/src/llama.cpp/src/llama-arch.cpp
Line | Count | Source |
1 | | #include "llama-arch.h" |
2 | | |
3 | | #include "llama-impl.h" |
4 | | |
5 | | #include <map> |
6 | | #include <set> |
7 | | #include <vector> |
8 | | |
9 | | static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = { |
10 | | { LLM_ARCH_CLIP, "clip" }, // dummy, only used by llama-quantize |
11 | | { LLM_ARCH_LLAMA, "llama" }, |
12 | | { LLM_ARCH_LLAMA4, "llama4" }, |
13 | | { LLM_ARCH_DECI, "deci" }, |
14 | | { LLM_ARCH_FALCON, "falcon" }, |
15 | | { LLM_ARCH_GROK, "grok" }, |
16 | | { LLM_ARCH_GPT2, "gpt2" }, |
17 | | { LLM_ARCH_GPTJ, "gptj" }, |
18 | | { LLM_ARCH_GPTNEOX, "gptneox" }, |
19 | | { LLM_ARCH_MPT, "mpt" }, |
20 | | { LLM_ARCH_BAICHUAN, "baichuan" }, |
21 | | { LLM_ARCH_STARCODER, "starcoder" }, |
22 | | { LLM_ARCH_REFACT, "refact" }, |
23 | | { LLM_ARCH_BERT, "bert" }, |
24 | | { LLM_ARCH_MODERN_BERT, "modern-bert" }, |
25 | | { LLM_ARCH_NOMIC_BERT, "nomic-bert" }, |
26 | | { LLM_ARCH_NOMIC_BERT_MOE, "nomic-bert-moe" }, |
27 | | { LLM_ARCH_NEO_BERT, "neo-bert" }, |
28 | | { LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" }, |
29 | | { LLM_ARCH_JINA_BERT_V3, "jina-bert-v3" }, |
30 | | { LLM_ARCH_EUROBERT, "eurobert" }, |
31 | | { LLM_ARCH_BLOOM, "bloom" }, |
32 | | { LLM_ARCH_STABLELM, "stablelm" }, |
33 | | { LLM_ARCH_QWEN, "qwen" }, |
34 | | { LLM_ARCH_QWEN2, "qwen2" }, |
35 | | { LLM_ARCH_QWEN2MOE, "qwen2moe" }, |
36 | | { LLM_ARCH_QWEN2VL, "qwen2vl" }, |
37 | | { LLM_ARCH_QWEN3, "qwen3" }, |
38 | | { LLM_ARCH_QWEN3MOE, "qwen3moe" }, |
39 | | { LLM_ARCH_QWEN3NEXT, "qwen3next" }, |
40 | | { LLM_ARCH_QWEN3VL, "qwen3vl" }, |
41 | | { LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" }, |
42 | | { LLM_ARCH_QWEN35, "qwen35" }, |
43 | | { LLM_ARCH_QWEN35MOE, "qwen35moe" }, |
44 | | { LLM_ARCH_PHI2, "phi2" }, |
45 | | { LLM_ARCH_PHI3, "phi3" }, |
46 | | { LLM_ARCH_PHIMOE, "phimoe" }, |
47 | | { LLM_ARCH_PLAMO, "plamo" }, |
48 | | { LLM_ARCH_PLAMO2, "plamo2" }, |
49 | | { LLM_ARCH_PLAMO3, "plamo3" }, |
50 | | { LLM_ARCH_CODESHELL, "codeshell" }, |
51 | | { LLM_ARCH_ORION, "orion" }, |
52 | | { LLM_ARCH_INTERNLM2, "internlm2" }, |
53 | | { LLM_ARCH_MINICPM, "minicpm" }, |
54 | | { LLM_ARCH_MINICPM3, "minicpm3" }, |
55 | | { LLM_ARCH_GEMMA, "gemma" }, |
56 | | { LLM_ARCH_GEMMA2, "gemma2" }, |
57 | | { LLM_ARCH_GEMMA3, "gemma3" }, |
58 | | { LLM_ARCH_GEMMA3N, "gemma3n" }, |
59 | | { LLM_ARCH_GEMMA4, "gemma4" }, |
60 | | { LLM_ARCH_GEMMA_EMBEDDING, "gemma-embedding" }, |
61 | | { LLM_ARCH_STARCODER2, "starcoder2" }, |
62 | | { LLM_ARCH_MAMBA, "mamba" }, |
63 | | { LLM_ARCH_MAMBA2, "mamba2" }, |
64 | | { LLM_ARCH_JAMBA, "jamba" }, |
65 | | { LLM_ARCH_FALCON_H1, "falcon-h1" }, |
66 | | { LLM_ARCH_XVERSE, "xverse" }, |
67 | | { LLM_ARCH_COMMAND_R, "command-r" }, |
68 | | { LLM_ARCH_COHERE2, "cohere2" }, |
69 | | { LLM_ARCH_DBRX, "dbrx" }, |
70 | | { LLM_ARCH_OLMO, "olmo" }, |
71 | | { LLM_ARCH_OLMO2, "olmo2" }, |
72 | | { LLM_ARCH_OLMOE, "olmoe" }, |
73 | | { LLM_ARCH_OPENELM, "openelm" }, |
74 | | { LLM_ARCH_ARCTIC, "arctic" }, |
75 | | { LLM_ARCH_DEEPSEEK, "deepseek" }, |
76 | | { LLM_ARCH_DEEPSEEK2, "deepseek2" }, |
77 | | { LLM_ARCH_DEEPSEEK2OCR, "deepseek2-ocr" }, |
78 | | { LLM_ARCH_CHATGLM, "chatglm" }, |
79 | | { LLM_ARCH_GLM4, "glm4" }, |
80 | | { LLM_ARCH_GLM4_MOE, "glm4moe" }, |
81 | | { LLM_ARCH_GLM_DSA, "glm-dsa" }, |
82 | | { LLM_ARCH_BITNET, "bitnet" }, |
83 | | { LLM_ARCH_T5, "t5" }, |
84 | | { LLM_ARCH_T5ENCODER, "t5encoder" }, |
85 | | { LLM_ARCH_JAIS, "jais" }, |
86 | | { LLM_ARCH_JAIS2, "jais2" }, |
87 | | { LLM_ARCH_NEMOTRON, "nemotron" }, |
88 | | { LLM_ARCH_NEMOTRON_H, "nemotron_h" }, |
89 | | { LLM_ARCH_NEMOTRON_H_MOE, "nemotron_h_moe" }, |
90 | | { LLM_ARCH_EXAONE, "exaone" }, |
91 | | { LLM_ARCH_EXAONE4, "exaone4" }, |
92 | | { LLM_ARCH_EXAONE_MOE, "exaone-moe" }, |
93 | | { LLM_ARCH_RWKV6, "rwkv6" }, |
94 | | { LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" }, |
95 | | { LLM_ARCH_RWKV7, "rwkv7" }, |
96 | | { LLM_ARCH_ARWKV7, "arwkv7" }, |
97 | | { LLM_ARCH_GRANITE, "granite" }, |
98 | | { LLM_ARCH_GRANITE_MOE, "granitemoe" }, |
99 | | { LLM_ARCH_GRANITE_HYBRID, "granitehybrid" }, |
100 | | { LLM_ARCH_CHAMELEON, "chameleon" }, |
101 | | { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" }, |
102 | | { LLM_ARCH_PLM, "plm" }, |
103 | | { LLM_ARCH_BAILINGMOE, "bailingmoe" }, |
104 | | { LLM_ARCH_BAILINGMOE2, "bailingmoe2" }, |
105 | | { LLM_ARCH_DOTS1, "dots1" }, |
106 | | { LLM_ARCH_ARCEE, "arcee" }, |
107 | | { LLM_ARCH_AFMOE, "afmoe" }, |
108 | | { LLM_ARCH_ERNIE4_5, "ernie4_5" }, |
109 | | { LLM_ARCH_ERNIE4_5_MOE, "ernie4_5-moe" }, |
110 | | { LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" }, |
111 | | { LLM_ARCH_HUNYUAN_DENSE, "hunyuan-dense" }, |
112 | | { LLM_ARCH_SMOLLM3, "smollm3" }, |
113 | | { LLM_ARCH_OPENAI_MOE, "gpt-oss" }, |
114 | | { LLM_ARCH_LFM2, "lfm2" }, |
115 | | { LLM_ARCH_LFM2MOE, "lfm2moe" }, |
116 | | { LLM_ARCH_DREAM, "dream" }, |
117 | | { LLM_ARCH_SMALLTHINKER, "smallthinker" }, |
118 | | { LLM_ARCH_LLADA, "llada" }, |
119 | | { LLM_ARCH_LLADA_MOE, "llada-moe" }, |
120 | | { LLM_ARCH_SEED_OSS, "seed_oss" }, |
121 | | { LLM_ARCH_GROVEMOE, "grovemoe" }, |
122 | | { LLM_ARCH_APERTUS, "apertus" }, |
123 | | { LLM_ARCH_MINIMAX_M2, "minimax-m2" }, |
124 | | { LLM_ARCH_COGVLM, "cogvlm" }, |
125 | | { LLM_ARCH_RND1, "rnd1" }, |
126 | | { LLM_ARCH_PANGU_EMBED, "pangu-embedded" }, |
127 | | { LLM_ARCH_MISTRAL3, "mistral3" }, |
128 | | { LLM_ARCH_MISTRAL4, "mistral4" }, |
129 | | { LLM_ARCH_PADDLEOCR, "paddleocr" }, |
130 | | { LLM_ARCH_MIMO2, "mimo2" }, |
131 | | { LLM_ARCH_STEP35, "step35" }, |
132 | | { LLM_ARCH_LLAMA_EMBED, "llama-embed" }, |
133 | | { LLM_ARCH_MAINCODER, "maincoder" }, |
134 | | { LLM_ARCH_KIMI_LINEAR, "kimi-linear" }, |
135 | | { LLM_ARCH_UNKNOWN, "(unknown)" }, |
136 | | }; |
137 | | |
138 | | static const std::map<llm_kv, const char *> LLM_KV_NAMES = { |
139 | | { LLM_KV_GENERAL_TYPE, "general.type" }, |
140 | | { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, |
141 | | { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, |
142 | | { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, |
143 | | { LLM_KV_GENERAL_FILE_TYPE, "general.file_type" }, |
144 | | { LLM_KV_GENERAL_SAMPLING_SEQUENCE, "general.sampling.sequence" }, |
145 | | { LLM_KV_GENERAL_SAMPLING_TOP_K, "general.sampling.top_k" }, |
146 | | { LLM_KV_GENERAL_SAMPLING_TOP_P, "general.sampling.top_p" }, |
147 | | { LLM_KV_GENERAL_SAMPLING_MIN_P, "general.sampling.min_p" }, |
148 | | { LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY, "general.sampling.xtc_probability" }, |
149 | | { LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD, "general.sampling.xtc_threshold" }, |
150 | | { LLM_KV_GENERAL_SAMPLING_TEMP, "general.sampling.temp" }, |
151 | | { LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N, "general.sampling.penalty_last_n" }, |
