Hugging Face's logo

WaveCut
/
LingBot-Video-MoE-30B-A3B-SDNQ-uint4-static

Text-to-Video
Diffusers
Safetensors
LingBotVideoPipeline
lingbot-video
sdnq
uint4
8-bit precision
Model card Files
xet
 Community
WaveCut commited on
Commit
8609f64
·
verified ·
1 Parent(s): 322d0a8

Add files using upload-large-folder tool

Browse files
Files changed (12) hide show
  1. transformer/.quantization_complete +1 -0
  2. transformer/component.SHA256SUMS +11 -0
  3. transformer/config.json +260 -0
  4. transformer/coverage.json +0 -0
  5. transformer/diffusion_pytorch_model-00001-of-00005.safetensors +3 -0
  6. transformer/diffusion_pytorch_model-00002-of-00005.safetensors +3 -0
  7. transformer/diffusion_pytorch_model-00003-of-00005.safetensors +3 -0
  8. transformer/diffusion_pytorch_model-00004-of-00005.safetensors +3 -0
  9. transformer/diffusion_pytorch_model-00005-of-00005.safetensors +3 -0
  10. transformer/diffusion_pytorch_model.safetensors.index.json +0 -0
  11. transformer/quantization_config.json +215 -0
  12. transformer/sdnq_experts.json +0 -0
transformer/.quantization_complete ADDED
@@ -0,0 +1 @@
 
 
1
+ ok
transformer/component.SHA256SUMS ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dc51b8c96c2d745df3bd5590d990230a482fd247123599548e0632fdbf97fc22 ./.quantization_complete
2
+ 8f3e9ff1d5ec390da0d466f8ed8869042ae36e704a5fc7ecf8725ba689175da8 ./config.json
3
+ 8bc6949365153e7e5779142a96de01ab23b8979f6c92a4d8ede7eea99e6ae6a5 ./coverage.json
4
+ 380d327ee8172fd93869930380e705916558dd25341d957893f0f371b1203993 ./diffusion_pytorch_model-00001-of-00005.safetensors
5
+ 7b3d48e0cbca336738276847ef5e44de42f03ada1cd762e56e9e86fcd93a49b6 ./diffusion_pytorch_model-00002-of-00005.safetensors
6
+ ed3461cb304ab56829a6cd34f379361578ef6eccabbdac87d028b39e4268ea25 ./diffusion_pytorch_model-00003-of-00005.safetensors
7
+ fc1c9feafa182672e3e33872a76575f53a4727f52e9df3bb81a36ab2b47be84a ./diffusion_pytorch_model-00004-of-00005.safetensors
8
+ 37422bbc28d1d9b8793ab43ad584b22271db07fcf5957f3c66ee46a888bb750c ./diffusion_pytorch_model-00005-of-00005.safetensors
9
+ 48d8cb723c47742c0e875083a3e699b8a4541d7fc1036b25a1a160746efe051d ./diffusion_pytorch_model.safetensors.index.json
10
+ f594ca9c58c2343b92e543cc7c37cba81c96aba36bc7410f1d645c82c80417bc ./quantization_config.json
11
+ b49557484202c5499185e6af092f850d4b69788fddf2ccda57c06f5493887027 ./sdnq_experts.json
transformer/config.json ADDED
@@ -0,0 +1,260 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_class_name": "LingBotVideoTransformer3DModel",
3
+ "_diffusers_version": "0.39.0",
4
+ "_name_or_path": "models/moe",
5
+ "axes_dims": [
6
+ 32,
7
+ 48,
8
+ 48
9
+ ],
10
+ "axes_lens": [
