Orion-zhen commited on
Commit
2c6668c
·
verified ·
1 Parent(s): 007472b

Upload 2 files

Browse files
models/DeepSeek-V3/config.json CHANGED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "DeepseekV3ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_deepseek.DeepseekV3Config",
9
+ "AutoModel": "modeling_deepseek.DeepseekV3Model",
10
+ "AutoModelForCausalLM": "modeling_deepseek.DeepseekV3ForCausalLM"
11
+ },
12
+ "bos_token_id": 0,
13
+ "eos_token_id": 1,
14
+ "ep_size": 1,
15
+ "first_k_dense_replace": 3,
16
+ "hidden_act": "silu",
17
+ "hidden_size": 7168,
18
+ "initializer_range": 0.02,
19
+ "intermediate_size": 18432,
20
+ "kv_lora_rank": 512,
21
+ "max_position_embeddings": 163840,
22
+ "model_type": "deepseek_v3",
23
+ "moe_intermediate_size": 2048,
24
+ "moe_layer_freq": 1,
25
+ "n_group": 8,
26
+ "n_routed_experts": 256,
27
+ "n_shared_experts": 1,
28
+ "norm_topk_prob": true,
29
+ "num_attention_heads": 128,
30
+ "num_experts_per_tok": 8,
31
+ "num_hidden_layers": 61,
32
+ "num_key_value_heads": 128,
33
+ "num_nextn_predict_layers": 1,
34
+ "q_lora_rank": 1536,
35
+ "qk_nope_head_dim": 128,
36
+ "qk_rope_head_dim": 64,
37
+ "quantization_config": {
38
+ "activation_scheme": "dynamic",
39
+ "fmt": "e4m3",
40
+ "quant_method": "fp8",
41
+ "weight_block_size": [
42
+ 128,
43
+ 128
44
+ ]
45
+ },
46
+ "rms_norm_eps": 1e-06,
47
+ "rope_scaling": {
48
+ "beta_fast": 32,
49
+ "beta_slow": 1,
50
+ "factor": 40,
51
+ "mscale": 1.0,
52
+ "mscale_all_dim": 1.0,
53
+ "original_max_position_embeddings": 4096,
54
+ "type": "yarn"
55
+ },
56
+ "rope_theta": 10000,
57
+ "routed_scaling_factor": 2.5,
58
+ "scoring_func": "sigmoid",
59
+ "tie_word_embeddings": false,
60
+ "topk_group": 4,
61
+ "topk_method": "noaux_tc",
62
+ "torch_dtype": "bfloat16",
63
+ "transformers_version": "4.46.3",
64
+ "use_cache": true,
65
+ "v_head_dim": 128,
66
+ "vocab_size": 129280
67
+ }
models/DeepSeek-V3/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff