_finetuned; // needed for round trip save
20    std::fill(hparams.rope_pattern.begin(), hparams.rope_pattern.end(), rope_finetuned);
21
22    // A layer is recurrent IFF the n_head_kv value is set to 0
23    for (uint32_t i = 0; i < hparams.n_layer(); ++i) {
24        hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;
25    }
26
27    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
28
29    switch (hparams.n_embd) {
30        case 768: type = LLM_TYPE_350M; break;
31        case 1536: type = (hparams.n_ff() == 512 ? LLM_TYPE_7B_A1B : LLM_TYPE_1B); break;
32        case 2048: case 2560: type = LLM_TYPE_3B; break;
33        case 4096: type = LLM_TYPE_32B_A9B; break;
34        default: type = LLM_TYPE_UNKNOWN;
35    }
36
37    // For Granite MoE Shared
38    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
39}
40
41void llama_model_granite_hybrid::load_arch_tensors(llama_model_loader &) {
42    LLAMA_LOAD_LOCALS;
43
44    // mamba2 Mixer SSM params
45    // NOTE: int64_t for tensor dimensions
46    const int64_t d_conv     = hparams.ssm_d_conv;
47    const int64_t d_inner    = hparams.ssm_d_inner;
48    const int64_t d_state    = hparams.ssm_d_state;
49    const int64_t n_ssm_head = hparams.ssm_dt_rank;
50    const int64_t n_group    = hparams.ssm_n_group;
51    const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
52
53    // only an expansion factor of 2 is supported for now
54    GGML_ASSERT(2 * n_embd == d_inner);
55
56    // embeddings
57    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
58
59    // output
60    {
61        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
62        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
63        // if output is NULL, init from the input tok embed, duplicated to allow offloading
64        if (output == NULL) {
65            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
66        }
67    }
68
69    for (int i = 0; i < n_layer; ++i) {
70        auto & layer = layers[i];
71
72        // norm
73        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
74
75        if (hparams.is_recr(i)) {
76            // ssm layers
77            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
78
79            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
80            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED);
81
82            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0);
83
84            // no "weight" suffix for these
85            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0);
86            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0);
87
88            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
89
90            // out_proj
91            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
92        } else {
93            // attention layers (with optional bias)
94            const int64_t n_head_i = hparams.n_head(i);
95            const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
96            const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
97            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, 0);
98            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
99            layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
100        }
101
102        // feed forward (w/ optional biases)
103        if (n_expert > 0) {
104            // MoE FFN
105            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
106            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
107            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
108            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
109            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
110            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
111
112            // For Granite MoE Shared
113            if (hparams.n_ff_shexp > 0) {
114                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
115                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
116                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
117            }
118        } else {
119            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
120            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
121            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
122            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
123            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
124            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
125            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
126            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
127        }
128    }
129}
130
131std::unique_ptr<llm_graph_context> llama_model_granite_hybrid::build_arch_graph(const llm_graph_params & params) const {
132    return std::make_unique<graph>(*this, params);
133}
134
135llama_model_granite_hybrid::graph::graph(const llama_model & model, const llm_graph_params & params) :
136    llm_build_mamba_base(params) {
137    const int64_t n_embd_head = hparams.n_embd_head_v();
138    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
139
140    ggml_tensor * cur;
141    ggml_tensor * inpL;
142
143    inpL = build_inp_embd(model.tok_embd);
144
145    auto * inp = build_inp_mem_hybrid();
146
147    ggml_tensor * inp_out_ids = build_inp_out_ids();
148
149    // Positional embeddings populated if rope enabled
150    ggml_tensor * inp_pos = nullptr;
151    if (hparams.has_rope(0)) {