11
+ 4096,
12
+ 512,
13
+ 512
14
+ ],
15
+ "decoder_sparse_step": 1,
16
+ "depth": 48,
17
+ "freq_dim": 256,
18
+ "hidden_size": 2048,
19
+ "in_channels": 16,
20
+ "intermediate_size": 6144,
21
+ "mlp_only_layers": [],
22
+ "moe_intermediate_size": 768,
23
+ "n_group": 4,
24
+ "n_shared_experts": 1,
25
+ "norm_eps": 1e-06,
26
+ "norm_topk_prob": true,
27
+ "num_attention_heads": 16,
28
+ "num_experts": 128,
29
+ "num_experts_per_tok": 8,
30
+ "out_bias": true,
31
+ "out_channels": 16,
32
+ "patch_embed_bias": true,
33
+ "patch_size": [
34
+ 1,
35
+ 2,
36
+ 2
37
+ ],
38
+ "qkv_bias": false,
39
+ "quantization_config": {
40
+ "add_skip_keys": true,
41
+ "dequantize_fp32": false,
42
+ "dynamic_loss_threshold": null,
43
+ "group_size": 0,
44
+ "hadamard_group_size": 256,
45
+ "is_integer": true,
46
+ "is_training": false,
47
+ "minimum_allowed_numel": 16384,
48
+ "modules_dtype_dict": {},
49
+ "modules_quant_config": {},
50
+ "modules_to_not_convert": [
51
+ "wte",
52
+ "norm_post_attn",
53
+ "norm2",
54
+ ".proj_out",
55
+ "patch_embed",
56
+ ".vid_out",
57
+ ".time_embed",
58
+ "patch_emb",
59
+ ".final_layer",
60
+ ".condition_embedder",
61
+ "norm_post_ffn",
62
+ "norm_out",
63
+ ".norm_out",
64
+ ".txt_in",
65
+ "norm_out_modulation",
66
+ ".vid_in",
67
+ "time_embedder",
68
+ "multi_modal_projector",
69
+ ".emb_in",
70
+ ".txt_out",
71
+ "patch_embedding",
72
+ "norm_q",
73
+ "router",
74
+ ".t_embedder",
75
+ "norm1",
76
+ ".img_in",
77
+ "scale_shift_table",
78
+ "lm_head",
79
+ ".x_embedder",
80
+ "norm",
81
+ "time_modulation",
82
+ "norm_k",
83
+ ".y_embedder",
84
+ ".emb_out",
85
+ ".img_out",
86
+ "time_text_embed",
87
+ ".context_embedder",
88
+ "blocks.0.ffn.experts.w1_sdnq.weight",
89
+ "blocks.0.ffn.experts.w2_sdnq.weight",
90
+ "blocks.0.ffn.experts.w3_sdnq.weight",
91
+ "blocks.1.ffn.experts.w1_sdnq.weight",
92
+ "blocks.1.ffn.experts.w2_sdnq.weight",
93
+ "blocks.1.ffn.experts.w3_sdnq.weight",
94
+ "blocks.2.ffn.experts.w1_sdnq.weight",
95
+ "blocks.2.ffn.experts.w2_sdnq.weight",
96
+ "blocks.2.ffn.experts.w3_sdnq.weight",
97
+ "blocks.3.ffn.experts.w1_sdnq.weight",
98
+ "blocks.3.ffn.experts.w2_sdnq.weight",
99
+ "blocks.3.ffn.experts.w3_sdnq.weight",
100
+ "blocks.4.ffn.experts.w1_sdnq.weight",
101
+ "blocks.4.ffn.experts.w2_sdnq.weight",
102
+ "blocks.4.ffn.experts.w3_sdnq.weight",
103
+ "blocks.5.ffn.experts.w1_sdnq.weight",
104
+ "blocks.5.ffn.experts.w2_sdnq.weight",
105
+ "blocks.5.ffn.experts.w3_sdnq.weight",
106
+ "blocks.6.ffn.experts.w1_sdnq.weight",
107
+ "blocks.6.ffn.experts.w2_sdnq.weight",
108
+ "blocks.6.ffn.experts.w3_sdnq.weight",
109
+ "blocks.7.ffn.experts.w1_sdnq.weight",
