Hugging Face's logo

WaveCut
/
FLUX.2-klein-4B-OrbitQuant-W2A4

Diffusers
Safetensors
OrbitQuantComponentArtifact
orbitquant
quantized
diffusion-transformer
8-bit precision
Model card Files
xet
 Community
WaveCut commited on
Commit
a4957b7
·
verified ·
1 Parent(s): 426ea38

Upload FLUX.2 Klein OrbitQuant W2A4 artifact

Browse files
Files changed (15) hide show
  1. README.md +38 -0
  2. SHA256SUMS +14 -0
  3. assets/.gitkeep +0 -0
  4. benchmark/orbitquant.metrics.csv +1 -0
  5. benchmark/orbitquant.metrics.jsonl +0 -0
  6. benchmark/original.metrics.csv +1 -0
  7. benchmark/original.metrics.jsonl +0 -0
  8. benchmark/summary.json +10 -0
  9. model.safetensors +3 -0
  10. model_index.json +19 -0
  11. orbitquant_codebooks.safetensors +3 -0
  12. orbitquant_manifest.json +977 -0
  13. orbitquant_rotations.safetensors +3 -0
  14. prompts.json +57 -0
  15. quantization_config.json +22 -0
README.md ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: black-forest-labs/FLUX.2-klein-4B
3
+ tags:
4
+ - orbitquant
5
+ - quantized
6
+ - diffusers
7
+ ---
8
+
9
+ # black-forest-labs/FLUX.2-klein-4B OrbitQuant W2A4
10
+
11
+ This is an OrbitQuant artifact generated from the source model listed above.
12
+
13
+ ## Quantization
14
+
15
+ - Method: `orbitquant`
16
+ - Bits: `W2A4`
17
+ - Runtime mode: `dequant_bf16`
18
+ - Activation kernel backend: `triton_cuda`
19
+ - Target policy: `flux2`
20
+ - Rotation: `rpbh`
21
+ - Rotation seed: `0`
22
+ - Block size: `paper`
23
+ - Block size policy: `largest_power_of_two_dividing_dim`
24
+ - Codebook: `lloyd_max`
25
+ - Codebook version: `1`
26
+ - Calibration data: none
27
+ - Text encoders and VAE: left in source precision by default
28
+
29
+ ## Source
30
+
31
+ - Model: `black-forest-labs/FLUX.2-klein-4B`
32
+ - Revision: `e7b7dc27f91deacad38e78976d1f2b499d76a294`
33
+ - Source license: `apache-2.0`
34
+ - OrbitQuant paper: https://arxiv.org/abs/2607.02461
35
+
36
+ ## Limitations
37
+
38
+ The initial runtime may dequantize packed weights before BF16 matmul. Disk artifacts are compact; current CUDA/MPS kernels optimize selected activation codebook lookup/rescale stages and packed weight dequantization; full fused low-bit kernels are separate work.
SHA256SUMS ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 470f0cd03ade830add0b90d49ba15255552a8bc00b6c4d7524f30ab44f006e04 README.md
2
+ e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 assets/.gitkeep
3
+ 90d54eb9870d76c7e992e1ca94341cc7a58c931a079162ef1774b7f2b571ac73 benchmark/orbitquant.metrics.csv
4
+ e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 benchmark/orbitquant.metrics.jsonl
5
+ 90d54eb9870d76c7e992e1ca94341cc7a58c931a079162ef1774b7f2b571ac73 benchmark/original.metrics.csv
6
+ e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 benchmark/original.metrics.jsonl
7
+ 34a4410c42ca64bd864be4b9f9abef076ad2f2be9c05d9b65e74b83489b03406 benchmark/summary.json
8
+ b5c4654d1a0a00d47a40b75953298796da8249726b3ad7f58700de615585ec0b model.safetensors
9
+ a49141221a603fae77e65aab145dc2e6b6794592eb1e2e7d20a6d8d85414a4e5 model_index.json
10
+ 5767b046e6b0721c56bad6306657708f60f3d779721920f8ddaab51e34a9cd10 orbitquant_codebooks.safetensors
11
+ 62034df7d1b4de9c53307e1597b6bdc59df8344dc69b6f15ae26b31506c29c77 orbitquant_manifest.json
12
+ 3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17 orbitquant_rotations.safetensors
13
+ 9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd prompts.json
14
+ 0e689c046ea3fbf76a8128b71b4686f11196ba7b835fe5cafdc71e7498801b7f quantization_config.json
assets/.gitkeep ADDED
File without changes
benchmark/orbitquant.metrics.csv ADDED
@@ -0,0 +1 @@
 
 
1
+ metric,value
benchmark/orbitquant.metrics.jsonl ADDED
File without changes
benchmark/original.metrics.csv ADDED
@@ -0,0 +1 @@
 
 
1
+ metric,value
benchmark/original.metrics.jsonl ADDED
File without changes
benchmark/summary.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "not_run",
3
+ "source_model_id": "black-forest-labs/FLUX.2-klein-4B",
4
+ "source_revision": "e7b7dc27f91deacad38e78976d1f2b499d76a294",
5
+ "weight_bits": 2,
6
+ "activation_bits": 4,
7
+ "target_policy": "flux2",
8
+ "runtime_mode": "dequant_bf16",
9
+ "activation_kernel_backend": "triton_cuda"
10
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b5c4654d1a0a00d47a40b75953298796da8249726b3ad7f58700de615585ec0b
3
+ size 1104236392
model_index.json ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_class_name": "OrbitQuantComponentArtifact",
3
+ "artifact_format": "orbitquant-v1",
4
+ "quant_method": "orbitquant",
5
+ "source_model_id": "black-forest-labs/FLUX.2-klein-4B",
6
+ "source_revision": "e7b7dc27f91deacad38e78976d1f2b499d76a294",
7
+ "source_license": "apache-2.0",
8
+ "component": "transformer",
9
+ "weight_name": "model.safetensors",
10
+ "quantization_config": "quantization_config.json",
