Skip to content

Commit 335a1f9

Browse files
ikawrakowIwan Kawrakow
ikawrakow
and
Iwan Kawrakow
authored
Refactor file llama.cpp (ikawrakow#823)
* llama_model and llama_hparams * llama_build_context Surprisingly small reduction in llama.cpp compile time given the reduction in LOCs (22k -> 14k) * LLM_TN llama.cpp compilation: 50 s -> 33 s * llama_quantize * arch names * All graph building is now in llm-build-context.cpp * hparams loading llama.cpp is now just 9300 LOC, but still takes 32 seconds to compile. * We are now at 6 seconds to build the src folder * load -> create We are not actually loading the tensors, but just creating them. --------- Co-authored-by: Iwan Kawrakow <[email protected]>
1 parent f649e36 commit 335a1f9

16 files changed

+18129
-17594
lines changed

‎src/CMakeLists.txt‎

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -19,6 +19,16 @@ add_library(llama
1919
llama-sampling.cpp
2020
llama-mmap.cpp
2121
llama-model-loader.cpp
22+
llama-load-tensors.cpp
23+
llama-build-context.h
24+
llama-build-context.cpp
25+
llama-model.h
26+
llama-model.cpp
27+
llama-quantize.cpp
28+
llama-arch.h
29+
llama-arch.cpp
30+
llama-hparams.h
31+
llama-hparams.cpp
2232
unicode.h
2333
unicode.cpp
2434
unicode-data.cpp