152        inp_pos = build_inp_pos();
153    }
154
155    for (int il = 0; il < n_layer; ++il) {
156        struct ggml_tensor * inpSA = inpL;
157
158        // norm
159        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
160        cb(cur, "attn_norm", il);
161
162        if (hparams.is_recr(il)) {
163            // ssm layer //
164            cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
165        } else {
166            // attention layer //
167            cur = build_attention_layer(cur, inp_pos, inp->get_attn(), model, n_embd_head, il);
168        }
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
175        // ffn
176        cur = build_layer_ffn(cur, inpSA, model, il);
177
178        // input for next layer
179        inpL = cur;
180    }
181
182    cur = inpL;
183
184    cur = build_norm(cur, model.output_norm, NULL, 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    if (hparams.f_logit_scale) {
194        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
195    }
196    cb(cur, "result_output", -1);
197    res->t_logits = cur;
198
199    ggml_build_forward_expand(gf, cur);
200}
201
202ggml_tensor * llama_model_granite_hybrid::graph::build_attention_layer(ggml_tensor *             cur,
203                                                              ggml_tensor *             inp_pos,
204                                                              llm_graph_input_attn_kv * inp_attn,
205                                                              const llama_model &       model,
206                                                              const int64_t             n_embd_head,
207                                                              const int                 il) {
208    auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, hparams.n_head(il), hparams.n_head_kv(il), il);
209
210    if (hparams.has_rope(il)) {
211        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
212        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
213                             ext_factor, attn_factor, beta_fast, beta_slow);
214
215        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
216                             ext_factor, attn_factor, beta_fast, beta_slow);
217    }
218
219    cb(Qcur, "Qcur", il);
220    cb(Kcur, "Kcur", il);
221    cb(Vcur, "Vcur", il);
222
223    const float kq_scale =
224        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
225    cur = build_attn(inp_attn,
226            model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
227            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
228    cb(cur, "attn_out", il);
229    return cur;
230}
231
232ggml_tensor * llama_model_granite_hybrid::graph::build_layer_ffn(ggml_tensor *       cur,
233                                                        ggml_tensor *       inpSA,
234                                                        const llama_model & model,
235                                                        const int           il) {
236    // For Granite architectures - scale residual
237    if (hparams.f_residual_scale) {
238        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
239    }
240    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
241    cb(ffn_inp, "ffn_inp", il);
242
243    // feed-forward network (non-MoE)
244    if (model.layers[il].ffn_gate_inp == nullptr) {
245        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
246        cb(cur, "ffn_norm", il);
247
248        cur = build_ffn(cur,
249                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
250                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
251                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
252                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
253        cb(cur, "ffn_out", il);
254
255    } else {
256        // MoE branch
257        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
258        cb(cur, "ffn_norm", il);
259
260        ggml_tensor * moe_out =
261            build_moe_ffn(cur,
262                model.layers[il].ffn_gate_inp,
263                model.layers[il].ffn_up_exps,
264                model.layers[il].ffn_gate_exps,
265                model.layers[il].ffn_down_exps,
266                nullptr,
267                n_expert, n_expert_used,
268                LLM_FFN_SILU, true,
269                hparams.expert_weights_scale,
270                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
271                il);
272        cb(moe_out, "ffn_moe_out", il);
273
274        // For Granite MoE Shared
275        if (hparams.n_ff_shexp > 0) {
276            ggml_tensor * ffn_shexp =
277                build_ffn(cur,
278                    model.layers[il].ffn_up_shexp, NULL, NULL,
279                    model.layers[il].ffn_gate_shexp, NULL, NULL,
280                    model.layers[il].ffn_down_shexp, NULL, NULL,
281                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
282            cb(ffn_shexp, "ffn_shexp", il);
283
284            cur = ggml_add(ctx0, moe_out, ffn_shexp);
285            cb(cur, "ffn_out", il);
286        } else {
287            cur = moe_out;
288        }
289    }
290
291    // For Granite architectures - scale residual
292    if (hparams.f_residual_scale) {
293        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
294    }
295    cur = ggml_add(ctx0, cur, ffn_inp);
296    cb(cur, "ffn_out", il);
297
298    cur = build_cvec(cur, il);
299    cb(cur, "l_out", il);
300
301    return cur;
302}