11
+ "manifest": "orbitquant_manifest.json",
12
+ "codebooks": "orbitquant_codebooks.safetensors",
13
+ "rotations": "orbitquant_rotations.safetensors",
14
+ "weight_bits": 2,
15
+ "activation_bits": 4,
16
+ "target_policy": "flux2",
17
+ "runtime_mode": "dequant_bf16",
18
+ "activation_kernel_backend": "triton_cuda"
19
+ }
orbitquant_codebooks.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5767b046e6b0721c56bad6306657708f60f3d779721920f8ddaab51e34a9cd10
3
+ size 1424
orbitquant_manifest.json ADDED
@@ -0,0 +1,977 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifact_format": "orbitquant-v1",
3
+ "source_model_id": "black-forest-labs/FLUX.2-klein-4B",
4
+ "source_revision": "e7b7dc27f91deacad38e78976d1f2b499d76a294",
5
+ "source_license": "apache-2.0",
6
+ "quant_method": "orbitquant",
7
+ "paper": "https://arxiv.org/abs/2607.02461",
8
+ "weight_bits": 2,
9
+ "activation_bits": 4,
10
+ "rotation": "rpbh",
11
+ "rotation_seed": 0,
12
+ "block_size": "paper",
13
+ "block_size_policy": "largest_power_of_two_dividing_dim",
14
+ "codebook": "lloyd_max",
15
+ "codebook_version": 1,
16
+ "row_norm_dtype": "bfloat16",
17
+ "runtime_mode": "dequant_bf16",
18
+ "activation_kernel_backend": "triton_cuda",
19
+ "target_policy": "flux2",
20
+ "adaln_policy": "int4_rtn_group64_bf16_activation",
21
+ "quantized_modules": [
22
+ "transformer_blocks.0.attn.to_q",
23
+ "transformer_blocks.0.attn.to_k",
24
+ "transformer_blocks.0.attn.to_v",
25
+ "transformer_blocks.0.attn.to_out.0",
26
+ "transformer_blocks.0.attn.add_q_proj",
27
+ "transformer_blocks.0.attn.add_k_proj",
28
+ "transformer_blocks.0.attn.add_v_proj",
29
+ "transformer_blocks.0.attn.to_add_out",
30
+ "transformer_blocks.0.ff.linear_in",
31
+ "transformer_blocks.0.ff.linear_out",
32
+ "transformer_blocks.0.ff_context.linear_in",
33
+ "transformer_blocks.0.ff_context.linear_out",
34
+ "transformer_blocks.1.attn.to_q",
35
+ "transformer_blocks.1.attn.to_k",
36
+ "transformer_blocks.1.attn.to_v",
37
+ "transformer_blocks.1.attn.to_out.0",
38
+ "transformer_blocks.1.attn.add_q_proj",
39
+ "transformer_blocks.1.attn.add_k_proj",
40
+ "transformer_blocks.1.attn.add_v_proj",
41
+ "transformer_blocks.1.attn.to_add_out",
42
+ "transformer_blocks.1.ff.linear_in",
43
+ "transformer_blocks.1.ff.linear_out",
44
+ "transformer_blocks.1.ff_context.linear_in",
45
+ "transformer_blocks.1.ff_context.linear_out",
46
+ "transformer_blocks.2.attn.to_q",
47
+ "transformer_blocks.2.attn.to_k",
48
+ "transformer_blocks.2.attn.to_v",
49
+ "transformer_blocks.2.attn.to_out.0",
50
+ "transformer_blocks.2.attn.add_q_proj",
51
+ "transformer_blocks.2.attn.add_k_proj",
52
+ "transformer_blocks.2.attn.add_v_proj",
53
+ "transformer_blocks.2.attn.to_add_out",
54
+ "transformer_blocks.2.ff.linear_in",
55
+ "transformer_blocks.2.ff.linear_out",
56
+ "transformer_blocks.2.ff_context.linear_in",
57
+ "transformer_blocks.2.ff_context.linear_out",
58
+ "transformer_blocks.3.attn.to_q",
59
+ "transformer_blocks.3.attn.to_k",
60
+ "transformer_blocks.3.attn.to_v",
61
+ "transformer_blocks.3.attn.to_out.0",
62
+ "transformer_blocks.3.attn.add_q_proj",
63
+ "transformer_blocks.3.attn.add_k_proj",
64
+ "transformer_blocks.3.attn.add_v_proj",
65
+ "transformer_blocks.3.attn.to_add_out",
66
+ "transformer_blocks.3.ff.linear_in",
67
+ "transformer_blocks.3.ff.linear_out",
68
+ "transformer_blocks.3.ff_context.linear_in",
69
+ "transformer_blocks.3.ff_context.linear_out",
70
+ "transformer_blocks.4.attn.to_q",
71
+ "transformer_blocks.4.attn.to_k",
72
+ "transformer_blocks.4.attn.to_v",
73
+ "transformer_blocks.4.attn.to_out.0",
74
+ "transformer_blocks.4.attn.add_q_proj",
75
+ "transformer_blocks.4.attn.add_k_proj",
76
+ "transformer_blocks.4.attn.add_v_proj",
77
+ "transformer_blocks.4.attn.to_add_out",
78
+ "transformer_blocks.4.ff.linear_in",
79
+ "transformer_blocks.4.ff.linear_out",
80
+ "transformer_blocks.4.ff_context.linear_in",
81
+ "transformer_blocks.4.ff_context.linear_out",
82
+ "single_transformer_blocks.0.attn.to_qkv_mlp_proj",
83
+ "single_transformer_blocks.0.attn.to_out",
84
+ "single_transformer_blocks.1.attn.to_qkv_mlp_proj",
85
+ "single_transformer_blocks.1.attn.to_out",
86
+ "single_transformer_blocks.2.attn.to_qkv_mlp_proj",
87
+ "single_transformer_blocks.2.attn.to_out",
88
+ "single_transformer_blocks.3.attn.to_qkv_mlp_proj",
89
+ "single_transformer_blocks.3.attn.to_out",
90
+ "single_transformer_blocks.4.attn.to_qkv_mlp_proj",
91
+ "single_transformer_blocks.4.attn.to_out",
92
+ "single_transformer_blocks.5.attn.to_qkv_mlp_proj",
93
+ "single_transformer_blocks.5.attn.to_out",
94
+ "single_transformer_blocks.6.attn.to_qkv_mlp_proj",
95
+ "single_transformer_blocks.6.attn.to_out",
96
+ "single_transformer_blocks.7.attn.to_qkv_mlp_proj",
97
+ "single_transformer_blocks.7.attn.to_out",
98