110
+ "blocks.7.ffn.experts.w2_sdnq.weight",
111
+ "blocks.7.ffn.experts.w3_sdnq.weight",
112
+ "blocks.8.ffn.experts.w1_sdnq.weight",
113
+ "blocks.8.ffn.experts.w2_sdnq.weight",
114
+ "blocks.8.ffn.experts.w3_sdnq.weight",
115
+ "blocks.9.ffn.experts.w1_sdnq.weight",
116
+ "blocks.9.ffn.experts.w2_sdnq.weight",
117
+ "blocks.9.ffn.experts.w3_sdnq.weight",
118
+ "blocks.10.ffn.experts.w1_sdnq.weight",
119
+ "blocks.10.ffn.experts.w2_sdnq.weight",
120
+ "blocks.10.ffn.experts.w3_sdnq.weight",
121
+ "blocks.11.ffn.experts.w1_sdnq.weight",
122
+ "blocks.11.ffn.experts.w2_sdnq.weight",
123
+ "blocks.11.ffn.experts.w3_sdnq.weight",
124
+ "blocks.12.ffn.experts.w1_sdnq.weight",
125
+ "blocks.12.ffn.experts.w2_sdnq.weight",
126
+ "blocks.12.ffn.experts.w3_sdnq.weight",
127
+ "blocks.13.ffn.experts.w1_sdnq.weight",
128
+ "blocks.13.ffn.experts.w2_sdnq.weight",
129
+ "blocks.13.ffn.experts.w3_sdnq.weight",
130
+ "blocks.14.ffn.experts.w1_sdnq.weight",
131
+ "blocks.14.ffn.experts.w2_sdnq.weight",
132
+ "blocks.14.ffn.experts.w3_sdnq.weight",
133
+ "blocks.15.ffn.experts.w1_sdnq.weight",
134
+ "blocks.15.ffn.experts.w2_sdnq.weight",
135
+ "blocks.15.ffn.experts.w3_sdnq.weight",
136
+ "blocks.16.ffn.experts.w1_sdnq.weight",
137
+ "blocks.16.ffn.experts.w2_sdnq.weight",
138
+ "blocks.16.ffn.experts.w3_sdnq.weight",
139
+ "blocks.17.ffn.experts.w1_sdnq.weight",
140
+ "blocks.17.ffn.experts.w2_sdnq.weight",
141
+ "blocks.17.ffn.experts.w3_sdnq.weight",
142
+ "blocks.18.ffn.experts.w1_sdnq.weight",
143
+ "blocks.18.ffn.experts.w2_sdnq.weight",
144
+ "blocks.18.ffn.experts.w3_sdnq.weight",
145
+ "blocks.19.ffn.experts.w1_sdnq.weight",
146
+ "blocks.19.ffn.experts.w2_sdnq.weight",
147
+ "blocks.19.ffn.experts.w3_sdnq.weight",
148
+ "blocks.20.ffn.experts.w1_sdnq.weight",
149
+ "blocks.20.ffn.experts.w2_sdnq.weight",
150
+ "blocks.20.ffn.experts.w3_sdnq.weight",
151
+ "blocks.21.ffn.experts.w1_sdnq.weight",
152
+ "blocks.21.ffn.experts.w2_sdnq.weight",
153
+ "blocks.21.ffn.experts.w3_sdnq.weight",
154
+ "blocks.22.ffn.experts.w1_sdnq.weight",
155
+ "blocks.22.ffn.experts.w2_sdnq.weight",
156
+ "blocks.22.ffn.experts.w3_sdnq.weight",
157
+ "blocks.23.ffn.experts.w1_sdnq.weight",
158
+ "blocks.23.ffn.experts.w2_sdnq.weight",
159
+ "blocks.23.ffn.experts.w3_sdnq.weight",
160
+ "blocks.24.ffn.experts.w1_sdnq.weight",
161
+ "blocks.24.ffn.experts.w2_sdnq.weight",
162
+ "blocks.24.ffn.experts.w3_sdnq.weight",
163
+ "blocks.25.ffn.experts.w1_sdnq.weight",
164
+ "blocks.25.ffn.experts.w2_sdnq.weight",
165
+ "blocks.25.ffn.experts.w3_sdnq.weight",
166
+ "blocks.26.ffn.experts.w1_sdnq.weight",
167
+ "blocks.26.ffn.experts.w2_sdnq.weight",
168