152 | | { LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT, "general.sampling.penalty_repeat" }, |
153 | | { LLM_KV_GENERAL_SAMPLING_MIROSTAT, "general.sampling.mirostat" }, |
154 | | { LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU, "general.sampling.mirostat_tau" }, |
155 | | { LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA, "general.sampling.mirostat_eta" }, |
156 | | { LLM_KV_GENERAL_NAME, "general.name" }, |
157 | | { LLM_KV_GENERAL_AUTHOR, "general.author" }, |
158 | | { LLM_KV_GENERAL_VERSION, "general.version" }, |
159 | | { LLM_KV_GENERAL_URL, "general.url" }, |
160 | | { LLM_KV_GENERAL_DESCRIPTION, "general.description" }, |
161 | | { LLM_KV_GENERAL_LICENSE, "general.license" }, |
162 | | { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" }, |
163 | | { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" }, |
164 | | |
165 | | { LLM_KV_VOCAB_SIZE, "%s.vocab_size" }, |
166 | | { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, |
167 | | { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" }, |
168 | | { LLM_KV_EMBEDDING_LENGTH_OUT, "%s.embedding_length_out" }, |
169 | | { LLM_KV_EMBEDDING_LENGTH_PER_LAYER, "%s.embedding_length_per_layer_input" }, |
170 | | { LLM_KV_FEATURES_LENGTH, "%s.features_length" }, |
171 | | { LLM_KV_BLOCK_COUNT, "%s.block_count" }, |
172 | | { LLM_KV_LEADING_DENSE_BLOCK_COUNT, "%s.leading_dense_block_count" }, |
173 | | { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" }, |
174 | | { LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" }, |
175 | | { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" }, |
176 | | { LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, "%s.expert_chunk_feed_forward_length" }, |
177 | | { LLM_KV_SWIGLU_CLAMP_EXP, "%s.swiglu_clamp_exp" }, |
178 | | { LLM_KV_SWIGLU_CLAMP_SHEXP, "%s.swiglu_clamp_shexp" }, |
179 | | { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, |
180 | | { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, |
181 | | { LLM_KV_EXPERT_COUNT, "%s.expert_count" }, |
182 | | { LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" }, |
183 | | { LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" }, |
184 | | { LLM_KV_EXPERT_GROUP_COUNT, "%s.expert_group_count" }, |
185 | | { LLM_KV_EXPERT_GROUP_USED_COUNT, "%s.expert_group_used_count" }, |
186 | | { LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" }, |
187 | | { LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" }, |
188 | | { LLM_KV_EXPERT_GATING_FUNC, "%s.expert_gating_func" }, |
189 | | { LLM_KV_EXPERT_GROUP_SCALE, "%s.expert_group_scale" }, |
190 | | { LLM_KV_EXPERTS_PER_GROUP, "%s.experts_per_group" }, |
191 | | { LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" }, |
192 | | { LLM_KV_MOE_LATENT_SIZE, "%s.moe_latent_size" }, |
193 | | { LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" }, |
194 | | { LLM_KV_NUM_DEEPSTACK_LAYERS, "%s.n_deepstack_layers" }, |
195 | | { LLM_KV_POOLING_TYPE, "%s.pooling_type" }, |
196 | | { LLM_KV_LOGIT_SCALE, "%s.logit_scale" }, |
197 | | { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" }, |
198 | | { LLM_KV_DECODER_BLOCK_COUNT, "%s.decoder_block_count" }, |
199 | | { LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" }, |
200 | | { LLM_KV_ROUTER_LOGIT_SOFTCAPPING, "%s.router_logit_softcapping" }, |
201 | | { LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" }, |
202 | | { LLM_KV_SWIN_NORM, "%s.swin_norm" }, |
203 | | { LLM_KV_RESCALE_EVERY_N_LAYERS, "%s.rescale_every_n_layers" }, |
204 | | { LLM_KV_TIME_MIX_EXTRA_DIM, "%s.time_mix_extra_dim" }, |
205 | | { LLM_KV_TIME_DECAY_EXTRA_DIM, "%s.time_decay_extra_dim" }, |
206 | | { LLM_KV_RESIDUAL_SCALE, "%s.residual_scale" }, |
207 | | { LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" }, |
208 | | { LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" }, |
209 | | { LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" }, |
210 | | { LLM_KV_FULL_ATTENTION_INTERVAL, "%s.full_attention_interval" }, |
211 | | |
212 | | { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, |
213 | | { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, |
214 | | { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, |
215 | | { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, |
216 | | { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" }, |
217 | | { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" }, |
218 | | { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, |
219 | | { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, |
220 | | { LLM_KV_ATTENTION_GROUPNORM_EPS, "%s.attention.group_norm_epsilon" }, |
221 | | { LLM_KV_ATTENTION_GROUPNORM_GROUPS, "%s.attention.group_norm_groups" }, |
222 | | { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" }, |
223 | | { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, |
224 | | { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, |
225 | | { LLM_KV_ATTENTION_DECAY_LORA_RANK, "%s.attention.decay_lora_rank" }, |
226 | | { LLM_KV_ATTENTION_ICLR_LORA_RANK, "%s.attention.iclr_lora_rank" }, |
227 | | { LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" }, |
228 | | { LLM_KV_ATTENTION_GATE_LORA_RANK, "%s.attention.gate_lora_rank" }, |
229 | | { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, |
230 | | { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, |
231 | | { LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, "%s.attention.sliding_window_pattern" }, |
232 | | { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" }, |
233 | | { LLM_KV_ATTENTION_OUTPUT_SCALE, "%s.attention.output_scale" }, |
234 | | { LLM_KV_ATTENTION_TEMPERATURE_LENGTH, "%s.attention.temperature_length" }, |
235 | | { LLM_KV_ATTENTION_TEMPERATURE_SCALE, "%s.attention.temperature_scale" }, |
236 | | { LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" }, |
237 | | { LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" }, |
238 | | { LLM_KV_ATTENTION_KEY_LENGTH_SWA, "%s.attention.key_length_swa" }, |
239 | | { LLM_KV_ATTENTION_VALUE_LENGTH_SWA, "%s.attention.value_length_swa" }, |
240 | | { LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, "%s.attention.indexer.head_count" }, |
241 | | { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, "%s.attention.indexer.key_length" }, |
242 | | { LLM_KV_ATTENTION_INDEXER_TOP_K, "%s.attention.indexer.top_k" }, |
243 | | { LLM_KV_ATTENTION_SHARED_KV_LAYERS, "%s.attention.shared_kv_layers" }, |
244 | | |
245 | | { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, |
246 | | { LLM_KV_ROPE_DIMENSION_COUNT_SWA, "%s.rope.dimension_count_swa" }, |
247 | | { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, |
248 | | { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, |
249 | | { LLM_KV_ROPE_FREQ_BASE_SWA, "%s.rope.freq_base_swa" }, |
250 | | { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, |