+ "single_transformer_blocks.8.attn.to_qkv_mlp_proj",
99
+ "single_transformer_blocks.8.attn.to_out",
100
+ "single_transformer_blocks.9.attn.to_qkv_mlp_proj",
101
+ "single_transformer_blocks.9.attn.to_out",
102
+ "single_transformer_blocks.10.attn.to_qkv_mlp_proj",
103
+ "single_transformer_blocks.10.attn.to_out",
104
+ "single_transformer_blocks.11.attn.to_qkv_mlp_proj",
105
+ "single_transformer_blocks.11.attn.to_out",
106
+ "single_transformer_blocks.12.attn.to_qkv_mlp_proj",
107
+ "single_transformer_blocks.12.attn.to_out",
108
+ "single_transformer_blocks.13.attn.to_qkv_mlp_proj",
109
+ "single_transformer_blocks.13.attn.to_out",
110
+ "single_transformer_blocks.14.attn.to_qkv_mlp_proj",
111
+ "single_transformer_blocks.14.attn.to_out",
112
+ "single_transformer_blocks.15.attn.to_qkv_mlp_proj",
113
+ "single_transformer_blocks.15.attn.to_out",
114
+ "single_transformer_blocks.16.attn.to_qkv_mlp_proj",
115
+ "single_transformer_blocks.16.attn.to_out",
116
+ "single_transformer_blocks.17.attn.to_qkv_mlp_proj",
117
+ "single_transformer_blocks.17.attn.to_out",
118
+ "single_transformer_blocks.18.attn.to_qkv_mlp_proj",
119
+ "single_transformer_blocks.18.attn.to_out",
120
+ "single_transformer_blocks.19.attn.to_qkv_mlp_proj",
121
+ "single_transformer_blocks.19.attn.to_out"
122
+ ],
123
+ "adaln_modules": [
124
+ "double_stream_modulation_img.linear",
125
+ "double_stream_modulation_txt.linear",
126
+ "single_stream_modulation.linear"
127
+ ],
128
+ "skipped_modules": [
129
+ "time_guidance_embed.timestep_embedder.linear_1",
130
+ "time_guidance_embed.timestep_embedder.linear_2",
131
+ "x_embedder",
132
+ "context_embedder",
133
+ "norm_out.linear",
134
+ "proj_out"
135
+ ],
136
+ "module_shapes": {
137
+ "time_guidance_embed.timestep_embedder.linear_1.weight": [
138
+ 3072,
139
+ 256
140
+ ],
141
+ "time_guidance_embed.timestep_embedder.linear_2.weight": [
142
+ 3072,
143
+ 3072
144
+ ],
145
+ "double_stream_modulation_img.linear.packed_weight": [
146
+ 28311552
147
+ ],
148
+ "double_stream_modulation_img.linear.scales": [
149
+ 18432,
150
+ 48
151
+ ],
152
+ "double_stream_modulation_txt.linear.packed_weight": [
153
+ 28311552
154
+ ],
155
+ "double_stream_modulation_txt.linear.scales": [
156
+ 18432,
157
+ 48
158
+ ],
159
+ "single_stream_modulation.linear.packed_weight": [
160
+ 14155776
161
+ ],
162
+ "single_stream_modulation.linear.scales": [
163
+ 9216,
164
+ 48
165
+ ],
166
+ "x_embedder.weight": [
167
+ 3072,
168
+ 128
169
+ ],
170
+ "context_embedder.weight": [
171
+ 3072,
172
+ 7680
173
+ ],
174
+ "transformer_blocks.0.attn.to_q.packed_weight_indices": [
175
+ 2359296
176
+ ],
177
+ "transformer_blocks.0.attn.to_q.row_norms": [
178
+ 3072
179
+ ],
180
+ "transformer_blocks.0.attn.to_k.packed_weight_indices": [
181
+ 2359296
182
+ ],
183
+ "transformer_blocks.0.attn.to_k.row_norms": [
184
+ 3072
185
+ ],
186
+ "transformer_blocks.0.attn.to_v.packed_weight_indices": [
187
+ 2359296
188
+ ],
189
+ "transformer_blocks.0.attn.to_v.row_norms": [
190
+ 3072
191
+ ],
192
+ "transformer_blocks.0.attn.norm_q.weight": [
193
+ 128
194
+ ],
195
+ "transformer_blocks.0.attn.norm_k.weight": [
196
+ 128
197
+ ],
198
+ "transformer_blocks.0.attn.to_out.0.packed_weight_indices": [
199
+ 2359296
200
+ ],
201
+ "transformer_blocks.0.attn.to_out.0.row_norms": [
202
+ 3072
203
+ ],
204
+ "transformer_blocks.0.attn.norm_added_q.weight": [
205
+ 128
206
+ ],
207
+ "transformer_blocks.0.attn.norm_added_k.weight": [
208
+ 128
209
+ ],
210
+ "transformer_blocks.0.attn.add_q_proj.packed_weight_indices": [
211
+ 2359296
212
+ ],
213
+ "transformer_blocks.0.attn.add_q_proj.row_norms": [
214
+ 3072
215
+ ],
216
+ "transformer_blocks.0.attn.add_k_proj.packed_weight_indices": [
217
+ 2359296
218
+ ],
219
+ "transformer_blocks.0.attn.add_k_proj.row_norms": [
220
+ 3072
221
+ ],
222
+ "transformer_blocks.0.attn.add_v_proj.packed_weight_indices": [
223
+ 2359296
224
+ ],
225
+ "transformer_blocks.0.attn.add_v_proj.row_norms": [
226
+ 3072
227
+ ],
228
+ "transformer_blocks.0.attn.to_add_out.packed_weight_indices": [
229
+ 2359296
230
+ ],
231
+ "transformer_blocks.0.attn.to_add_out.row_norms": [
232
+ 3072
233
+ ],
234
+ "transformer_blocks.0.ff.linear_in.packed_weight_indices": [
235
+ 14155776
236
+ ],
237
+ "transformer_blocks.0.ff.linear_in.row_norms": [
238
+ 18432
239
+ ],
240
+ "transformer_blocks.0.ff.linear_out.packed_weight_indices": [
241
+ 7077888
242
+ ],
243
+ "transformer_blocks.0.ff.linear_out.row_norms": [
244
+ 3072
245
+ ],
246
+ "transformer_blocks.0.ff_context.linear_in.packed_weight_indices": [
247
+ 14155776
248
+ ],
249
+ "transformer_blocks.0.ff_context.linear_in.row_norms": [
250
+ 18432
251
+ ],
252
+ "transformer_blocks.0.ff_context.linear_out.packed_weight_indices": [
253
+ 7077888
254
+ ],
255
+ "transformer_blocks.0.ff_context.linear_out.row_norms": [
256
+ 3072
257
+ ],
258
+ "transformer_blocks.1.attn.to_q.packed_weight_indices": [
259
+ 2359296
260
+ ],
261
+ "transformer_blocks.1.attn.to_q.row_norms": [
262
+ 3072
263
+ ],
264
+ "transformer_blocks.1.attn.to_k.packed_weight_indices": [
265
+ 2359296
266
+ ],
267
+ "transformer_blocks.1.attn.to_k.row_norms": [
268
+ 3072
269
+ ],
270
+ "transformer_blocks.1.attn.to_v.packed_weight_indices": [
271
+ 2359296
272
+ ],
273
+ "transformer_blocks.1.attn.to_v.row_norms": [
274
+ 3072
275
+ ],
276
+ "transformer_blocks.1.attn.norm_q.weight": [
277
+ 128
278
+ ],
279
+ "transformer_blocks.1.attn.norm_k.weight": [
280
+ 128
281
+ ],
282
+ "transformer_blocks.1.attn.to_out.0.packed_weight_indices": [
283
+ 2359296
284
+ ],
285
+ "transformer_blocks.1.attn.to_out.0.row_norms": [
286
+ 3072
287
+ ],
288
+ "transformer_blocks.1.attn.norm_added_q.weight": [
289
+ 128
290
+ ],
291
+ "transformer_blocks.1.attn.norm_added_k.weight": [
292
+ 128
293
+ ],
294
+ "transformer_blocks.1.attn.add_q_proj.packed_weight_indices": [
295
+ 2359296
296
+ ],
297
+ "transformer_blocks.1.attn.add_q_proj.row_norms": [
298
+ 3072
299
+ ],
300
+ "transformer_blocks.1.attn.add_k_proj.packed_weight_indices": [
301
+ 2359296
302
+ ],
303
+ "transformer_blocks.1.attn.add_k_proj.row_norms": [
304
+ 3072
305
+ ],
306
+ "transformer_blocks.1.attn.add_v_proj.packed_weight_indices": [
307
+ 2359296
308
+ ],
309
+ "transformer_blocks.1.attn.add_v_proj.row_norms": [
310
+ 3072
311
+ ],
312
+ "transformer_blocks.1.attn.to_add_out.packed_weight_indices": [
313
+ 2359296
314
+ ],
315
+ "transformer_blocks.1.attn.to_add_out.row_norms": [
316
+ 3072
317
+ ],
318
+ "transformer_blocks.1.ff.linear_in.packed_weight_indices": [
319
+ 14155776
320
+ ],
321
+ "transformer_blocks.1.ff.linear_in.row_norms": [
322
+ 18432
323
+ ],
324
+ "transformer_blocks.1.ff.linear_out.packed_weight_indices": [
325
+ 7077888
326
+ ],
327
+ "transformer_blocks.1.ff.linear_out.row_norms": [
328
+ 3072
329
+ ],
330
+ "transformer_blocks.1.ff_context.linear_in.packed_weight_indices": [
331
+ 14155776
332
+ ],
333
+ "transformer_blocks.1.ff_context.linear_in.row_norms": [
334
+ 18432
335
+ ],
336
+ "transformer_blocks.1.ff_context.linear_out.packed_weight_indices": [
337
+ 7077888
338
+ ],
339
+ "transformer_blocks.1.ff_context.linear_out.row_norms": [
340
+ 3072
341
+ ],
342
+ "transformer_blocks.2.attn.to_q.packed_weight_indices": [
343
+ 2359296
344
+ ],
345
+ "transformer_blocks.2.attn.to_q.row_norms": [
346
+ 3072
347
+ ],
348
+ "transformer_blocks.2.attn.to_k.packed_weight_indices": [
349
+ 2359296
350
+ ],
351
+ "transformer_blocks.2.attn.to_k.row_norms": [
352
+ 3072
353
+ ],
354
+ "transformer_blocks.2.attn.to_v.packed_weight_indices": [
355
+ 2359296
356
+ ],
357
+ "transformer_blocks.2.attn.to_v.row_norms": [
358
+ 3072
359
+ ],
360
+ "transformer_blocks.2.attn.norm_q.weight": [
361
+ 128
362
+ ],
363
+ "transformer_blocks.2.attn.norm_k.weight": [
364
+ 128
365
+ ],
366
+ "transformer_blocks.2.attn.to_out.0.packed_weight_indices": [
367
+ 2359296
368
+ ],
369
+ "transformer_blocks.2.attn.to_out.0.row_norms": [
370
+ 3072
371
+ ],
372
+ "transformer_blocks.2.attn.norm_added_q.weight": [
373
+ 128
374
+ ],
375
+ "transformer_blocks.2.attn.norm_added_k.weight": [
376
+ 128
377
+ ],
378
+ "transformer_blocks.2.attn.add_q_proj.packed_weight_indices": [
379
+ 2359296
380
+ ],
381
+ "transformer_blocks.2.attn.add_q_proj.row_norms": [
382
+ 3072
383
+ ],
384
+ "transformer_blocks.2.attn.add_k_proj.packed_weight_indices": [
385
+ 2359296
386
+ ],
387
+ "transformer_blocks.2.attn.add_k_proj.row_norms": [
388
+ 3072
389
+ ],
390
+ "transformer_blocks.2.attn.add_v_proj.packed_weight_indices": [
391
+ 2359296
392
+ ],
393
+ "transformer_blocks.2.attn.add_v_proj.row_norms": [
394
+ 3072
395
+ ],
396
+ "transformer_blocks.2.attn.to_add_out.packed_weight_indices": [
397
+ 2359296
398
+ ],
399
+ "transformer_blocks.2.attn.to_add_out.row_norms": [
400
+ 3072
401
+ ],
402
+ "transformer_blocks.2.ff.linear_in.packed_weight_indices": [
403
+ 14155776
404
+ ],
405
+ "transformer_blocks.2.ff.linear_in.row_norms": [
406
+ 18432
407
+ ],
408
+ "transformer_blocks.2.ff.linear_out.packed_weight_indices": [
409
+ 7077888
410
+ ],
411
+ "transformer_blocks.2.ff.linear_out.row_norms": [
412
+ 3072
413
+ ],
414
+ "transformer_blocks.2.ff_context.linear_in.packed_weight_indices": [
415
+ 14155776
416
+ ],
417
+ "transformer_blocks.2.ff_context.linear_in.row_norms": [
418
+ 18432
419
+ ],
420
+ "transformer_blocks.2.ff_context.linear_out.packed_weight_indices": [
421
+ 7077888
422
+ ],
423
+ "transformer_blocks.2.ff_context.linear_out.row_norms": [
424
+ 3072
425
+ ],
426
+ "transformer_blocks.3.attn.to_q.packed_weight_indices": [
427
+ 2359296
428
+ ],
429
+ "transformer_blocks.3.attn.to_q.row_norms": [
430
+ 3072
431
+ ],
432
+ "transformer_blocks.3.attn.to_k.packed_weight_indices": [
433
+ 2359296
434
+ ],
435
+ "transformer_blocks.3.attn.to_k.row_norms": [
436
+ 3072
437
+ ],
438
+ "transformer_blocks.3.attn.to_v.packed_weight_indices": [
439
+ 2359296
440
+ ],
441
+ "transformer_blocks.3.attn.to_v.row_norms": [
442
+ 3072
443
+ ],
444
+ "transformer_blocks.3.attn.norm_q.weight": [
445
+ 128
446
+ ],
447
+ "transformer_blocks.3.attn.norm_k.weight": [
448
+ 128
449
+ ],
450
+ "transformer_blocks.3.attn.to_out.0.packed_weight_indices": [
451
+ 2359296
452
+ ],
453
+ "transformer_blocks.3.attn.to_out.0.row_norms": [
454
+ 3072
455
+ ],
456
+ "transformer_blocks.3.attn.norm_added_q.weight": [
457
+ 128
458
+ ],
459
+ "transformer_blocks.3.attn.norm_added_k.weight": [
460
+ 128
461
+ ],
462
+ "transformer_blocks.3.attn.add_q_proj.packed_weight_indices": [
463
+ 2359296
464
+ ],
465
+ "transformer_blocks.3.attn.add_q_proj.row_norms": [
466
+ 3072
467
+ ],
468
+ "transformer_blocks.3.attn.add_k_proj.packed_weight_indices": [
469
+ 2359296
470
+ ],
471
+ "transformer_blocks.3.attn.add_k_proj.row_norms": [
472
+ 3072
473
+ ],
474
+ "transformer_blocks.3.attn.add_v_proj.packed_weight_indices": [
475
+ 2359296
476
+ ],
477
+ "transformer_blocks.3.attn.add_v_proj.row_norms": [
478
+ 3072
479
+ ],
480
+ "transformer_blocks.3.attn.to_add_out.packed_weight_indices": [
481
+ 2359296
482
+ ],
483
+ "transformer_blocks.3.attn.to_add_out.row_norms": [
484
+ 3072
485
+ ],
486
+ "transformer_blocks.3.ff.linear_in.packed_weight_indices": [
487
+ 14155776
488
+ ],
489
+ "transformer_blocks.3.ff.linear_in.row_norms": [
490
+ 18432
491
+ ],
492
+ "transformer_blocks.3.ff.linear_out.packed_weight_indices": [
493
+ 7077888
494
+ ],
495
+ "transformer_blocks.3.ff.linear_out.row_norms": [
496
+ 3072
497
+ ],
498
+ "transformer_blocks.3.ff_context.linear_in.packed_weight_indices": [
499
+ 14155776
500
+ ],
501
+ "transformer_blocks.3.ff_context.linear_in.row_norms": [
502
+ 18432
503
+ ],
504
+ "transformer_blocks.3.ff_context.linear_out.packed_weight_indices": [
505
+ 7077888
506
+ ],
507
+ "transformer_blocks.3.ff_context.linear_out.row_norms": [
508
+ 3072
509
+ ],
510
+ "transformer_blocks.4.attn.to_q.packed_weight_indices": [
511
+ 2359296
512
+ ],
513
+ "transformer_blocks.4.attn.to_q.row_norms": [
514
+ 3072
515
+ ],
516
+ "transformer_blocks.4.attn.to_k.packed_weight_indices": [
517
+ 2359296
518
+ ],
519
+ "transformer_blocks.4.attn.to_k.row_norms": [
520
+ 3072
521
+ ],
522
+ "transformer_blocks.4.attn.to_v.packed_weight_indices": [
523
+ 2359296
524
+ ],
525
+ "transformer_blocks.4.attn.to_v.row_norms": [
526
+ 3072
527
+ ],
528
+ "transformer_blocks.4.attn.norm_q.weight": [
529
+ 128
530
+ ],
531
+ "transformer_blocks.4.attn.norm_k.weight": [
532
+ 128
533
+ ],
534
+ "transformer_blocks.4.attn.to_out.0.packed_weight_indices": [
535
+ 2359296
536
+ ],
537
+ "transformer_blocks.4.attn.to_out.0.row_norms": [
538
+ 3072
539
+ ],
540
+ "transformer_blocks.4.attn.norm_added_q.weight": [
541
+ 128
542
+ ],
543
+ "transformer_blocks.4.attn.norm_added_k.weight": [
544
+ 128
545
+ ],
546
+ "transformer_blocks.4.attn.add_q_proj.packed_weight_indices": [
547
+ 2359296
548
+ ],
549
+ "transformer_blocks.4.attn.add_q_proj.row_norms": [
550
+ 3072
551
+ ],
552
+ "transformer_blocks.4.attn.add_k_proj.packed_weight_indices": [
553
+ 2359296
554
+ ],
555
+ "transformer_blocks.4.attn.add_k_proj.row_norms": [
556
+ 3072
557
+ ],
558
+ "transformer_blocks.4.attn.add_v_proj.packed_weight_indices": [
559
+ 2359296
560
+ ],
561
+ "transformer_blocks.4.attn.add_v_proj.row_norms": [
562
+ 3072
563
+ ],
564
+ "transformer_blocks.4.attn.to_add_out.packed_weight_indices": [
565
+ 2359296
566
+ ],
567
+ "transformer_blocks.4.attn.to_add_out.row_norms": [
568
+ 3072
569
+ ],
570
+ "transformer_blocks.4.ff.linear_in.packed_weight_indices": [
571
+ 14155776
572
+ ],
573
+ "transformer_blocks.4.ff.linear_in.row_norms": [
574
+ 18432
575
+ ],
576
+ "transformer_blocks.4.ff.linear_out.packed_weight_indices": [
577
+ 7077888
578
+ ],
579
+ "transformer_blocks.4.ff.linear_out.row_norms": [
580
+ 3072
581
+ ],
582
+ "transformer_blocks.4.ff_context.linear_in.packed_weight_indices": [
583
+ 14155776
584
+ ],
585
+ "transformer_blocks.4.ff_context.linear_in.row_norms": [
586
+ 18432
587
+ ],
588
+ "transformer_blocks.4.ff_context.linear_out.packed_weight_indices": [
589
+ 7077888
590
+ ],
591
+ "transformer_blocks.4.ff_context.linear_out.row_norms": [
592
+ 3072
593
+ ],
594
+ "single_transformer_blocks.0.attn.to_qkv_mlp_proj.packed_weight_indices": [
595
+ 21233664
596
+ ],
597
+ "single_transformer_blocks.0.attn.to_qkv_mlp_proj.row_norms": [
598
+ 27648
599
+ ],
600
+ "single_transformer_blocks.0.attn.norm_q.weight": [
601
+ 128
602
+ ],
603
+ "single_transformer_blocks.0.attn.norm_k.weight": [
604
+ 128
605
+ ],
606
+ "single_transformer_blocks.0.attn.to_out.packed_weight_indices": [
