Hugging Face's logo

inference-optimization
/
DeepSeek-V4-Flash-bf16-dequantized-5layers

Safetensors
deepseek_v4
Model card Files
xet
 Community
kylesayrs commited on
Commit
34c667c
·
verified ·
1 Parent(s): 90ffbfe

Upload folder using huggingface_hub

Browse files
Files changed (6) hide show
  1. config.json +142 -0
  2. generation_config.json +9 -0
  3. model.safetensors +3 -0
  4. model.safetensors.index.json +157 -0
  5. tokenizer.json +0 -0
  6. tokenizer_config.json +14 -0
config.json ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "DeepseekV4ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 0,
8
+ "compress_ratios": [
9
+ 0,
10
+ 0,
11
+ 4,
12
+ 128,
13
+ 4
14
+ ],
15
+ "compress_rope_theta": 160000.0,
16
+ "eos_token_id": 1,
17
+ "hc_eps": 1e-06,
18
+ "hc_mult": 4,
19
+ "hc_sinkhorn_iters": 20,
20
+ "head_dim": 512,
21
+ "hidden_act": "silu",
22
+ "hidden_size": 4096,
23
+ "index_head_dim": 128,
24
+ "index_n_heads": 64,
25
+ "index_topk": 512,
26
+ "initializer_range": 0.02,
27
+ "max_position_embeddings": 1048576,
28
+ "model_type": "deepseek_v4",
29
+ "moe_intermediate_size": 2048,
30
+ "n_routed_experts": 256,
31
+ "n_shared_experts": 1,
32
+ "norm_topk_prob": true,
33
+ "num_attention_heads": 64,
34
+ "num_experts_per_tok": 6,
35
+ "num_hash_layers": 3,
36
+ "num_hidden_layers": 5,
37
+ "num_key_value_heads": 1,
38
+ "num_nextn_predict_layers": 0,
39
+ "o_groups": 8,
40
+ "o_lora_rank": 1024,
41
+ "q_lora_rank": 1024,
42
+ "qk_rope_head_dim": 64,
43
+ "quantization": {
44
+ "group_size": 32,
45
+ "bits": 4,
46
+ "mode": "mxfp4",
47
+ "model.layers.0.ffn.switch_mlp.gate_proj": {
48
+ "group_size": 32,
49
+ "bits": 4,
50
+ "mode": "mxfp4"
51
+ },
52
+ "model.layers.0.ffn.switch_mlp.up_proj": {
53
+ "group_size": 32,
54
+ "bits": 4,
55
+ "mode": "mxfp4"
56
+ },
57
+ "model.layers.0.ffn.switch_mlp.down_proj": {
58
+ "group_size": 32,
59
+ "bits": 4,
60
+ "mode": "mxfp4"
61
+ },
62
+ "model.layers.1.ffn.switch_mlp.gate_proj": {
63
+ "group_size": 32,
64
+ "bits": 4,
65
+ "mode": "mxfp4"
66
+ },
67
+ "model.layers.1.ffn.switch_mlp.up_proj": {
68
+ "group_size": 32,
69
+ "bits": 4,
70
+ "mode": "mxfp4"
71
+ },
72
+ "model.layers.1.ffn.switch_mlp.down_proj": {
73
+ "group_size": 32,
74
+ "bits": 4,
75
+ "mode": "mxfp4"
76
+ },
77
+ "model.layers.2.ffn.switch_mlp.gate_proj": {
78
+ "group_size": 32,
79
+ "bits": 4,
80
+ "mode": "mxfp4"
81
+ },
82
+ "model.layers.2.ffn.switch_mlp.up_proj": {
83
+ "group_size": 32,
84
+ "bits": 4,
85
+ "mode": "mxfp4"
86
+ },
87
+ "model.layers.2.ffn.switch_mlp.down_proj": {
88
+ "group_size": 32,
89
+ "bits": 4,
90
+ "mode": "mxfp4"
91
+ },
92
+ "model.layers.3.ffn.switch_mlp.gate_proj": {
93
+ "group_size": 32,
94
+ "bits": 4,
95
+ "mode": "mxfp4"
96
+ },
97
+ "model.layers.3.ffn.switch_mlp.up_proj": {
98
+ "group_size": 32,
99
+ "bits": 4,
100
+ "mode": "mxfp4"
101
+ },
102
+ "model.layers.3.ffn.switch_mlp.down_proj": {
103
+ "group_size": 32,
104
+ "bits": 4,
105
+ "mode": "mxfp4"
106
+ },
107
+ "model.layers.4.ffn.switch_mlp.gate_proj": {
108
+ "group_size": 32,
109
+ "bits": 4,
110
+ "mode": "mxfp4"
111
+ },
112
+ "model.layers.4.ffn.switch_mlp.up_proj": {
113
+ "group_size": 32,
114
+ "bits": 4,
115
+ "mode": "mxfp4"
116
+ },
117
+ "model.layers.4.ffn.switch_mlp.down_proj": {
118
+ "group_size": 32,
119
+ "bits": 4,
120
+ "mode": "mxfp4"
121
+ }
122
+ },
123
+ "rms_norm_eps": 1e-06,
124
+ "rope_scaling": {
125
+ "beta_fast": 32,
126
+ "beta_slow": 1,
127
+ "factor": 16,
128
+ "original_max_position_embeddings": 65536,
129
+ "type": "yarn"
130
+ },
131
+ "rope_theta": 10000.0,
132
+ "routed_scaling_factor": 1.5,
133
+ "scoring_func": "sqrtsoftplus",
134
+ "sliding_window": 128,
135
+ "swiglu_limit": 10.0,
136
+ "tie_word_embeddings": false,
137
+ "topk_method": "noaux_tc",
138
+ "torch_dtype": "bfloat16",
139
+ "transformers_version": "4.57.1",
140
+ "use_cache": true,
141
+ "vocab_size": 129280
142
+ }
generation_config.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 0,
4
+ "eos_token_id": 1,
5
+ "do_sample": true,
6
+ "temperature": 1.0,
7
+ "top_p": 1.0,
8
+ "transformers_version": "4.46.3"
9
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6aef629b91b2830c6cdcc53310a06ff535e1e37d3600339083fe9a43106b997e
3
+ size 67994148004
model.safetensors.index.json ADDED
@@ -0,0 +1,157 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 67994131276
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "model.safetensors",
7
+ "model.embed_tokens.weight": "model.safetensors",
8
+ "model.hc_head.base": "model.safetensors",
9
+ "model.hc_head.fn": "model.safetensors",
10
+ "model.hc_head.scale": "model.safetensors",
11
+ "model.layers.0.attn.attn_sink": "model.safetensors",
12
+ "model.layers.0.attn.kv_norm.weight": "model.safetensors",
13
+ "model.layers.0.attn.q_norm.weight": "model.safetensors",
14
+ "model.layers.0.attn.wkv.weight": "model.safetensors",
15
+ "model.layers.0.attn.wo_a.weight": "model.safetensors",
16
+ "model.layers.0.attn.wo_b.weight": "model.safetensors",
17
+ "model.layers.0.attn.wq_a.weight": "model.safetensors",
18
+ "model.layers.0.attn.wq_b.weight": "model.safetensors",
19
+ "model.layers.0.attn_hc.base": "model.safetensors",
20
+ "model.layers.0.attn_hc.fn": "model.safetensors",
21
+ "model.layers.0.attn_hc.scale": "model.safetensors",
22
+ "model.layers.0.attn_norm.weight": "model.safetensors",
23
+ "model.layers.0.ffn.gate.tid2eid": "model.safetensors",
24
+ "model.layers.0.ffn.gate.weight": "model.safetensors",
25
+ "model.layers.0.ffn.shared_experts.down_proj.weight": "model.safetensors",
26
+ "model.layers.0.ffn.shared_experts.gate_proj.weight": "model.safetensors",
27
+ "model.layers.0.ffn.shared_experts.up_proj.weight": "model.safetensors",
28
+ "model.layers.0.ffn.switch_mlp.down_proj.weight": "model.safetensors",
29
+ "model.layers.0.ffn.switch_mlp.gate_proj.weight": "model.safetensors",
30
+ "model.layers.0.ffn.switch_mlp.up_proj.weight": "model.safetensors",
31
+ "model.layers.0.ffn_hc.base": "model.safetensors",
32
+ "model.layers.0.ffn_hc.fn": "model.safetensors",
33
+ "model.layers.0.ffn_hc.scale": "model.safetensors",
34
+ "model.layers.0.ffn_norm.weight": "model.safetensors",
35
+ "model.layers.1.attn.attn_sink": "model.safetensors",
36
+ "model.layers.1.attn.kv_norm.weight": "model.safetensors",
37
+ "model.layers.1.attn.q_norm.weight": "model.safetensors",
38
+ "model.layers.1.attn.wkv.weight": "model.safetensors",
39
+ "model.layers.1.attn.wo_a.weight": "model.safetensors",
40
+ "model.layers.1.attn.wo_b.weight": "model.safetensors",
41
+ "model.layers.1.attn.wq_a.weight": "model.safetensors",
42
+ "model.layers.1.attn.wq_b.weight": "model.safetensors",
43
+ "model.layers.1.attn_hc.base": "model.safetensors",
44
+ "model.layers.1.attn_hc.fn": "model.safetensors",
45
+ "model.layers.1.attn_hc.scale": "model.safetensors",
46
+ "model.layers.1.attn_norm.weight": "model.safetensors",
47
+ "model.layers.1.ffn.gate.tid2eid": "model.safetensors",
48
+ "model.layers.1.ffn.gate.weight": "model.safetensors",
49
+ "model.layers.1.ffn.shared_experts.down_proj.weight": "model.safetensors",
50
+ "model.layers.1.ffn.shared_experts.gate_proj.weight": "model.safetensors",
51
+ "model.layers.1.ffn.shared_experts.up_proj.weight": "model.safetensors",
52
+ "model.layers.1.ffn.switch_mlp.down_proj.weight": "model.safetensors",
53
+ "model.layers.1.ffn.switch_mlp.gate_proj.weight": "model.safetensors",
54
+ "model.layers.1.ffn.switch_mlp.up_proj.weight": "model.safetensors",
55
+ "model.layers.1.ffn_hc.base": "model.safetensors",
56
+ "model.layers.1.ffn_hc.fn": "model.safetensors",
57
+ "model.layers.1.ffn_hc.scale": "model.safetensors",
58
+ "model.layers.1.ffn_norm.weight": "model.safetensors",
59
+ "model.layers.2.attn.attn_sink": "model.safetensors",
60
+ "model.layers.2.attn.compressor.ape": "model.safetensors",
61
+ "model.layers.2.attn.compressor.norm.weight": "model.safetensors",
62
+ "model.layers.2.attn.compressor.wgate.weight": "model.safetensors",
63
+ "model.layers.2.attn.compressor.wkv.weight": "model.safetensors",
64
+ "model.layers.2.attn.indexer.compressor.ape": "model.safetensors",
65
+ "model.layers.2.attn.indexer.compressor.norm.weight": "model.safetensors",
66
+ "model.layers.2.attn.indexer.compressor.wgate.weight": "model.safetensors",
67
+ "model.layers.2.attn.indexer.compressor.wkv.weight": "model.safetensors",
68
+ "model.layers.2.attn.indexer.weights_proj.weight": "model.safetensors",
69
+ "model.layers.2.attn.indexer.wq_b.weight": "model.safetensors",
70
+ "model.layers.2.attn.kv_norm.weight": "model.safetensors",
71
+ "model.layers.2.attn.q_norm.weight": "model.safetensors",
72
+ "model.layers.2.attn.wkv.weight": "model.safetensors",
73
+ "model.layers.2.attn.wo_a.weight": "model.safetensors",
74
+ "model.layers.2.attn.wo_b.weight": "model.safetensors",
75
+ "model.layers.2.attn.wq_a.weight": "model.safetensors",
76
+ "model.layers.2.attn.wq_b.weight": "model.safetensors",
77
+ "model.layers.2.attn_hc.base": "model.safetensors",
78
+ "model.layers.2.attn_hc.fn": "model.safetensors",
79
+ "model.layers.2.attn_hc.scale": "model.safetensors",
80
+ "model.layers.2.attn_norm.weight": "model.safetensors",
81
+ "model.layers.2.ffn.gate.tid2eid": "model.safetensors",
82
+ "model.layers.2.ffn.gate.weight": "model.safetensors",
83
+ "model.layers.2.ffn.shared_experts.down_proj.weight": "model.safetensors",
84
+ "model.layers.2.ffn.shared_experts.gate_proj.weight": "model.safetensors",
85
+ "model.layers.2.ffn.shared_experts.up_proj.weight": "model.safetensors",
86
+ "model.layers.2.ffn.switch_mlp.down_proj.weight": "model.safetensors",
87
+ "model.layers.2.ffn.switch_mlp.gate_proj.weight": "model.safetensors",
88
+ "model.layers.2.ffn.switch_mlp.up_proj.weight": "model.safetensors",
89
+ "model.layers.2.ffn_hc.base": "model.safetensors",
90
+ "model.layers.2.ffn_hc.fn": "model.safetensors",
91
+ "model.layers.2.ffn_hc.scale": "model.safetensors",
92
+ "model.layers.2.ffn_norm.weight": "model.safetensors",
93
+ "model.layers.3.attn.attn_sink": "model.safetensors",
94
+ "model.layers.3.attn.compressor.ape": "model.safetensors",
95
+ "model.layers.3.attn.compressor.norm.weight": "model.safetensors",
96
+ "model.layers.3.attn.compressor.wgate.weight": "model.safetensors",
97
+ "model.layers.3.attn.compressor.wkv.weight": "model.safetensors",
98
+ "model.layers.3.attn.kv_norm.weight": "model.safetensors",
99
+ "model.layers.3.attn.q_norm.weight": "model.safetensors",
100
+ "model.layers.3.attn.wkv.weight": "model.safetensors",
101
+ "model.layers.3.attn.wo_a.weight": "model.safetensors",
102
+ "model.layers.3.attn.wo_b.weight": "model.safetensors",
103
+ "model.layers.3.attn.wq_a.weight": "model.safetensors",
104
+ "model.layers.3.attn.wq_b.weight": "model.safetensors",
105
+ "model.layers.3.attn_hc.base": "model.safetensors",
106
+ "model.layers.3.attn_hc.fn": "model.safetensors",
107
+ "model.layers.3.attn_hc.scale": "model.safetensors",
108
+ "model.layers.3.attn_norm.weight": "model.safetensors",
109
+ "model.layers.3.ffn.gate.e_score_correction_bias": "model.safetensors",
110
+ "model.layers.3.ffn.gate.weight": "model.safetensors",
111
+ "model.layers.3.ffn.shared_experts.down_proj.weight": "model.safetensors",
112
+ "model.layers.3.ffn.shared_experts.gate_proj.weight": "model.safetensors",
113
+ "model.layers.3.ffn.shared_experts.up_proj.weight": "model.safetensors",
114
+ "model.layers.3.ffn.switch_mlp.down_proj.weight": "model.safetensors",
115
+ "model.layers.3.ffn.switch_mlp.gate_proj.weight": "model.safetensors",
116
+ "model.layers.3.ffn.switch_mlp.up_proj.weight": "model.safetensors",
117
+ "model.layers.3.ffn_hc.base": "model.safetensors",
118
+ "model.layers.3.ffn_hc.fn": "model.safetensors",