251 | | { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" }, |
252 | | { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" }, |
253 | | { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" }, |
254 | | { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, |
255 | | { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, |
256 | | { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" }, |
257 | | { LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, "%s.rope.scaling.yarn_ext_factor" }, |
258 | | { LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor" }, |
259 | | { LLM_KV_ROPE_SCALING_YARN_BETA_FAST, "%s.rope.scaling.yarn_beta_fast" }, |
260 | | { LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, "%s.rope.scaling.yarn_beta_slow" }, |
261 | | |
262 | | { LLM_KV_SPLIT_NO, "split.no" }, |
263 | | { LLM_KV_SPLIT_COUNT, "split.count" }, |
264 | | { LLM_KV_SPLIT_TENSORS_COUNT, "split.tensors.count" }, |
265 | | |
266 | | { LLM_KV_SSM_CONV_KERNEL, "%s.ssm.conv_kernel" }, |
267 | | { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" }, |
268 | | { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" }, |
269 | | { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" }, |
270 | | { LLM_KV_SSM_GROUP_COUNT, "%s.ssm.group_count" }, |
271 | | { LLM_KV_SSM_DT_B_C_RMS, "%s.ssm.dt_b_c_rms" }, |
272 | | |
273 | | { LLM_KV_KDA_HEAD_DIM, "%s.kda.head_dim" }, |
274 | | |
275 | | { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" }, |
276 | | |
277 | | { LLM_KV_POSNET_EMBEDDING_LENGTH, "%s.posnet.embedding_length" }, |
278 | | { LLM_KV_POSNET_BLOCK_COUNT, "%s.posnet.block_count" }, |
279 | | |
280 | | { LLM_KV_CONVNEXT_EMBEDDING_LENGTH, "%s.convnext.embedding_length" }, |
281 | | { LLM_KV_CONVNEXT_BLOCK_COUNT, "%s.convnext.block_count" }, |
282 | | |
283 | | { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" }, |
284 | | |
285 | | { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" }, |
286 | | // sentence-transformers dense modules feature dims |
287 | | { LLM_KV_DENSE_2_FEAT_IN, "%s.dense_2_feat_in" }, |
288 | | { LLM_KV_DENSE_2_FEAT_OUT, "%s.dense_2_feat_out" }, |
289 | | { LLM_KV_DENSE_3_FEAT_IN, "%s.dense_3_feat_in" }, |
290 | | { LLM_KV_DENSE_3_FEAT_OUT, "%s.dense_3_feat_out" }, |
291 | | |
292 | | { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, |
293 | | { LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" }, |
294 | | { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, |
295 | | { LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" }, |
296 | | { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, "tokenizer.ggml.token_type_count" }, |
297 | | { LLM_KV_TOKENIZER_SCORES, "tokenizer.ggml.scores" }, |
298 | | { LLM_KV_TOKENIZER_MERGES, "tokenizer.ggml.merges" }, |
299 | | { LLM_KV_TOKENIZER_BOS_ID, "tokenizer.ggml.bos_token_id" }, |
300 | | { LLM_KV_TOKENIZER_EOS_ID, "tokenizer.ggml.eos_token_id" }, |
301 | | { LLM_KV_TOKENIZER_EOT_ID, "tokenizer.ggml.eot_token_id" }, |
302 | | { LLM_KV_TOKENIZER_EOM_ID, "tokenizer.ggml.eom_token_id" }, |
303 | | { LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" }, |
304 | | { LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" }, |
305 | | { LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" }, |
306 | | { LLM_KV_TOKENIZER_CLS_ID, "tokenizer.ggml.cls_token_id" }, |
307 | | { LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_token_id" }, |
308 | | { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" }, |
309 | | { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" }, |
310 | | { LLM_KV_TOKENIZER_ADD_SEP, "tokenizer.ggml.add_sep_token" }, |
311 | | { LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" }, |
312 | | { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, "tokenizer.ggml.remove_extra_whitespaces" }, |
313 | | { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap" }, |
314 | | { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" }, |
315 | | { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" }, |
316 | | { LLM_KV_TOKENIZER_CHAT_TEMPLATE, "tokenizer.chat_template" }, |
317 | | { LLM_KV_TOKENIZER_FIM_PRE_ID, "tokenizer.ggml.fim_pre_token_id" }, |
318 | | { LLM_KV_TOKENIZER_FIM_SUF_ID, "tokenizer.ggml.fim_suf_token_id" }, |
319 | | { LLM_KV_TOKENIZER_FIM_MID_ID, "tokenizer.ggml.fim_mid_token_id" }, |
320 | | { LLM_KV_TOKENIZER_FIM_PAD_ID, "tokenizer.ggml.fim_pad_token_id" }, |
321 | | { LLM_KV_TOKENIZER_FIM_REP_ID, "tokenizer.ggml.fim_rep_token_id" }, |
322 | | { LLM_KV_TOKENIZER_FIM_SEP_ID, "tokenizer.ggml.fim_sep_token_id" }, |
323 | | |
324 | | { LLM_KV_ADAPTER_TYPE, "adapter.type" }, |
325 | | { LLM_KV_ADAPTER_LORA_ALPHA, "adapter.lora.alpha" }, |
326 | | { LLM_KV_ADAPTER_LORA_TASK_NAME, "adapter.lora.task_name" }, |
327 | | { LLM_KV_ADAPTER_LORA_PROMPT_PREFIX, "adapter.lora.prompt_prefix" }, |
328 | | { LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS, "adapter.alora.invocation_tokens" }, |
329 | | |
330 | | { LLM_KV_XIELU_ALPHA_N, "xielu.alpha_n" }, |
331 | | { LLM_KV_XIELU_ALPHA_P, "xielu.alpha_p" }, |
332 | | { LLM_KV_XIELU_BETA, "xielu.beta" }, |
333 | | { LLM_KV_XIELU_EPS, "xielu.eps" }, |
334 | | |
335 | | // deprecated |
336 | | { LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" }, |
337 | | { LLM_KV_TOKENIZER_SUFFIX_ID, "tokenizer.ggml.suffix_token_id" }, |
338 | | { LLM_KV_TOKENIZER_MIDDLE_ID, "tokenizer.ggml.middle_token_id" }, |
339 | | }; |
340 | | |
341 | | static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = { |
342 | | { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, |
343 | | { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, |
344 | | { LLM_TENSOR_OUTPUT_NORM_LFM2, "token_embd_norm" }, // fix for wrong tensor name |
345 | | { LLM_TENSOR_OUTPUT, "output" }, |
346 | | { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, |
347 | | { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, |
348 | | { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, |
349 | | { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, |
350 | | { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, |
351 | | { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, |
352 | | { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, |
353 | | { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, |
354 | | { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, |
355 | | { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, |
356 | | { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, |
357 | | { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, |
358 | | { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" }, |
359 | | { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" }, |
360 | | { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, |
361 | | { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, |