607
+ 9437184
608
+ ],
609
+ "single_transformer_blocks.0.attn.to_out.row_norms": [
610
+ 3072
611
+ ],
612
+ "single_transformer_blocks.1.attn.to_qkv_mlp_proj.packed_weight_indices": [
613
+ 21233664
614
+ ],
615
+ "single_transformer_blocks.1.attn.to_qkv_mlp_proj.row_norms": [
616
+ 27648
617
+ ],
618
+ "single_transformer_blocks.1.attn.norm_q.weight": [
619
+ 128
620
+ ],
621
+ "single_transformer_blocks.1.attn.norm_k.weight": [
622
+ 128
623
+ ],
624
+ "single_transformer_blocks.1.attn.to_out.packed_weight_indices": [
625
+ 9437184
626
+ ],
627
+ "single_transformer_blocks.1.attn.to_out.row_norms": [
628
+ 3072
629
+ ],
630
+ "single_transformer_blocks.2.attn.to_qkv_mlp_proj.packed_weight_indices": [
631
+ 21233664
632
+ ],
633
+ "single_transformer_blocks.2.attn.to_qkv_mlp_proj.row_norms": [
634
+ 27648
635
+ ],
636
+ "single_transformer_blocks.2.attn.norm_q.weight": [
637
+ 128
638
+ ],
639
+ "single_transformer_blocks.2.attn.norm_k.weight": [
640
+ 128
641
+ ],
642
+ "single_transformer_blocks.2.attn.to_out.packed_weight_indices": [
643
+ 9437184
644
+ ],
645
+ "single_transformer_blocks.2.attn.to_out.row_norms": [
646
+ 3072
647
+ ],
648
+ "single_transformer_blocks.3.attn.to_qkv_mlp_proj.packed_weight_indices": [
649
+ 21233664
650
+ ],
651
+ "single_transformer_blocks.3.attn.to_qkv_mlp_proj.row_norms": [
652
+ 27648
653
+ ],
654
+ "single_transformer_blocks.3.attn.norm_q.weight": [
655
+ 128
656
+ ],
657
+ "single_transformer_blocks.3.attn.norm_k.weight": [
658
+ 128
659
+ ],
660
+ "single_transformer_blocks.3.attn.to_out.packed_weight_indices": [
661
+ 9437184
662
+ ],
663
+ "single_transformer_blocks.3.attn.to_out.row_norms": [
664
+ 3072
665
+ ],
666
+ "single_transformer_blocks.4.attn.to_qkv_mlp_proj.packed_weight_indices": [
667
+ 21233664
668
+ ],
669
+ "single_transformer_blocks.4.attn.to_qkv_mlp_proj.row_norms": [
670
+ 27648
671
+ ],
672
+ "single_transformer_blocks.4.attn.norm_q.weight": [
673
+ 128
674
+ ],
675
+ "single_transformer_blocks.4.attn.norm_k.weight": [
676
+ 128
677
+ ],
678
+ "single_transformer_blocks.4.attn.to_out.packed_weight_indices": [
679
+ 9437184
680
+ ],
681
+ "single_transformer_blocks.4.attn.to_out.row_norms": [
682
+ 3072
683
+ ],
684
+ "single_transformer_blocks.5.attn.to_qkv_mlp_proj.packed_weight_indices": [
685
+ 21233664
686
+ ],
687
+ "single_transformer_blocks.5.attn.to_qkv_mlp_proj.row_norms": [
688
+ 27648
689
+ ],
690
+ "single_transformer_blocks.5.attn.norm_q.weight": [
691
+ 128
692
+ ],
693
+ "single_transformer_blocks.5.attn.norm_k.weight": [
694
+ 128
695
+ ],
696
+ "single_transformer_blocks.5.attn.to_out.packed_weight_indices": [
697
+ 9437184
698
+ ],
699
+ "single_transformer_blocks.5.attn.to_out.row_norms": [
700
+ 3072
701
+ ],
702
+ "single_transformer_blocks.6.attn.to_qkv_mlp_proj.packed_weight_indices": [
703
+ 21233664
704
+ ],
705
+ "single_transformer_blocks.6.attn.to_qkv_mlp_proj.row_norms": [
706
+ 27648
707
+ ],
708
+ "single_transformer_blocks.6.attn.norm_q.weight": [
709
+ 128
710
+ ],
711
+ "single_transformer_blocks.6.attn.norm_k.weight": [
712
+ 128
713
+ ],
714
+ "single_transformer_blocks.6.attn.to_out.packed_weight_indices": [
715
+ 9437184
716
+ ],
717
+ "single_transformer_blocks.6.attn.to_out.row_norms": [
718
+ 3072
719
+ ],
720
+ "single_transformer_blocks.7.attn.to_qkv_mlp_proj.packed_weight_indices": [
721
+ 21233664
722
+ ],
723
+ "single_transformer_blocks.7.attn.to_qkv_mlp_proj.row_norms": [
724
+ 27648
725
+ ],
726
+ "single_transformer_blocks.7.attn.norm_q.weight": [
727
+ 128
728
+ ],
729
+ "single_transformer_blocks.7.attn.norm_k.weight": [
730
+ 128
731
+ ],
732
+ "single_transformer_blocks.7.attn.to_out.packed_weight_indices": [
733
+ 9437184
734
+ ],
735
+ "single_transformer_blocks.7.attn.to_out.row_norms": [
736
+ 3072
737
+ ],
738
+ "single_transformer_blocks.8.attn.to_qkv_mlp_proj.packed_weight_indices": [
739
+ 21233664
740
+ ],
741
+ "single_transformer_blocks.8.attn.to_qkv_mlp_proj.row_norms": [
742
+ 27648
743
+ ],
744
+ "single_transformer_blocks.8.attn.norm_q.weight": [
745
+ 128
746
+ ],
747
+ "single_transformer_blocks.8.attn.norm_k.weight": [
748
+ 128
749
+ ],
750
+ "single_transformer_blocks.8.attn.to_out.packed_weight_indices": [
751
+ 9437184
752
+ ],
753
+ "single_transformer_blocks.8.attn.to_out.row_norms": [
754
+ 3072
755
+ ],
756
+ "single_transformer_blocks.9.attn.to_qkv_mlp_proj.packed_weight_indices": [
757
+ 21233664
758
+ ],
759
+ "single_transformer_blocks.9.attn.to_qkv_mlp_proj.row_norms": [
760
+ 27648
761
+ ],
762
+ "single_transformer_blocks.9.attn.norm_q.weight": [
763
+ 128
764
+ ],
765
+ "single_transformer_blocks.9.attn.norm_k.weight": [
766
+ 128
767
+ ],
768
+ "single_transformer_blocks.9.attn.to_out.packed_weight_indices": [
769
+ 9437184
770
+ ],
771
+ "single_transformer_blocks.9.attn.to_out.row_norms": [