119
+ "model.layers.3.ffn_hc.scale": "model.safetensors",
120
+ "model.layers.3.ffn_norm.weight": "model.safetensors",
121
+ "model.layers.4.attn.attn_sink": "model.safetensors",
122
+ "model.layers.4.attn.compressor.ape": "model.safetensors",
123
+ "model.layers.4.attn.compressor.norm.weight": "model.safetensors",
124
+ "model.layers.4.attn.compressor.wgate.weight": "model.safetensors",
125
+ "model.layers.4.attn.compressor.wkv.weight": "model.safetensors",
126
+ "model.layers.4.attn.indexer.compressor.ape": "model.safetensors",
127
+ "model.layers.4.attn.indexer.compressor.norm.weight": "model.safetensors",
128
+ "model.layers.4.attn.indexer.compressor.wgate.weight": "model.safetensors",
129
+ "model.layers.4.attn.indexer.compressor.wkv.weight": "model.safetensors",
130
+ "model.layers.4.attn.indexer.weights_proj.weight": "model.safetensors",
131
+ "model.layers.4.attn.indexer.wq_b.weight": "model.safetensors",
132
+ "model.layers.4.attn.kv_norm.weight": "model.safetensors",
133
+ "model.layers.4.attn.q_norm.weight": "model.safetensors",
134
+ "model.layers.4.attn.wkv.weight": "model.safetensors",
135
+ "model.layers.4.attn.wo_a.weight": "model.safetensors",
136
+ "model.layers.4.attn.wo_b.weight": "model.safetensors",
137
+ "model.layers.4.attn.wq_a.weight": "model.safetensors",
138
+ "model.layers.4.attn.wq_b.weight": "model.safetensors",
139
+ "model.layers.4.attn_hc.base": "model.safetensors",
140
+ "model.layers.4.attn_hc.fn": "model.safetensors",
141
+ "model.layers.4.attn_hc.scale": "model.safetensors",
142
+ "model.layers.4.attn_norm.weight": "model.safetensors",
143
+ "model.layers.4.ffn.gate.e_score_correction_bias": "model.safetensors",
144
+ "model.layers.4.ffn.gate.weight": "model.safetensors",
145
+ "model.layers.4.ffn.shared_experts.down_proj.weight": "model.safetensors",
146
+ "model.layers.4.ffn.shared_experts.gate_proj.weight": "model.safetensors",
147
+ "model.layers.4.ffn.shared_experts.up_proj.weight": "model.safetensors",
148
+ "model.layers.4.ffn.switch_mlp.down_proj.weight": "model.safetensors",
149
+ "model.layers.4.ffn.switch_mlp.gate_proj.weight": "model.safetensors",
150
+ "model.layers.4.ffn.switch_mlp.up_proj.weight": "model.safetensors",
151
+ "model.layers.4.ffn_hc.base": "model.safetensors",
152
+ "model.layers.4.ffn_hc.fn": "model.safetensors",
153
+ "model.layers.4.ffn_hc.scale": "model.safetensors",
154
+ "model.layers.4.ffn_norm.weight": "model.safetensors",
155
+ "model.norm.weight": "model.safetensors"
156
+ }
157
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "bos_token": "<｜begin▁of▁sentence｜>",
4
+ "clean_up_tokenization_spaces": false,
5
+ "eos_token": "<｜end▁of▁sentence｜>",
6
+ "is_local": true,
7
+ "legacy": true,
8
+ "local_files_only": false,
9
+ "model_max_length": 1048576,
10
+ "pad_token": "<｜end▁of▁sentence｜>",
11
+ "sp_model_kwargs": {},
12
+ "tokenizer_class": "TokenizersBackend",
13
+ "unk_token": null
14
+ }