362 | | { LLM_TENSOR_FFN_GATE_UP_EXPS, "blk.%d.ffn_gate_up_exps" }, |
363 | | { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, |
364 | | { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, |
365 | | { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, |
366 | | { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, |
367 | | { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, |
368 | | { LLM_TENSOR_ATTN_GATE, "blk.%d.attn_gate" }, |
369 | | { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, |
370 | | { LLM_TENSOR_FFN_POST_NORM_1, "blk.%d.post_ffw_norm_1" }, |
371 | | { LLM_TENSOR_FFN_POST_NORM_2, "blk.%d.post_ffw_norm_2" }, |
372 | | { LLM_TENSOR_FFN_PRE_NORM_2, "blk.%d.pre_ffw_norm_2" }, |
373 | | { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, |
374 | | { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, |
375 | | { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, |
376 | | { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" }, |
377 | | { LLM_TENSOR_FFN_LATENT_DOWN, "blk.%d.ffn_latent_down" }, |
378 | | { LLM_TENSOR_FFN_LATENT_UP, "blk.%d.ffn_latent_up" }, |
379 | | { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, |
380 | | { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, |
381 | | { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, |
382 | | { LLM_TENSOR_LAYER_OUT_SCALE, "blk.%d.layer_output_scale" }, |
383 | | { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, |
384 | | { LLM_TENSOR_POS_EMBD, "position_embd" }, |
385 | | { LLM_TENSOR_FFN_ACT, "blk.%d.ffn.act" }, |
386 | | { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, |
387 | | { LLM_TENSOR_TOKEN_TYPES, "token_types" }, |
388 | | { LLM_TENSOR_CLS, "cls" }, |
389 | | { LLM_TENSOR_CLS_OUT, "cls.output" }, |
390 | | { LLM_TENSOR_CLS_NORM, "cls.norm" }, |
391 | | { LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" }, |
392 | | { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" }, |
393 | | { LLM_TENSOR_SSM_A_NOSCAN, "blk.%d.ssm_a" }, |
394 | | { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, |
395 | | { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, |
396 | | { LLM_TENSOR_SSM_BETA_ALPHA, "blk.%d.ssm_ba" }, |
397 | | { LLM_TENSOR_SSM_ALPHA, "blk.%d.ssm_alpha" }, |
398 | | { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, |
399 | | { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, |
400 | | { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, |
401 | | { LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" }, |
402 | | { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" }, |
403 | | { LLM_TENSOR_SSM_X, "blk.%d.ssm_x" }, |
404 | | { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, |
405 | | { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" }, |
406 | | { LLM_TENSOR_SSM_DT_NORM, "blk.%d.ssm_dt_norm" }, |
407 | | { LLM_TENSOR_SSM_B_NORM, "blk.%d.ssm_b_norm" }, |
408 | | { LLM_TENSOR_SSM_C_NORM, "blk.%d.ssm_c_norm" }, |
409 | | { LLM_TENSOR_SSM_CONV1D_Q, "blk.%d.ssm_conv1d_q" }, |
410 | | { LLM_TENSOR_SSM_CONV1D_K, "blk.%d.ssm_conv1d_k" }, |
411 | | { LLM_TENSOR_SSM_CONV1D_V, "blk.%d.ssm_conv1d_v" }, |
412 | | { LLM_TENSOR_SSM_F_A, "blk.%d.ssm_f_a" }, |
413 | | { LLM_TENSOR_SSM_F_B, "blk.%d.ssm_f_b" }, |
414 | | { LLM_TENSOR_SSM_BETA, "blk.%d.ssm_beta" }, |
415 | | { LLM_TENSOR_SSM_G_A, "blk.%d.ssm_g_a" }, |
416 | | { LLM_TENSOR_SSM_G_B, "blk.%d.ssm_g_b" }, |
417 | | { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, |
418 | | { LLM_TENSOR_ATTN_Q_A_NORM, "blk.%d.attn_q_a_norm" }, |
419 | | { LLM_TENSOR_ATTN_KV_A_NORM, "blk.%d.attn_kv_a_norm" }, |
420 | | { LLM_TENSOR_ATTN_Q_A, "blk.%d.attn_q_a" }, |
421 | | { LLM_TENSOR_ATTN_Q_B, "blk.%d.attn_q_b" }, |
422 | | { LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" }, |
423 | | { LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" }, |
424 | | { LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "per_layer_token_embd" }, |
425 | | { LLM_TENSOR_PER_LAYER_MODEL_PROJ, "per_layer_model_proj" }, |
426 | | { LLM_TENSOR_PER_LAYER_PROJ_NORM, "per_layer_proj_norm" }, |
427 | | { LLM_TENSOR_ALTUP_UNEMBD_PROJ, "altup_unembd_proj" }, |
428 | | { LLM_TENSOR_ALTUP_PROJ, "altup_proj" }, |
429 | | { LLM_TENSOR_PER_LAYER_INP_GATE, "blk.%d.inp_gate" }, |
430 | | { LLM_TENSOR_PER_LAYER_PROJ, "blk.%d.proj" }, |
431 | | { LLM_TENSOR_PER_LAYER_POST_NORM, "blk.%d.post_norm" }, |
432 | | { LLM_TENSOR_ALTUP_CORRECT_COEF, "blk.%d.altup_correct_coef" }, |
433 | | { LLM_TENSOR_ALTUP_CORRECT_SCALE, "blk.%d.altup_correct_scale" }, |
434 | | { LLM_TENSOR_ALTUP_PREDICT_COEF, "blk.%d.altup_predict_coef" }, |
435 | | { LLM_TENSOR_ALTUP_ROUTER, "blk.%d.altup_router" }, |
436 | | { LLM_TENSOR_ALTUP_ROUTER_NORM, "blk.%d.altup_router_norm" }, |
437 | | { LLM_TENSOR_LAUREL_L, "blk.%d.laurel_l" }, |
438 | | { LLM_TENSOR_LAUREL_R, "blk.%d.laurel_r" }, |
439 | | { LLM_TENSOR_LAUREL_POST_NORM, "blk.%d.laurel_post_norm" }, |
440 | | { LLM_TENSOR_DENSE_2_OUT, "dense_2" }, |
441 | | { LLM_TENSOR_DENSE_3_OUT, "dense_3" }, |
442 | | { LLM_TENSOR_FFN_NORM_EXPS, "blk.%d.ffn_norm_exps" }, |
443 | | { LLM_TENSOR_ATTN_K_B, "blk.%d.attn_k_b" }, |
444 | | { LLM_TENSOR_ATTN_V_B, "blk.%d.attn_v_b" }, |
445 | | { LLM_TENSOR_NEXTN_EH_PROJ, "blk.%d.nextn.eh_proj" }, |
446 | | { LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" }, |
447 | | { LLM_TENSOR_NEXTN_ENORM, "blk.%d.nextn.enorm" }, |
448 | | { LLM_TENSOR_NEXTN_HNORM, "blk.%d.nextn.hnorm" }, |
449 | | { LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" }, |
450 | | { LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" }, |
451 | | { LLM_TENSOR_ATTN_SUB_NORM, "blk.%d.attn_sub_norm" }, |
452 | | { LLM_TENSOR_FFN_SUB_NORM, "blk.%d.ffn_sub_norm" }, |
453 | | { LLM_TENSOR_DEC_OUTPUT_NORM, "dec.output_norm" }, |
454 | | { LLM_TENSOR_DEC_ATTN_NORM, "dec.blk.%d.attn_norm" }, |
455 | | { LLM_TENSOR_DEC_ATTN_Q, "dec.blk.%d.attn_q" }, |
456 | | { LLM_TENSOR_DEC_ATTN_K, "dec.blk.%d.attn_k" }, |
457 | | { LLM_TENSOR_DEC_ATTN_V, "dec.blk.%d.attn_v" }, |
458 | | { LLM_TENSOR_DEC_ATTN_OUT, "dec.blk.%d.attn_o" }, |
459 | | { LLM_TENSOR_DEC_ATTN_REL_B, "dec.blk.%d.attn_rel_b" }, |
460 | | { LLM_TENSOR_DEC_CROSS_ATTN_NORM, "dec.blk.%d.cross_attn_norm" }, |
461 | | { LLM_TENSOR_DEC_CROSS_ATTN_Q, "dec.blk.%d.cross_attn_q" }, |
462 | | { LLM_TENSOR_DEC_CROSS_ATTN_K, "dec.blk.%d.cross_attn_k" }, |
463 | | { LLM_TENSOR_DEC_CROSS_ATTN_V, "dec.blk.%d.cross_attn_v" }, |
464 | | { LLM_TENSOR_DEC_CROSS_ATTN_OUT, "dec.blk.%d.cross_attn_o" }, |
465 | | { LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "dec.blk.%d.cross_attn_rel_b" }, |
466 | | { LLM_TENSOR_DEC_FFN_NORM, "dec.blk.%d.ffn_norm" }, |
467 | | { LLM_TENSOR_DEC_FFN_GATE, "dec.blk.%d.ffn_gate" }, |
468 | | { LLM_TENSOR_DEC_FFN_DOWN, "dec.blk.%d.ffn_down" }, |
469 | | { LLM_TENSOR_DEC_FFN_UP, "dec.blk.%d.ffn_up" }, |
470 | | { LLM_TENSOR_ENC_ATTN_NORM, "enc.blk.%d.attn_norm" }, |