‎src/llama-arch.cpp‎

Lines changed: 217 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,217 @@
1+
#include "llama-arch.h"
2+
#include "llama-impl.h"
3+
4+
#include <map>
5+
6+
static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
7+
{ LLM_ARCH_LLAMA, "llama" },
8+
{ LLM_ARCH_LLAMA4, "llama4" },
9+
{ LLM_ARCH_DECI, "deci" },
10+
{ LLM_ARCH_FALCON, "falcon" },
11+
{ LLM_ARCH_GROK, "grok" },
12+
{ LLM_ARCH_GPT2, "gpt2" },
13+
{ LLM_ARCH_GPTJ, "gptj" },
14+
{ LLM_ARCH_GPTNEOX, "gptneox" },
15+
{ LLM_ARCH_MPT, "mpt" },
16+
{ LLM_ARCH_BAICHUAN, "baichuan" },
17+
{ LLM_ARCH_STARCODER, "starcoder" },
18+
{ LLM_ARCH_REFACT, "refact" },
19+
{ LLM_ARCH_BERT, "bert" },
20+
{ LLM_ARCH_NOMIC_BERT, "nomic-bert" },
21+
{ LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" },
22+
{ LLM_ARCH_BLOOM, "bloom" },
23+
{ LLM_ARCH_STABLELM, "stablelm" },
24+
{ LLM_ARCH_QWEN, "qwen" },
25+
{ LLM_ARCH_QWEN2, "qwen2" },
26+
{ LLM_ARCH_QWEN2MOE, "qwen2moe" },
27+
{ LLM_ARCH_QWEN2VL, "qwen2vl" },
28+
{ LLM_ARCH_QWEN3, "qwen3" },
29+
{ LLM_ARCH_QWEN3MOE, "qwen3moe" },
30+
{ LLM_ARCH_PHI2, "phi2" },
31+
{ LLM_ARCH_PHI3, "phi3" },
32+
{ LLM_ARCH_PLAMO, "plamo" },
33+
{ LLM_ARCH_CODESHELL, "codeshell" },
34+
{ LLM_ARCH_ORION, "orion" },
35+
{ LLM_ARCH_INTERNLM2, "internlm2" },
36+
{ LLM_ARCH_MINICPM, "minicpm" },
37+
{ LLM_ARCH_GEMMA, "gemma" },
38+
{ LLM_ARCH_GEMMA2, "gemma2" },
39+
{ LLM_ARCH_GEMMA3, "gemma3" },
40+
{ LLM_ARCH_STARCODER2, "starcoder2" },
41+
{ LLM_ARCH_MAMBA, "mamba" },
42+
{ LLM_ARCH_XVERSE, "xverse" },
43+
{ LLM_ARCH_COMMAND_R, "command-r" },
44+
{ LLM_ARCH_DBRX, "dbrx" },
45+
{ LLM_ARCH_OLMO, "olmo" },
46+
{ LLM_ARCH_OPENELM, "openelm" },
47+
{ LLM_ARCH_ARCTIC, "arctic" },
48+
{ LLM_ARCH_DEEPSEEK2, "deepseek2" },
49+
{ LLM_ARCH_CHATGLM, "chatglm" },
50+
{ LLM_ARCH_GLM4, "glm4" },
51+
{ LLM_ARCH_GLM4_MOE, "glm4moe" },
52+
{ LLM_ARCH_BITNET, "bitnet" },
53+
{ LLM_ARCH_BITNET_25, "bitnet-25" },
54+
{ LLM_ARCH_BITNET_B158, "bitnet-b1.58" },
55+
{ LLM_ARCH_T5, "t5" },
56+
{ LLM_ARCH_T5ENCODER, "t5encoder" },
57+
{ LLM_ARCH_JAIS, "jais" },
58+
{ LLM_ARCH_GRANITE, "granite" },
59+
{ LLM_ARCH_GRANITE_MOE, "granitemoe" },
60+
{ LLM_ARCH_COHERE2, "cohere2" },
61+
{ LLM_ARCH_DOTS1, "dots1" },
62+
{ LLM_ARCH_ERNIE4_5, "ernie4_5" },
63+
{ LLM_ARCH_ERNIE4_5_MOE, "ernie4_5-moe" },
64+
{ LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" },
65+
{ LLM_ARCH_OPENAI_MOE, "gpt-oss" },
66+
{ LLM_ARCH_UNKNOWN, "(unknown)" },
67+
};
68+
69+
llm_arch llm_arch_from_string(const std::string & name) {
70+
for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT
71+
if (kv.second == name) {
72+
return kv.first;
73+
}
74+
}
75+
76+
return LLM_ARCH_UNKNOWN;
77+
}
78+
79+
static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
80+
{ LLM_KV_GENERAL_TYPE, "general.type" },
81+
{ LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" },
82+
{ LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" },
83+
{ LLM_KV_GENERAL_ALIGNMENT, "general.alignment" },
84+
{ LLM_KV_GENERAL_NAME, "general.name" },
85+
{ LLM_KV_GENERAL_AUTHOR, "general.author" },
86+
{ LLM_KV_GENERAL_VERSION, "general.version" },
87+
{ LLM_KV_GENERAL_URL, "general.url" },
88+
{ LLM_KV_GENERAL_DESCRIPTION, "general.description" },
89+
{ LLM_KV_GENERAL_LICENSE, "general.license" },
90+
{ LLM_KV_GENERAL_SOURCE_URL, "general.source.url" },
91+
{ LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" },
92+
93+
{ LLM_KV_VOCAB_SIZE, "%s.vocab_size" },
94+
{ LLM_KV_CONTEXT_LENGTH, "%s.context_length" },
95+
{ LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" },
96+
{ LLM_KV_BLOCK_COUNT, "%s.block_count" },
97+
{ LLM_KV_LEADING_DENSE_BLOCK_COUNT, "%s.leading_dense_block_count" },
98+
{ LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" },
99+
{ LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" },
100+
{ LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
101+
{ LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" },
102+
{ LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" },
103+
{ LLM_KV_EXPERT_COUNT, "%s.expert_count" },
104+
{ LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" },
105+
{ LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" },
106+
{ LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" },
107+
{ LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" },
108+
{ LLM_KV_EXPERT_GATING_FUNC, "%s.expert_gating_func" },
109+
{ LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" },
110+
{ LLM_KV_POOLING_TYPE , "%s.pooling_type" },
111+
{ LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
112+
{ LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" },
113+
{ LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" },
114+
{ LLM_KV_ROUTER_LOGIT_SOFTCAPPING, "%s.router_logit_softcapping" },
115+
{ LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" },
116+
{ LLM_KV_RESIDUAL_SCALE, "%s.residual_scale" },
117+
{ LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" },
118+
{ LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" },
119+
{ LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" },
120+
121+
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
122+
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
123+
{ LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" },
124+
{ LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" },
125+
{ LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" },
126+
{ LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" },
127+
{ LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" },
128+
{ LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" },
129+
{ LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" },
130+
{ LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" },
131+
{ LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" },
132+
{ LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
133+
{ LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
134+
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
135+
{ LLM_KV_ATTENTION_OUTPUT_SCALE, "%s.attention.output_scale" },
136+
{ LLM_KV_ATTENTION_TEMPERATURE_LENGTH, "%s.attention.temperature_length" },
137+
138+
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
139+
{ LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
140+
{ LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
141+
{ LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
142+
{ LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
143+
{ LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
144+
{ LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
145+
{ LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
146+
{ LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
147+
{ LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" },
148+
{ LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, "%s.rope.scaling.yarn_ext_factor" },
149+
{ LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor" },
150+
{ LLM_KV_ROPE_SCALING_YARN_BETA_FAST, "%s.rope.scaling.yarn_beta_fast" },
151+
{ LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, "%s.rope.scaling.yarn_beta_slow" },
152+
153+
{ LLM_KV_SPLIT_NO, "split.no" },
154+
{ LLM_KV_SPLIT_COUNT, "split.count" },
155+
{ LLM_KV_SPLIT_TENSORS_COUNT, "split.tensors.count" },
156+
157+
{ LLM_KV_SSM_CONV_KERNEL, "%s.ssm.conv_kernel" },
158+
{ LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" },
159+
{ LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" },
160+
{ LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" },
161+
162+
{ LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" },
163+
{ LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" },
164+
{ LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" },
165+
{ LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" },
166+
{ LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, "tokenizer.ggml.token_type_count" },
167+
{ LLM_KV_TOKENIZER_SCORES, "tokenizer.ggml.scores" },
168+
{ LLM_KV_TOKENIZER_MERGES, "tokenizer.ggml.merges" },
169+
{ LLM_KV_TOKENIZER_BOS_ID, "tokenizer.ggml.bos_token_id" },
170+
{ LLM_KV_TOKENIZER_EOS_ID, "tokenizer.ggml.eos_token_id" },
171+
{ LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" },
172+
{ LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" },
173+
{ LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" },
174+
{ LLM_KV_TOKENIZER_CLS_ID, "tokenizer.ggml.cls_token_id" },
175+
{ LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_token_id" },
176+
{ LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" },
177+
{ LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" },
178+
{ LLM_KV_TOKENIZER_ADD_SEP, "tokenizer.ggml.add_sep_token" },
179+
{ LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" },
180+
{ LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, "tokenizer.ggml.remove_extra_whitespaces" },
181+
{ LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap" },
182+
{ LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" },
183+
{ LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" },
184+
{ LLM_KV_TOKENIZER_CHAT_TEMPLATE, "tokenizer.chat_template" },
185+
{ LLM_KV_TOKENIZER_CHAT_TEMPLATE_N, "tokenizer.chat_template.%s" },
186+
{ LLM_KV_TOKENIZER_FIM_PRE_ID, "tokenizer.ggml.fim_pre_token_id" },
187+
{ LLM_KV_TOKENIZER_FIM_SUF_ID, "tokenizer.ggml.fim_suf_token_id" },
188+
{ LLM_KV_TOKENIZER_FIM_MID_ID, "tokenizer.ggml.fim_mid_token_id" },
189+
{ LLM_KV_TOKENIZER_FIM_PAD_ID, "tokenizer.ggml.fim_pad_token_id" },
190+
{ LLM_KV_TOKENIZER_FIM_REP_ID, "tokenizer.ggml.fim_rep_token_id" },
191+
{ LLM_KV_TOKENIZER_FIM_SEP_ID, "tokenizer.ggml.fim_sep_token_id" },
192+
193+
{ LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" },
194+
{ LLM_KV_TOKENIZER_SUFFIX_ID, "tokenizer.ggml.suffix_token_id" },
195+
{ LLM_KV_TOKENIZER_MIDDLE_ID, "tokenizer.ggml.middle_token_id" },
196+
{ LLM_KV_TOKENIZER_EOT_ID, "tokenizer.ggml.eot_token_id" },
197+
{ LLM_KV_TOKENIZER_EOM_ID, "tokenizer.ggml.eom_token_id" },
198+
199+
{ LLM_KV_ADAPTER_TYPE, "adapter.type" },
200+
{ LLM_KV_ADAPTER_LORA_ALPHA, "adapter.lora.alpha" },
201+
};
202+
203+
LLM_KV::LLM_KV(llm_arch arch, const char* suffix) : arch(arch), suffix(suffix) {}
204+
205+
std::string LLM_KV::operator()(llm_kv kv) const {
206+
return suffix ? ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch), suffix)
207+
: ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
208+
}
209+
210+
const char * llama_model_arch_name(llm_arch arch) {
211+
auto it = LLM_ARCH_NAMES.find(arch);
212+
if (it == LLM_ARCH_NAMES.end()) {
213+
return "unknown";
214+
}
215+
return it->second;
216+
}
217+

‎src/llama-arch.h‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -298,3 +298,5 @@ enum llm_tensor {
298298
};
299299

300300
llm_arch llm_arch_from_string(const std::string & name);
301+
302+
const char * llama_model_arch_name(llm_arch arch);

0 commit comments

Comments
 (0)