+ "blocks.26.ffn.experts.w3_sdnq.weight",
169
+ "blocks.27.ffn.experts.w1_sdnq.weight",
170
+ "blocks.27.ffn.experts.w2_sdnq.weight",
171
+ "blocks.27.ffn.experts.w3_sdnq.weight",
172
+ "blocks.28.ffn.experts.w1_sdnq.weight",
173
+ "blocks.28.ffn.experts.w2_sdnq.weight",
174
+ "blocks.28.ffn.experts.w3_sdnq.weight",
175
+ "blocks.29.ffn.experts.w1_sdnq.weight",
176
+ "blocks.29.ffn.experts.w2_sdnq.weight",
177
+ "blocks.29.ffn.experts.w3_sdnq.weight",
178
+ "blocks.30.ffn.experts.w1_sdnq.weight",
179
+ "blocks.30.ffn.experts.w2_sdnq.weight",
180
+ "blocks.30.ffn.experts.w3_sdnq.weight",
181
+ "blocks.31.ffn.experts.w1_sdnq.weight",
182
+ "blocks.31.ffn.experts.w2_sdnq.weight",
183
+ "blocks.31.ffn.experts.w3_sdnq.weight",
184
+ "blocks.32.ffn.experts.w1_sdnq.weight",
185
+ "blocks.32.ffn.experts.w2_sdnq.weight",
186
+ "blocks.32.ffn.experts.w3_sdnq.weight",
187
+ "blocks.33.ffn.experts.w1_sdnq.weight",
188
+ "blocks.33.ffn.experts.w2_sdnq.weight",
189
+ "blocks.33.ffn.experts.w3_sdnq.weight",
190
+ "blocks.34.ffn.experts.w1_sdnq.weight",
191
+ "blocks.34.ffn.experts.w2_sdnq.weight",
192
+ "blocks.34.ffn.experts.w3_sdnq.weight",
193
+ "blocks.35.ffn.experts.w1_sdnq.weight",
194
+ "blocks.35.ffn.experts.w2_sdnq.weight",
195
+ "blocks.35.ffn.experts.w3_sdnq.weight",
196
+ "blocks.36.ffn.experts.w1_sdnq.weight",
197
+ "blocks.36.ffn.experts.w2_sdnq.weight",
198
+ "blocks.36.ffn.experts.w3_sdnq.weight",
199
+ "blocks.37.ffn.experts.w1_sdnq.weight",
200
+ "blocks.37.ffn.experts.w2_sdnq.weight",
201
+ "blocks.37.ffn.experts.w3_sdnq.weight",
202
+ "blocks.38.ffn.experts.w1_sdnq.weight",
203
+ "blocks.38.ffn.experts.w2_sdnq.weight",
204
+ "blocks.38.ffn.experts.w3_sdnq.weight",
205
+ "blocks.39.ffn.experts.w1_sdnq.weight",
206
+ "blocks.39.ffn.experts.w2_sdnq.weight",
207
+ "blocks.39.ffn.experts.w3_sdnq.weight",
208
+ "blocks.40.ffn.experts.w1_sdnq.weight",
209
+ "blocks.40.ffn.experts.w2_sdnq.weight",
210
+ "blocks.40.ffn.experts.w3_sdnq.weight",
211
+ "blocks.41.ffn.experts.w1_sdnq.weight",
212
+ "blocks.41.ffn.experts.w2_sdnq.weight",
213
+ "blocks.41.ffn.experts.w3_sdnq.weight",
214
+ "blocks.42.ffn.experts.w1_sdnq.weight",
215
+ "blocks.42.ffn.experts.w2_sdnq.weight",
216
+ "blocks.42.ffn.experts.w3_sdnq.weight",
217
+ "blocks.43.ffn.experts.w1_sdnq.weight",
218
+ "blocks.43.ffn.experts.w2_sdnq.weight",
219
+ "blocks.43.ffn.experts.w3_sdnq.weight",
220
+ "blocks.44.ffn.experts.w1_sdnq.weight",
221
+ "blocks.44.ffn.experts.w2_sdnq.weight",
222
+ "blocks.44.ffn.experts.w3_sdnq.weight",
223
+ "blocks.45.ffn.experts.w1_sdnq.weight",
224
+ "blocks.45.ffn.experts.w2_sdnq.weight",
225
+ "blocks.45.ffn.experts.w3_sdnq.weight",