471 | | { LLM_TENSOR_ENC_ATTN_Q, "enc.blk.%d.attn_q" }, |
472 | | { LLM_TENSOR_ENC_ATTN_K, "enc.blk.%d.attn_k" }, |
473 | | { LLM_TENSOR_ENC_ATTN_V, "enc.blk.%d.attn_v" }, |
474 | | { LLM_TENSOR_ENC_ATTN_OUT, "enc.blk.%d.attn_o" }, |
475 | | { LLM_TENSOR_ENC_ATTN_REL_B, "enc.blk.%d.attn_rel_b" }, |
476 | | { LLM_TENSOR_ENC_FFN_NORM, "enc.blk.%d.ffn_norm" }, |
477 | | { LLM_TENSOR_ENC_FFN_GATE, "enc.blk.%d.ffn_gate" }, |
478 | | { LLM_TENSOR_ENC_FFN_DOWN, "enc.blk.%d.ffn_down" }, |
479 | | { LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" }, |
480 | | { LLM_TENSOR_TIME_MIX_W1, "blk.%d.time_mix_w1" }, |
481 | | { LLM_TENSOR_TIME_MIX_W2, "blk.%d.time_mix_w2" }, |
482 | | { LLM_TENSOR_TIME_MIX_LERP_X, "blk.%d.time_mix_lerp_x" }, |
483 | | { LLM_TENSOR_TIME_MIX_LERP_W, "blk.%d.time_mix_lerp_w" }, |
484 | | { LLM_TENSOR_TIME_MIX_LERP_K, "blk.%d.time_mix_lerp_k" }, |
485 | | { LLM_TENSOR_TIME_MIX_LERP_V, "blk.%d.time_mix_lerp_v" }, |
486 | | { LLM_TENSOR_TIME_MIX_LERP_R, "blk.%d.time_mix_lerp_r" }, |
487 | | { LLM_TENSOR_TIME_MIX_LERP_G, "blk.%d.time_mix_lerp_g" }, |
488 | | { LLM_TENSOR_TIME_MIX_LERP_FUSED, "blk.%d.time_mix_lerp_fused" }, |
489 | | { LLM_TENSOR_TIME_MIX_FIRST, "blk.%d.time_mix_first" }, |
490 | | { LLM_TENSOR_TIME_MIX_DECAY, "blk.%d.time_mix_decay" }, |
491 | | { LLM_TENSOR_TIME_MIX_DECAY_W1, "blk.%d.time_mix_decay_w1" }, |
492 | | { LLM_TENSOR_TIME_MIX_DECAY_W2, "blk.%d.time_mix_decay_w2" }, |
493 | | { LLM_TENSOR_TIME_MIX_KEY, "blk.%d.time_mix_key" }, |
494 | | { LLM_TENSOR_TIME_MIX_VALUE, "blk.%d.time_mix_value" }, |
495 | | { LLM_TENSOR_TIME_MIX_RECEPTANCE, "blk.%d.time_mix_receptance" }, |
496 | | { LLM_TENSOR_TIME_MIX_GATE, "blk.%d.time_mix_gate" }, |
497 | | { LLM_TENSOR_TIME_MIX_LN, "blk.%d.time_mix_ln" }, |
498 | | { LLM_TENSOR_TIME_MIX_OUTPUT, "blk.%d.time_mix_output" }, |
499 | | { LLM_TENSOR_CHANNEL_MIX_LERP_K, "blk.%d.channel_mix_lerp_k" }, |
500 | | { LLM_TENSOR_CHANNEL_MIX_LERP_R, "blk.%d.channel_mix_lerp_r" }, |
501 | | { LLM_TENSOR_CHANNEL_MIX_KEY, "blk.%d.channel_mix_key" }, |
502 | | { LLM_TENSOR_CHANNEL_MIX_VALUE, "blk.%d.channel_mix_value" }, |
503 | | { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE, "blk.%d.channel_mix_receptance" }, |
504 | | { LLM_TENSOR_TIME_MIX_W0, "blk.%d.time_mix_w0" }, |
505 | | { LLM_TENSOR_TIME_MIX_A0, "blk.%d.time_mix_a0" }, |
506 | | { LLM_TENSOR_TIME_MIX_A1, "blk.%d.time_mix_a1" }, |
507 | | { LLM_TENSOR_TIME_MIX_A2, "blk.%d.time_mix_a2" }, |
508 | | { LLM_TENSOR_TIME_MIX_V0, "blk.%d.time_mix_v0" }, |
509 | | { LLM_TENSOR_TIME_MIX_V1, "blk.%d.time_mix_v1" }, |
510 | | { LLM_TENSOR_TIME_MIX_V2, "blk.%d.time_mix_v2" }, |
511 | | { LLM_TENSOR_TIME_MIX_G1, "blk.%d.time_mix_g1" }, |
512 | | { LLM_TENSOR_TIME_MIX_G2, "blk.%d.time_mix_g2" }, |
513 | | { LLM_TENSOR_TIME_MIX_K_K, "blk.%d.time_mix_k_k" }, |
514 | | { LLM_TENSOR_TIME_MIX_K_A, "blk.%d.time_mix_k_a" }, |
515 | | { LLM_TENSOR_TIME_MIX_R_K, "blk.%d.time_mix_r_k" }, |
516 | | { LLM_TENSOR_CONV1D, "conv1d" }, |
517 | | { LLM_TENSOR_CONVNEXT_DW, "convnext.%d.dw" }, |
518 | | { LLM_TENSOR_CONVNEXT_NORM, "convnext.%d.norm" }, |
519 | | { LLM_TENSOR_CONVNEXT_PW1, "convnext.%d.pw1" }, |
520 | | { LLM_TENSOR_CONVNEXT_PW2, "convnext.%d.pw2" }, |
521 | | { LLM_TENSOR_CONVNEXT_GAMMA, "convnext.%d.gamma" }, |
522 | | { LLM_TENSOR_POS_NET_CONV1, "posnet.%d.conv1" }, |
523 | | { LLM_TENSOR_POS_NET_CONV2, "posnet.%d.conv2" }, |
524 | | { LLM_TENSOR_POS_NET_NORM, "posnet.%d.norm" }, |
525 | | { LLM_TENSOR_POS_NET_NORM1, "posnet.%d.norm1" }, |
526 | | { LLM_TENSOR_POS_NET_NORM2, "posnet.%d.norm2" }, |
527 | | { LLM_TENSOR_POS_NET_ATTN_NORM, "posnet.%d.attn_norm" }, |
528 | | { LLM_TENSOR_POS_NET_ATTN_Q, "posnet.%d.attn_q" }, |
529 | | { LLM_TENSOR_POS_NET_ATTN_K, "posnet.%d.attn_k" }, |
530 | | { LLM_TENSOR_POS_NET_ATTN_V, "posnet.%d.attn_v" }, |
531 | | { LLM_TENSOR_POS_NET_ATTN_OUT, "posnet.%d.attn_output" }, |
532 | | { LLM_TENSOR_ATTN_SINKS, "blk.%d.attn_sinks" }, |
533 | | { LLM_TENSOR_SHORTCONV_CONV, "blk.%d.shortconv.conv" }, |
534 | | { LLM_TENSOR_SHORTCONV_INPROJ, "blk.%d.shortconv.in_proj" }, |
535 | | { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" }, |
536 | | { LLM_TENSOR_FFN_GATE_CHEXPS, "blk.%d.ffn_gate_chexps" }, |
537 | | { LLM_TENSOR_FFN_DOWN_CHEXPS, "blk.%d.ffn_down_chexps" }, |
538 | | { LLM_TENSOR_FFN_UP_CHEXPS, "blk.%d.ffn_up_chexps" }, |
539 | | { LLM_TENSOR_VISEXP_ATTN_QKV, "blk.%d.vis_attn_qkv" }, |
540 | | { LLM_TENSOR_VISEXP_ATTN_OUT, "blk.%d.vis_attn_output" }, |
541 | | { LLM_TENSOR_VISEXP_FFN_GATE, "blk.%d.vis_gate" }, |
542 | | { LLM_TENSOR_VISEXP_FFN_DOWN, "blk.%d.vis_down" }, |
543 | | { LLM_TENSOR_VISEXP_FFN_UP, "blk.%d.vis_up" }, |
544 | | { LLM_TENSOR_INDEXER_K_NORM, "blk.%d.indexer.k_norm" }, |
545 | | { LLM_TENSOR_INDEXER_PROJ, "blk.%d.indexer.proj" }, |
546 | | { LLM_TENSOR_INDEXER_ATTN_K, "blk.%d.indexer.attn_k" }, |
547 | | { LLM_TENSOR_INDEXER_ATTN_Q_B, "blk.%d.indexer.attn_q_b" }, |
548 | | }; |
549 | | |
550 | | // declare information about the model weight tensors: |
551 | | // - the layer in which the tensor is going to be used. this is needed in order to assign the correct buffer type for the weight |
552 | | // - the operator which is going to use the weight. this is needed to determine if the respective backend supports the operator |
553 | | // |
554 | | // for example, input layers are usually assigned to CPU/host buffer types |
555 | | // |
556 | | // a mismatch between the declared information and the actual layer/op in which the tensor is used can lead to sub-optimal |
557 | | // assignment of the buffer types and extra overhead during computation |
558 | | // example: https://github.com/ggml-org/llama.cpp/pull/17548 |
559 | | // |
560 | | static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = { |
561 | | {LLM_TENSOR_TOKEN_EMBD, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}}, |
562 | | {LLM_TENSOR_POS_EMBD, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}}, |
563 | | {LLM_TENSOR_TOKEN_TYPES, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}}, |
564 | | {LLM_TENSOR_TOKEN_EMBD_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // do the norms on the first layer (not the input layer) |
565 | | {LLM_TENSOR_OUTPUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, |
566 | | {LLM_TENSOR_CLS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, |
567 | | {LLM_TENSOR_CLS_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, |
568 | | {LLM_TENSOR_CLS_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, |
569 | | {LLM_TENSOR_DENSE_2_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output |
570 | | {LLM_TENSOR_DENSE_3_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output |
571 | | {LLM_TENSOR_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, |
572 | | {LLM_TENSOR_OUTPUT_NORM_LFM2, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, |
573 | | {LLM_TENSOR_DEC_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, |
574 | | {LLM_TENSOR_ENC_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, |
575 | | {LLM_TENSOR_ROPE_FREQS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ROPE}}, |
576 | | {LLM_TENSOR_ROPE_FACTORS_LONG, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ROPE}}, |
577 | | {LLM_TENSOR_ROPE_FACTORS_SHORT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ROPE}}, |
578 | | {LLM_TENSOR_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
579 | | {LLM_TENSOR_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
580 | | {LLM_TENSOR_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
581 | | {LLM_TENSOR_ATTN_QKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
582 | | {LLM_TENSOR_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
583 | | {LLM_TENSOR_ATTN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
584 | | {LLM_TENSOR_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
585 | | {LLM_TENSOR_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
586 | | {LLM_TENSOR_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
587 | | {LLM_TENSOR_FFN_DOWN_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
588 | | {LLM_TENSOR_FFN_GATE_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
589 | | {LLM_TENSOR_FFN_UP_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
590 | | {LLM_TENSOR_ATTN_Q_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
591 | | {LLM_TENSOR_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
592 | | {LLM_TENSOR_ATTN_KV_A_MQA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
593 | | {LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
594 | | {LLM_TENSOR_ATTN_K_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
595 | | {LLM_TENSOR_ATTN_V_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
596 | | {LLM_TENSOR_ATTN_SINKS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SCALE}}, |
597 | | {LLM_TENSOR_DEC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
598 | | {LLM_TENSOR_DEC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
599 | | {LLM_TENSOR_DEC_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
600 | | {LLM_TENSOR_DEC_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
601 | | {LLM_TENSOR_DEC_CROSS_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
602 | | {LLM_TENSOR_DEC_CROSS_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
603 | | {LLM_TENSOR_DEC_CROSS_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
604 | | {LLM_TENSOR_DEC_CROSS_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
605 | | {LLM_TENSOR_DEC_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
606 | | {LLM_TENSOR_DEC_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
607 | | {LLM_TENSOR_DEC_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
608 | | {LLM_TENSOR_ENC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
609 | | {LLM_TENSOR_ENC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
610 | | {LLM_TENSOR_ENC_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
611 | | {LLM_TENSOR_ENC_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
612 | | {LLM_TENSOR_ENC_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
613 | | {LLM_TENSOR_ENC_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
614 | | {LLM_TENSOR_ENC_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
615 | | {LLM_TENSOR_FFN_GATE_INP_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
616 | | {LLM_TENSOR_FFN_GATE_INP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
617 | | {LLM_TENSOR_SSM_IN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
618 | | {LLM_TENSOR_SSM_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
619 | | {LLM_TENSOR_SSM_DT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
620 | | {LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
621 | | {LLM_TENSOR_SSM_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
622 | | {LLM_TENSOR_SSM_BETA_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
623 | | {LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
624 | | {LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
625 | | {LLM_TENSOR_TIME_MIX_A1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
626 | | {LLM_TENSOR_TIME_MIX_A2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
627 | | {LLM_TENSOR_TIME_MIX_V1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
628 | | {LLM_TENSOR_TIME_MIX_V2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
629 | | {LLM_TENSOR_TIME_MIX_G1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
630 | | {LLM_TENSOR_TIME_MIX_G2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
631 | | {LLM_TENSOR_TIME_MIX_DECAY_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
632 | | {LLM_TENSOR_TIME_MIX_DECAY_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
633 | | {LLM_TENSOR_TIME_MIX_KEY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
634 | | {LLM_TENSOR_TIME_MIX_VALUE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
635 | | {LLM_TENSOR_TIME_MIX_RECEPTANCE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
636 | | {LLM_TENSOR_TIME_MIX_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
637 | | {LLM_TENSOR_TIME_MIX_OUTPUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
638 | | {LLM_TENSOR_CHANNEL_MIX_KEY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
639 | | {LLM_TENSOR_CHANNEL_MIX_RECEPTANCE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
640 | | {LLM_TENSOR_CHANNEL_MIX_VALUE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
641 | | {LLM_TENSOR_FFN_ACT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}}, |
642 | | {LLM_TENSOR_SSM_CONV1D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}}, |
643 | | {LLM_TENSOR_SSM_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}}, |
644 | | {LLM_TENSOR_SSM_A_NOSCAN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // a version of SSM_A used for MUL instead of SSM_SCAN |
645 | | {LLM_TENSOR_SSM_DT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
646 | | {LLM_TENSOR_SSM_B_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
647 | | {LLM_TENSOR_SSM_C_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
648 | | {LLM_TENSOR_SSM_D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
649 | | {LLM_TENSOR_SSM_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
650 | | // Kimi KDA - Conv tensors are 4D [d_conv, 1, d_inner, 1], reshaped to 2D at runtime |
651 | | {LLM_TENSOR_SSM_CONV1D_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
652 | | {LLM_TENSOR_SSM_CONV1D_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
653 | | {LLM_TENSOR_SSM_CONV1D_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
654 | | {LLM_TENSOR_SSM_F_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
655 | | {LLM_TENSOR_SSM_F_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
656 | | {LLM_TENSOR_SSM_BETA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