772
+ 3072
773
+ ],
774
+ "single_transformer_blocks.10.attn.to_qkv_mlp_proj.packed_weight_indices": [
775
+ 21233664
776
+ ],
777
+ "single_transformer_blocks.10.attn.to_qkv_mlp_proj.row_norms": [
778
+ 27648
779
+ ],
780
+ "single_transformer_blocks.10.attn.norm_q.weight": [
781
+ 128
782
+ ],
783
+ "single_transformer_blocks.10.attn.norm_k.weight": [
784
+ 128
785
+ ],
786
+ "single_transformer_blocks.10.attn.to_out.packed_weight_indices": [
787
+ 9437184
788
+ ],
789
+ "single_transformer_blocks.10.attn.to_out.row_norms": [
790
+ 3072
791
+ ],
792
+ "single_transformer_blocks.11.attn.to_qkv_mlp_proj.packed_weight_indices": [
793
+ 21233664
794
+ ],
795
+ "single_transformer_blocks.11.attn.to_qkv_mlp_proj.row_norms": [
796
+ 27648
797
+ ],
798
+ "single_transformer_blocks.11.attn.norm_q.weight": [
799
+ 128
800
+ ],
801
+ "single_transformer_blocks.11.attn.norm_k.weight": [
802
+ 128
803
+ ],
804
+ "single_transformer_blocks.11.attn.to_out.packed_weight_indices": [
805
+ 9437184
806
+ ],
807
+ "single_transformer_blocks.11.attn.to_out.row_norms": [
808
+ 3072
809
+ ],
810
+ "single_transformer_blocks.12.attn.to_qkv_mlp_proj.packed_weight_indices": [
811
+ 21233664
812
+ ],
813
+ "single_transformer_blocks.12.attn.to_qkv_mlp_proj.row_norms": [
814
+ 27648
815
+ ],
816
+ "single_transformer_blocks.12.attn.norm_q.weight": [
817
+ 128
818
+ ],
819
+ "single_transformer_blocks.12.attn.norm_k.weight": [
820
+ 128
821
+ ],
822
+ "single_transformer_blocks.12.attn.to_out.packed_weight_indices": [
823
+ 9437184
824
+ ],
825
+ "single_transformer_blocks.12.attn.to_out.row_norms": [
826
+ 3072
827
+ ],
828
+ "single_transformer_blocks.13.attn.to_qkv_mlp_proj.packed_weight_indices": [
829
+ 21233664
830
+ ],
831
+ "single_transformer_blocks.13.attn.to_qkv_mlp_proj.row_norms": [
832
+ 27648
833
+ ],
834
+ "single_transformer_blocks.13.attn.norm_q.weight": [
835
+ 128
836
+ ],
837
+ "single_transformer_blocks.13.attn.norm_k.weight": [
838
+ 128
839
+ ],
840
+ "single_transformer_blocks.13.attn.to_out.packed_weight_indices": [
841
+ 9437184
842
+ ],
843
+ "single_transformer_blocks.13.attn.to_out.row_norms": [
844
+ 3072
845
+ ],
846
+ "single_transformer_blocks.14.attn.to_qkv_mlp_proj.packed_weight_indices": [
847
+ 21233664
848
+ ],
849
+ "single_transformer_blocks.14.attn.to_qkv_mlp_proj.row_norms": [
850
+ 27648
851
+ ],
852
+ "single_transformer_blocks.14.attn.norm_q.weight": [
853
+ 128
854
+ ],
855
+ "single_transformer_blocks.14.attn.norm_k.weight": [
856
+ 128
857
+ ],
858
+ "single_transformer_blocks.14.attn.to_out.packed_weight_indices": [
859
+ 9437184
860
+ ],
861
+ "single_transformer_blocks.14.attn.to_out.row_norms": [
862
+ 3072
863
+ ],
864
+ "single_transformer_blocks.15.attn.to_qkv_mlp_proj.packed_weight_indices": [
865
+ 21233664
866
+ ],
867
+ "single_transformer_blocks.15.attn.to_qkv_mlp_proj.row_norms": [
868
+ 27648
869
+ ],
870
+ "single_transformer_blocks.15.attn.norm_q.weight": [
871
+ 128
872
+ ],
873
+ "single_transformer_blocks.15.attn.norm_k.weight": [
874
+ 128
875
+ ],
876
+ "single_transformer_blocks.15.attn.to_out.packed_weight_indices": [
877
+ 9437184
878
+ ],
879
+ "single_transformer_blocks.15.attn.to_out.row_norms": [
880
+ 3072
881
+ ],
882
+ "single_transformer_blocks.16.attn.to_qkv_mlp_proj.packed_weight_indices": [
883
+ 21233664
884
+ ],
885
+ "single_transformer_blocks.16.attn.to_qkv_mlp_proj.row_norms": [
886
+ 27648
887
+ ],
888
+ "single_transformer_blocks.16.attn.norm_q.weight": [
889
+ 128
890
+ ],
891
+ "single_transformer_blocks.16.attn.norm_k.weight": [
892
+ 128
893
+ ],
894
+ "single_transformer_blocks.16.attn.to_out.packed_weight_indices": [
895
+ 9437184
896
+ ],
897
+ "single_transformer_blocks.16.attn.to_out.row_norms": [
898
+ 3072
899
+ ],
900
+ "single_transformer_blocks.17.attn.to_qkv_mlp_proj.packed_weight_indices": [
901
+ 21233664
902
+ ],
903
+ "single_transformer_blocks.17.attn.to_qkv_mlp_proj.row_norms": [
904
+ 27648
905
+ ],
906
+ "single_transformer_blocks.17.attn.norm_q.weight": [
907
+ 128
908
+ ],
909
+ "single_transformer_blocks.17.attn.norm_k.weight": [
910
+ 128
911
+ ],
912
+ "single_transformer_blocks.17.attn.to_out.packed_weight_indices": [
913
+ 9437184
914
+ ],
915
+ "single_transformer_blocks.17.attn.to_out.row_norms": [
916
+ 3072
917
+ ],
918
+ "single_transformer_blocks.18.attn.to_qkv_mlp_proj.packed_weight_indices": [
919
+ 21233664
920
+ ],
921
+ "single_transformer_blocks.18.attn.to_qkv_mlp_proj.row_norms": [
922
+ 27648
923
+ ],
924
+ "single_transformer_blocks.18.attn.norm_q.weight": [
925
+ 128
926
+ ],
927
+ "single_transformer_blocks.18.attn.norm_k.weight": [
928
+ 128
929
+ ],
930
+ "single_transformer_blocks.18.attn.to_out.packed_weight_indices": [
931
+ 9437184
932
+ ],
933
+ "single_transformer_blocks.18.attn.to_out.row_norms": [
934
+ 3072
935
+ ],
936
+ "single_transformer_blocks.19.attn.to_qkv_mlp_proj.packed_weight_indices": [
937
+ 21233664
938
+ ],
939
+ "single_transformer_blocks.19.attn.to_qkv_mlp_proj.row_norms": [
940
+ 27648
941
+ ],
942
+ "single_transformer_blocks.19.attn.norm_q.weight": [
943
+ 128
944
+ ],
945
+ "single_transformer_blocks.19.attn.norm_k.weight": [
946
+ 128
947
+ ],
948
+ "single_transformer_blocks.19.attn.to_out.packed_weight_indices": [
949
+ 9437184
950
+ ],
951
+ "single_transformer_blocks.19.attn.to_out.row_norms": [
952
+ 3072
953
+ ],
954
+ "norm_out.linear.weight": [
955
+ 6144,
956
+ 3072
957
+ ],
958
+ "proj_out.weight": [
959
+ 128,
960
+ 3072
961
+ ]
962
+ },
963
+ "checksums": {
964
+ "model.safetensors": "b5c4654d1a0a00d47a40b75953298796da8249726b3ad7f58700de615585ec0b",
965
+ "orbitquant_codebooks.safetensors": "5767b046e6b0721c56bad6306657708f60f3d779721920f8ddaab51e34a9cd10",
966
+ "orbitquant_rotations.safetensors": "3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17",
967
+ "quantization_config.json": "0e689c046ea3fbf76a8128b71b4686f11196ba7b835fe5cafdc71e7498801b7f",
968
+ "model_index.json": "a49141221a603fae77e65aab145dc2e6b6794592eb1e2e7d20a6d8d85414a4e5",
969
+ "prompts.json": "9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd",
970
+ "benchmark/summary.json": "34a4410c42ca64bd864be4b9f9abef076ad2f2be9c05d9b65e74b83489b03406",
971
+ "benchmark/original.metrics.jsonl": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
972
+ "benchmark/orbitquant.metrics.jsonl": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
973
+ "benchmark/original.metrics.csv": "90d54eb9870d76c7e992e1ca94341cc7a58c931a079162ef1774b7f2b571ac73",
974
+ "benchmark/orbitquant.metrics.csv": "90d54eb9870d76c7e992e1ca94341cc7a58c931a079162ef1774b7f2b571ac73",
975
+ "assets/.gitkeep": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
976
+ }
977
+ }
orbitquant_rotations.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17
3
+ size 419020
prompts.json ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "prompt_pack": "image_visual_v1",
3
+ "media_type": "image",
4
+ "target_policy": "flux2",
5
+ "prompts": [
6
+ {
7
+ "id": "simple-object",
8
+ "category": "simple_object",
9
+ "prompt": "A red ceramic mug on a wooden desk, soft daylight, shallow depth of field"
10
+ },
11
+ {
12
+ "id": "two-object-composition",
13
+ "category": "two_object_composition",
14
+ "prompt": "Two robots playing chess in a quiet library, one silver and one matte black"
15
+ },
16
+ {
17
+ "id": "counting",
18
+ "category": "counting",
19
+ "prompt": "Five glass marbles arranged in a straight line on white paper"
20
+ },
21
+ {
22
+ "id": "color-binding",
23
+ "category": "color_binding",
24
+ "prompt": "A blue cube beside a yellow sphere and a green pyramid on a gray table"
25
+ },
26
+ {
27
+ "id": "spatial-relationship",
28
+ "category": "spatial_relationship",
29
+ "prompt": "A blue cube to the left of a yellow sphere, both centered in the frame"
30
+ },
31
+ {
32
+ "id": "long-prompt",
33
+ "category": "long_prompt",
34
+ "prompt": "A precise editorial photo of a compact robotics workbench with labeled tools, a small oscilloscope, coiled cables, a half-assembled drone, and a warm desk lamp; the scene should remain readable without clutter"
35
+ },
36
+ {
37
+ "id": "english-text-rendering",
38
+ "category": "text_rendering",
39
+ "prompt": "A clean street sign with the exact text \"ORBIT QUANT\""
40
+ },
41
+ {
42
+ "id": "cyrillic-text-rendering",
43
+ "category": "text_rendering",
44
+ "prompt": "A handwritten note in Cyrillic that says \"\u041a\u0412\u0410\u041d\u0422\u041e\u0412\u0410\u041d\u0418\u0415\""
45
+ },
46
+ {
47
+ "id": "style-heavy",
48
+ "category": "style",
49
+ "prompt": "A crowded train platform at sunset, documentary photo style, high dynamic range, natural grain, realistic faces, detailed architecture"
50
+ },
51
+ {
52
+ "id": "occlusion-reflection",
53
+ "category": "occlusion_reflection",
54
+ "prompt": "A reflective chrome teapot partly hidden behind orange flowers, with the room visible as a coherent reflection on the metal"
55
+ }
56
+ ]
57
+ }
quantization_config.json ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "quant_method": "orbitquant",
3
+ "weight_bits": 2,
4
+ "activation_bits": 4,
5
+ "rotation": "rpbh",
6
+ "rotation_seed": 0,
7
+ "block_size": "paper",
8
+ "codebook": "lloyd_max",
9
+ "codebook_dtype": "float32",
10
+ "row_norm_dtype": "bfloat16",
11
+ "activation_norm_dtype": "float32",
12
+ "activation_eps": 1e-12,
13
+ "weight_pack_dtype": "uint8",
14
+ "target_policy": "flux2",
15
+ "adaln_policy": "int4_rtn",
16
+ "adaln_group_size": 64,
17
+ "modules_to_not_convert": [],
18
+ "modules_dtype_dict": {},
19
+ "artifact_format_version": 1,
20
+ "runtime_mode": "dequant_bf16",
21
+ "activation_kernel_backend": "triton_cuda"
22
+ }