226
+ "blocks.46.ffn.experts.w1_sdnq.weight",
227
+ "blocks.46.ffn.experts.w2_sdnq.weight",
228
+ "blocks.46.ffn.experts.w3_sdnq.weight",
229
+ "blocks.47.ffn.experts.w1_sdnq.weight",
230
+ "blocks.47.ffn.experts.w2_sdnq.weight",
231
+ "blocks.47.ffn.experts.w3_sdnq.weight"
232
+ ],
233
+ "modules_to_not_use_matmul": [],
234
+ "non_blocking": false,
235
+ "quant_conv": false,
236
+ "quant_embedding": false,
237
+ "quant_method": "sdnq",
238
+ "quantization_device": "cuda",
239
+ "quantized_matmul_dtype": null,
240
+ "return_device": "cpu",
241
+ "sdnq_version": "0.2.2",
242
+ "svd_rank": 32,
243
+ "svd_steps": 8,
244
+ "use_dynamic_quantization": false,
245
+ "use_grad_ckpt": true,
246
+ "use_hadamard": false,
247
+ "use_quantized_matmul": false,
248
+ "use_quantized_matmul_conv": false,
249
+ "use_static_quantization": true,
250
+ "use_stochastic_rounding": false,
251
+ "use_svd": false,
252
+ "weights_dtype": "uint4"
253
+ },
254
+ "rope_theta": 256.0,
255
+ "routed_scaling_factor": 2.5,
256
+ "score_func": "sigmoid",
257
+ "text_dim": 2560,
258
+ "timestep_mlp_bias": true,
259
+ "topk_group": 2
260
+ }
transformer/coverage.json ADDED
The diff for this file is too large to render. See raw diff
 
transformer/diffusion_pytorch_model-00001-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:380d327ee8172fd93869930380e705916558dd25341d957893f0f371b1203993
3
+ size 3903708600
transformer/diffusion_pytorch_model-00002-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b3d48e0cbca336738276847ef5e44de42f03ada1cd762e56e9e86fcd93a49b6
3
+ size 3907192200
transformer/diffusion_pytorch_model-00003-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ed3461cb304ab56829a6cd34f379361578ef6eccabbdac87d028b39e4268ea25
3
+ size 3907192216
transformer/diffusion_pytorch_model-00004-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fc1c9feafa182672e3e33872a76575f53a4727f52e9df3bb81a36ab2b47be84a
3
+ size 3907192216
transformer/diffusion_pytorch_model-00005-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:37422bbc28d1d9b8793ab43ad584b22271db07fcf5957f3c66ee46a888bb750c
3
+ size 3288080256
transformer/diffusion_pytorch_model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
transformer/quantization_config.json ADDED
@@ -0,0 +1,215 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_skip_keys": true,
3
+ "dequantize_fp32": false,
4
+ "dynamic_loss_threshold": null,
5
+ "group_size": 0,
6
+ "hadamard_group_size": 256,
7
+ "is_integer": true,
8
+ "is_training": false,
9
+ "minimum_allowed_numel": 16384,
10
+ "modules_dtype_dict": {},
11
+ "modules_quant_config": {},
12
+ "modules_to_not_convert": [
13
+ "wte",
14
+ "norm_post_attn",