657 | | {LLM_TENSOR_SSM_G_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
658 | | {LLM_TENSOR_SSM_G_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
659 | | {LLM_TENSOR_TIME_MIX_LERP_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
660 | | {LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
661 | | {LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
662 | | {LLM_TENSOR_CHANNEL_MIX_LERP_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
663 | | {LLM_TENSOR_TIME_MIX_K_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
664 | | {LLM_TENSOR_TIME_MIX_K_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
665 | | {LLM_TENSOR_TIME_MIX_R_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
666 | | {LLM_TENSOR_TIME_MIX_LERP_W, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
667 | | {LLM_TENSOR_TIME_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
668 | | {LLM_TENSOR_TIME_MIX_LERP_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
669 | | {LLM_TENSOR_TIME_MIX_LERP_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
670 | | {LLM_TENSOR_TIME_MIX_LERP_G, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
671 | | {LLM_TENSOR_TIME_MIX_LERP_FUSED, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
672 | | {LLM_TENSOR_TIME_MIX_DECAY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
673 | | {LLM_TENSOR_TIME_MIX_W0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
674 | | {LLM_TENSOR_TIME_MIX_A0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
675 | | {LLM_TENSOR_TIME_MIX_V0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
676 | | {LLM_TENSOR_TIME_MIX_FIRST, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}}, |
677 | | {LLM_TENSOR_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
678 | | {LLM_TENSOR_ATTN_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
679 | | {LLM_TENSOR_ATTN_OUT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
680 | | {LLM_TENSOR_ATTN_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
681 | | {LLM_TENSOR_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
682 | | {LLM_TENSOR_FFN_PRE_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
683 | | {LLM_TENSOR_FFN_POST_NORM_1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
684 | | {LLM_TENSOR_FFN_POST_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
685 | | {LLM_TENSOR_FFN_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
686 | | {LLM_TENSOR_FFN_NORM_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
687 | | {LLM_TENSOR_ATTN_Q_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
688 | | {LLM_TENSOR_ATTN_K_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
689 | | {LLM_TENSOR_LAYER_OUT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
690 | | {LLM_TENSOR_LAYER_OUT_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
691 | | {LLM_TENSOR_ATTN_Q_A_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
692 | | {LLM_TENSOR_ATTN_KV_A_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
693 | | {LLM_TENSOR_ATTN_SUB_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
694 | | {LLM_TENSOR_FFN_SUB_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
695 | | {LLM_TENSOR_DEC_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
696 | | {LLM_TENSOR_DEC_CROSS_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
697 | | {LLM_TENSOR_DEC_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
698 | | {LLM_TENSOR_ENC_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
699 | | {LLM_TENSOR_ENC_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
700 | | {LLM_TENSOR_DEC_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}}, |
701 | | {LLM_TENSOR_ENC_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}}, |
702 | | {LLM_TENSOR_FFN_DOWN_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, |
703 | | {LLM_TENSOR_FFN_GATE_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, |
704 | | {LLM_TENSOR_FFN_UP_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, |
705 | | {LLM_TENSOR_FFN_GATE_UP_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, |
706 | | {LLM_TENSOR_FFN_DOWN_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, |
707 | | {LLM_TENSOR_FFN_GATE_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, |
708 | | {LLM_TENSOR_FFN_UP_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, |
709 | | {LLM_TENSOR_FFN_EXP_PROBS_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, |
710 | | // altup / laurel (gemma 3n) |
711 | | {LLM_TENSOR_PER_LAYER_TOKEN_EMBD, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}}, |
712 | | {LLM_TENSOR_PER_LAYER_MODEL_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
713 | | {LLM_TENSOR_PER_LAYER_PROJ_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
714 | | {LLM_TENSOR_ALTUP_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, |
715 | | {LLM_TENSOR_ALTUP_UNEMBD_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, |
716 | | {LLM_TENSOR_PER_LAYER_INP_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
717 | | {LLM_TENSOR_PER_LAYER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
718 | | {LLM_TENSOR_PER_LAYER_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
719 | | {LLM_TENSOR_ALTUP_CORRECT_COEF, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
720 | | {LLM_TENSOR_ALTUP_CORRECT_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
721 | | {LLM_TENSOR_ALTUP_PREDICT_COEF, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
722 | | {LLM_TENSOR_ALTUP_ROUTER, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
723 | | {LLM_TENSOR_ALTUP_ROUTER_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
724 | | {LLM_TENSOR_LAUREL_L, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
725 | | {LLM_TENSOR_LAUREL_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
726 | | {LLM_TENSOR_LAUREL_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
727 | | // this tensor is loaded for T5, but never used |
728 | | {LLM_TENSOR_DEC_CROSS_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}}, |
729 | | {LLM_TENSOR_CONV1D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}}, |
730 | | {LLM_TENSOR_POS_NET_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
731 | | {LLM_TENSOR_POS_NET_NORM1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
732 | | {LLM_TENSOR_POS_NET_NORM2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
733 | | {LLM_TENSOR_POS_NET_CONV1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}}, |
734 | | {LLM_TENSOR_POS_NET_CONV2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}}, |
735 | | {LLM_TENSOR_POS_NET_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
736 | | {LLM_TENSOR_POS_NET_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
737 | | {LLM_TENSOR_POS_NET_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
738 | | {LLM_TENSOR_POS_NET_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
739 | | {LLM_TENSOR_POS_NET_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
740 | | {LLM_TENSOR_CONVNEXT_DW, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}}, |
741 | | {LLM_TENSOR_CONVNEXT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
742 | | {LLM_TENSOR_CONVNEXT_PW1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
743 | | {LLM_TENSOR_CONVNEXT_PW2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
744 | | {LLM_TENSOR_CONVNEXT_GAMMA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
745 | | {LLM_TENSOR_SHORTCONV_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}}, |
746 | | {LLM_TENSOR_SHORTCONV_INPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
747 | | {LLM_TENSOR_SHORTCONV_OUTPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
748 | | {LLM_TENSOR_VISEXP_ATTN_QKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
749 | | {LLM_TENSOR_VISEXP_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
750 | | {LLM_TENSOR_VISEXP_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
751 | | {LLM_TENSOR_VISEXP_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
752 | | {LLM_TENSOR_VISEXP_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
753 | | {LLM_TENSOR_INDEXER_K_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
754 | | {LLM_TENSOR_INDEXER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
755 | | {LLM_TENSOR_INDEXER_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
756 | | {LLM_TENSOR_INDEXER_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, |
757 | | // NextN/MTP tensors are currently ignored (reserved for future MTP support) |
758 | | // These tensors only exist in the last layer(s) and are treated as output tensors |
759 | | {LLM_TENSOR_NEXTN_EH_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, |
760 | | {LLM_TENSOR_NEXTN_EMBED_TOKENS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}}, |
761 | | {LLM_TENSOR_NEXTN_ENORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}}, |
762 | | {LLM_TENSOR_NEXTN_HNORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, |
763 | | {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, |
764 | | {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, |
765 | | // Nemotron 3 Super |
766 | | {LLM_TENSOR_FFN_LATENT_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
767 | | {LLM_TENSOR_FFN_LATENT_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, |
768 | | }; |
769 | | |
770 | 0 | LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {} |
771 | | |
772 | 0 | std::string LLM_KV::operator()(llm_kv kv) const { |
773 | 0 | std::string name = ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch)); |
774 | |
|
775 | 0 | if (suffix != nullptr) { |
776 | 0 | name += "."; |
777 | 0 | name += suffix; |
778 | 0 | } |
779 | |
|
780 | 0 | return name; |
781 | 0 | } |
782 | | |
783 | | LLM_TN_IMPL::LLM_TN_IMPL(llm_arch arch, llm_tensor tensor, const char * suffix, int bid, int xid) |
784 | 0 | : arch(arch), tensor(tensor), suffix(suffix), bid(bid), xid(xid) {} |
785 | | |
786 | 0 | std::string LLM_TN_IMPL::str() const { |
787 | 0 | if (LLM_TENSOR_NAMES.find(tensor) == LLM_TENSOR_NAMES.end()) { |
788 | 0 | GGML_ABORT("unknown tensor name for tensor id %d", static_cast<int>(tensor)); |
789 | 0 | } |
790 | |
|
791 | 0 | std::string name = ::format(LLM_TENSOR_NAMES.at(tensor), bid, xid); |
792 | 0 | if (suffix != nullptr) { |
793 | 0 | name += "."; |
794 | 0 | name += suffix; |
795 | 0 | } |
796 | |
|
797 | 0 | return name; |
798 | 0 | } |
799 | | |
800 | 0 | std::vector<llm_arch> llm_arch_all() { |
801 | 0 | std::vector<llm_arch> ret; |
802 | 0 | ret.reserve(LLM_ARCH_NAMES.size()); |
803 | 0 | for (const auto & [arch, _] : LLM_ARCH_NAMES) { |
804 | 0 | ret.push_back(arch); |
805 | 0 | } |
806 | 0 | return ret; |
807 | 0 | } |
808 | | |
809 | 0 | const char * llm_arch_name(llm_arch arch) { |
810 | 0 | auto it = LLM_ARCH_NAMES.find(arch); |
811 | 0 | if (it == LLM_ARCH_NAMES.end()) { |
812 | 0 | return "unknown"; |
813 | 0 | } |
814 | 0 | return it->second; |
815 | 0 | } |
816 | | |
817 | 0 | llm_arch llm_arch_from_string(const std::string & name) { |
818 | 0 | for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT |
819 | 0 | if (kv.second == name) { |
820 | 0 | return kv.first; |
821 | 0 | } |
822 | 0 | } |
823 | | |
824 | 0 | return LLM_ARCH_UNKNOWN; |
825 | 0 | } |
826 | | |
827 | 0 | const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) { |
828 | 0 | return LLM_TENSOR_INFOS.at(tensor); |
829 | 0 | } |
830 | | |
831 | 0 | bool llm_arch_is_recurrent(const llm_arch & arch) { |
832 | 0 | switch (arch) { |
833 | 0 | case LLM_ARCH_MAMBA: |
834 | 0 | case LLM_ARCH_MAMBA2: |
835 | 0 | case LLM_ARCH_RWKV6: |
836 | 0 | case LLM_ARCH_RWKV6QWEN2: |
837 | 0 | case LLM_ARCH_RWKV7: |
838 | 0 | case LLM_ARCH_ARWKV7: |
839 | 0 | return true; |
840 | 0 | default: |
841 | 0 | return false; |
842 | 0 | } |
843 | 0 | } |
844 | | |
845 | 0 | bool llm_arch_is_hybrid(const llm_arch & arch) { |
846 | 0 | switch (arch) { |
847 | 0 | case LLM_ARCH_JAMBA: |
848 | 0 | case LLM_ARCH_FALCON_H1: |
849 | 0 | case LLM_ARCH_PLAMO2: |
850 | 0 | case LLM_ARCH_GRANITE_HYBRID: |
851 | 0 | case LLM_ARCH_LFM2: |
852 | 0 | case LLM_ARCH_LFM2MOE: |
853 | 0 | case LLM_ARCH_NEMOTRON_H: |
854 | 0 | case LLM_ARCH_NEMOTRON_H_MOE: |
855 | 0 | case LLM_ARCH_QWEN3NEXT: |
856 | 0 | case LLM_ARCH_KIMI_LINEAR: |
857 | 0 | case LLM_ARCH_QWEN35: |
858 | 0 | case LLM_ARCH_QWEN35MOE: |
859 | 0 | return true; |
860 | 0 | default: |
861 | 0 | return false; |
862 | 0 | } |
863 | 0 | } |
864 | | |
865 | 0 | bool llm_arch_is_diffusion(const llm_arch & arch) { |
866 | 0 | switch (arch) { |
867 | 0 | case LLM_ARCH_DREAM: |
868 | 0 | case LLM_ARCH_LLADA: |
869 | 0 | case LLM_ARCH_LLADA_MOE: |
870 | 0 | case LLM_ARCH_RND1: |
871 | 0 | return true; |
872 | 0 | default: |
873 | 0 | return false; |
874 | 0 | } |
875 | 0 | } |
876 | | |
877 | 0 | bool llm_arch_supports_sm_tensor(const llm_arch & arch) { |
878 | 0 | switch (arch) { |
879 | 0 | case LLM_ARCH_GROK: |
880 | 0 | case LLM_ARCH_MPT: |
881 | 0 | case LLM_ARCH_PLAMO2: |
882 | 0 | case LLM_ARCH_MINICPM3: |
883 | 0 | case LLM_ARCH_GEMMA3N: |
884 | 0 | case LLM_ARCH_MAMBA: |
885 | 0 | case LLM_ARCH_MAMBA2: |
886 | 0 | case LLM_ARCH_JAMBA: |
887 | 0 | case LLM_ARCH_FALCON_H1: |
888 | 0 | case LLM_ARCH_OLMO2: |
889 | 0 | case LLM_ARCH_OLMOE: |
890 | 0 | case LLM_ARCH_DEEPSEEK2: |
891 | 0 | case LLM_ARCH_GLM_DSA: |
892 | 0 | case LLM_ARCH_BITNET: |
893 | 0 | case LLM_ARCH_T5: |
894 | 0 | case LLM_ARCH_NEMOTRON_H: |
895 | 0 | case LLM_ARCH_NEMOTRON_H_MOE: |
896 | 0 | case LLM_ARCH_GRANITE_HYBRID: |
897 | 0 | case LLM_ARCH_LFM2: |
898 | 0 | case LLM_ARCH_LFM2MOE: |
899 | 0 | case LLM_ARCH_MINIMAX_M2: |
900 | 0 | case LLM_ARCH_MISTRAL4: |
901 | 0 | case LLM_ARCH_KIMI_LINEAR: |
902 | 0 | return false; |
903 | 0 | default: |
904 | 0 | return true; |
905 | 0 | } |
906 | 0 | } |