15
+ "norm2",
16
+ ".proj_out",
17
+ "patch_embed",
18
+ ".vid_out",
19
+ ".time_embed",
20
+ "patch_emb",
21
+ ".final_layer",
22
+ ".condition_embedder",
23
+ "norm_post_ffn",
24
+ "norm_out",
25
+ ".norm_out",
26
+ ".txt_in",
27
+ "norm_out_modulation",
28
+ ".vid_in",
29
+ "time_embedder",
30
+ "multi_modal_projector",
31
+ ".emb_in",
32
+ ".txt_out",
33
+ "patch_embedding",
34
+ "norm_q",
35
+ "router",
36
+ ".t_embedder",
37
+ "norm1",
38
+ ".img_in",
39
+ "scale_shift_table",
40
+ "lm_head",
41
+ ".x_embedder",
42
+ "norm",
43
+ "time_modulation",
44
+ "norm_k",
45
+ ".y_embedder",
46
+ ".emb_out",
47
+ ".img_out",
48
+ "time_text_embed",
49
+ ".context_embedder",
50
+ "blocks.0.ffn.experts.w1_sdnq.weight",
51
+ "blocks.0.ffn.experts.w2_sdnq.weight",
52
+ "blocks.0.ffn.experts.w3_sdnq.weight",
53
+ "blocks.1.ffn.experts.w1_sdnq.weight",
54
+ "blocks.1.ffn.experts.w2_sdnq.weight",
55
+ "blocks.1.ffn.experts.w3_sdnq.weight",
56
+ "blocks.2.ffn.experts.w1_sdnq.weight",
57
+ "blocks.2.ffn.experts.w2_sdnq.weight",
58
+ "blocks.2.ffn.experts.w3_sdnq.weight",
59
+ "blocks.3.ffn.experts.w1_sdnq.weight",
60
+ "blocks.3.ffn.experts.w2_sdnq.weight",
61
+ "blocks.3.ffn.experts.w3_sdnq.weight",
62
+ "blocks.4.ffn.experts.w1_sdnq.weight",
63
+ "blocks.4.ffn.experts.w2_sdnq.weight",
64
+ "blocks.4.ffn.experts.w3_sdnq.weight",
65
+ "blocks.5.ffn.experts.w1_sdnq.weight",
66
+ "blocks.5.ffn.experts.w2_sdnq.weight",
67
+ "blocks.5.ffn.experts.w3_sdnq.weight",
68
+ "blocks.6.ffn.experts.w1_sdnq.weight",
69
+ "blocks.6.ffn.experts.w2_sdnq.weight",
70
+ "blocks.6.ffn.experts.w3_sdnq.weight",
71
+ "blocks.7.ffn.experts.w1_sdnq.weight",
72
+ "blocks.7.ffn.experts.w2_sdnq.weight",
73
+ "blocks.7.ffn.experts.w3_sdnq.weight",
74
+ "blocks.8.ffn.experts.w1_sdnq.weight",
75
+ "blocks.8.ffn.experts.w2_sdnq.weight",
76
+ "blocks.8.ffn.experts.w3_sdnq.weight",
77
+ "blocks.9.ffn.experts.w1_sdnq.weight",
78
+ "blocks.9.ffn.experts.w2_sdnq.weight",
79
+ "blocks.9.ffn.experts.w3_sdnq.weight",
80
+ "blocks.10.ffn.experts.w1_sdnq.weight",
81
+ "blocks.10.ffn.experts.w2_sdnq.weight",
82
+ "blocks.10.ffn.experts.w3_sdnq.weight",
83
+ "blocks.11.ffn.experts.w1_sdnq.weight",
84
+ "blocks.11.ffn.experts.w2_sdnq.weight",
85
+ "blocks.11.ffn.experts.w3_sdnq.weight",
86
+ "blocks.12.ffn.experts.w1_sdnq.weight",
87
+ "blocks.12.ffn.experts.w2_sdnq.weight",
88
+ "blocks.12.ffn.experts.w3_sdnq.weight",
89
+ "blocks.13.ffn.experts.w1_sdnq.weight",
90
+ "blocks.13.ffn.experts.w2_sdnq.weight",
91
+ "blocks.13.ffn.experts.w3_sdnq.weight",
92
+ "blocks.14.ffn.experts.w1_sdnq.weight",
93
+ "blocks.14.ffn.experts.w2_sdnq.weight",
94
+ "blocks.14.ffn.experts.w3_sdnq.weight",
95
+ "blocks.15.ffn.experts.w1_sdnq.weight",
96
+ "blocks.15.ffn.experts.w2_sdnq.weight",
97
+ "blocks.15.ffn.experts.w3_sdnq.weight",
98
+ "blocks.16.ffn.experts.w1_sdnq.weight",
99
+ "blocks.16.ffn.experts.w2_sdnq.weight",
100
+ "blocks.16.ffn.experts.w3_sdnq.weight",
101
+ "blocks.17.ffn.experts.w1_sdnq.weight",
102
+ "blocks.17.ffn.experts.w2_sdnq.weight",
103
+ "blocks.17.ffn.experts.w3_sdnq.weight",
104
+ "blocks.18.ffn.experts.w1_sdnq.weight",
105
+ "blocks.18.ffn.experts.w2_sdnq.weight",
106
+ "blocks.18.ffn.experts.w3_sdnq.weight",
107
+ "blocks.19.ffn.experts.w1_sdnq.weight",
108
+ "blocks.19.ffn.experts.w2_sdnq.weight",
109
+ "blocks.19.ffn.experts.w3_sdnq.weight",
110
+ "blocks.20.ffn.experts.w1_sdnq.weight",
111
+ "blocks.20.ffn.experts.w2_sdnq.weight",
112
+ "blocks.20.ffn.experts.w3_sdnq.weight",
113
+ "blocks.21.ffn.experts.w1_sdnq.weight",
114
+ "blocks.21.ffn.experts.w2_sdnq.weight",
115
+ "blocks.21.ffn.experts.w3_sdnq.weight",
116
+ "blocks.22.ffn.experts.w1_sdnq.weight",
117
+ "blocks.22.ffn.experts.w2_sdnq.weight",
118
+ "blocks.22.ffn.experts.w3_sdnq.weight",
119
+ "blocks.23.ffn.experts.w1_sdnq.weight",
120
+ "blocks.23.ffn.experts.w2_sdnq.weight",
121
+ "blocks.23.ffn.experts.w3_sdnq.weight",
122
+ "blocks.24.ffn.experts.w1_sdnq.weight",
123
+ "blocks.24.ffn.experts.w2_sdnq.weight",
124
+ "blocks.24.ffn.experts.w3_sdnq.weight",
125
+ "blocks.25.ffn.experts.w1_sdnq.weight",
126
+ "blocks.25.ffn.experts.w2_sdnq.weight",
127
+ "blocks.25.ffn.experts.w3_sdnq.weight",
128
+ "blocks.26.ffn.experts.w1_sdnq.weight",
129
+ "blocks.26.ffn.experts.w2_sdnq.weight",
130
+ "blocks.26.ffn.experts.w3_sdnq.weight",
131
+ "blocks.27.ffn.experts.w1_sdnq.weight",
132
+ "blocks.27.ffn.experts.w2_sdnq.weight",
133
+ "blocks.27.ffn.experts.w3_sdnq.weight",
134
+ "blocks.28.ffn.experts.w1_sdnq.weight",
135
+ "blocks.28.ffn.experts.w2_sdnq.weight",
136
+ "blocks.28.ffn.experts.w3_sdnq.weight",
137
+ "blocks.29.ffn.experts.w1_sdnq.weight",
138
+ "blocks.29.ffn.experts.w2_sdnq.weight",
139
+ "blocks.29.ffn.experts.w3_sdnq.weight",
140
+ "blocks.30.ffn.experts.w1_sdnq.weight",
141
+ "blocks.30.ffn.experts.w2_sdnq.weight",
142
+ "blocks.30.ffn.experts.w3_sdnq.weight",
143
+ "blocks.31.ffn.experts.w1_sdnq.weight",
144
+ "blocks.31.ffn.experts.w2_sdnq.weight",
145
+ "blocks.31.ffn.experts.w3_sdnq.weight",
146
+ "blocks.32.ffn.experts.w1_sdnq.weight",
147
+ "blocks.32.ffn.experts.w2_sdnq.weight",
148
+ "blocks.32.ffn.experts.w3_sdnq.weight",
149
+ "blocks.33.ffn.experts.w1_sdnq.weight",
150
+ "blocks.33.ffn.experts.w2_sdnq.weight",
151
+ "blocks.33.ffn.experts.w3_sdnq.weight",
152
+ "blocks.34.ffn.experts.w1_sdnq.weight",
153
+ "blocks.34.ffn.experts.w2_sdnq.weight",
154
+ "blocks.34.ffn.experts.w3_sdnq.weight",
155
+ "blocks.35.ffn.experts.w1_sdnq.weight",
156
+ "blocks.35.ffn.experts.w2_sdnq.weight",
157
+ "blocks.35.ffn.experts.w3_sdnq.weight",
158
+ "blocks.36.ffn.experts.w1_sdnq.weight",
159
+ "blocks.36.ffn.experts.w2_sdnq.weight",
160
+ "blocks.36.ffn.experts.w3_sdnq.weight",
161
+ "blocks.37.ffn.experts.w1_sdnq.weight",
162
+ "blocks.37.ffn.experts.w2_sdnq.weight",
163
+ "blocks.37.ffn.experts.w3_sdnq.weight",
164
+ "blocks.38.ffn.experts.w1_sdnq.weight",
165
+ "blocks.38.ffn.experts.w2_sdnq.weight",
166
+ "blocks.38.ffn.experts.w3_sdnq.weight",
167
+ "blocks.39.ffn.experts.w1_sdnq.weight",
168
+ "blocks.39.ffn.experts.w2_sdnq.weight",
169
+ "blocks.39.ffn.experts.w3_sdnq.weight",
170
+ "blocks.40.ffn.experts.w1_sdnq.weight",
171
+ "blocks.40.ffn.experts.w2_sdnq.weight",
172
+ "blocks.40.ffn.experts.w3_sdnq.weight",
173
+ "blocks.41.ffn.experts.w1_sdnq.weight",
174
+ "blocks.41.ffn.experts.w2_sdnq.weight",
175
+ "blocks.41.ffn.experts.w3_sdnq.weight",
176
+ "blocks.42.ffn.experts.w1_sdnq.weight",
177
+ "blocks.42.ffn.experts.w2_sdnq.weight",
178
+ "blocks.42.ffn.experts.w3_sdnq.weight",
179
+ "blocks.43.ffn.experts.w1_sdnq.weight",
180
+ "blocks.43.ffn.experts.w2_sdnq.weight",
181
+ "blocks.43.ffn.experts.w3_sdnq.weight",
182
+ "blocks.44.ffn.experts.w1_sdnq.weight",
183
+ "blocks.44.ffn.experts.w2_sdnq.weight",
184
+ "blocks.44.ffn.experts.w3_sdnq.weight",
185
+ "blocks.45.ffn.experts.w1_sdnq.weight",
186
+ "blocks.45.ffn.experts.w2_sdnq.weight",
187
+ "blocks.45.ffn.experts.w3_sdnq.weight",
188
+ "blocks.46.ffn.experts.w1_sdnq.weight",
189
+ "blocks.46.ffn.experts.w2_sdnq.weight",
190
+ "blocks.46.ffn.experts.w3_sdnq.weight",
191
+ "blocks.47.ffn.experts.w1_sdnq.weight",
192
+ "blocks.47.ffn.experts.w2_sdnq.weight",
193
+ "blocks.47.ffn.experts.w3_sdnq.weight"
194
+ ],
195
+ "modules_to_not_use_matmul": [],
196
+ "non_blocking": false,
197
+ "quant_conv": false,
198
+ "quant_embedding": false,
199
+ "quant_method": "sdnq",
200
+ "quantization_device": "cuda",
201
+ "quantized_matmul_dtype": null,
202
+ "return_device": "cpu",
203
+ "sdnq_version": "0.2.2",
204
+ "svd_rank": 32,
205
+ "svd_steps": 8,
206
+ "use_dynamic_quantization": false,
207
+ "use_grad_ckpt": true,
208
+ "use_hadamard": false,
209
+ "use_quantized_matmul": false,
210
+ "use_quantized_matmul_conv": false,
211
+ "use_static_quantization": true,
212
+ "use_stochastic_rounding": false,
213
+ "use_svd": false,
214
+ "weights_dtype": "uint4"
215
+ }
transformer/sdnq_experts.json ADDED
The diff for this file is too large to render. See raw diff