WaveCut commited on
Commit
65e8a90
·
verified ·
1 Parent(s): 94ccd0d

Replace with paper-conformant codebook v2 artifact

Browse files
README.md CHANGED
@@ -70,27 +70,12 @@ Use these settings when comparing this artifact against the BF16 source model or
70
  - Release-grade paper metrics: not applicable to this extra target.
71
  - The model card reports artifact-level validation status only.
72
 
73
- ## Native Validation Evidence
74
-
75
- The compact benchmark summary records native BF16-vs-OrbitQuant evidence for the comparison matrix below. Detailed per-sample generation records are retained outside this compact artifact.
76
-
77
- | Evidence | Value |
78
- | --- | --- |
79
- | Comparison matrix | `assets/image_generation_comparison_matrix.webp` |
80
- | Paired prompt/seed count | `1` |
81
- | BF16 source generated samples | `1` |
82
- | BF16 source generated frames | `0` |
83
- | BF16 source nonempty outputs | `1` |
84
- | OrbitQuant generated samples | `1` |
85
- | OrbitQuant generated frames | `0` |
86
- | OrbitQuant nonempty outputs | `1` |
87
-
88
  ## Quantization
89
 
90
  - Method: `orbitquant`
91
  - Bits: `W4A4`
92
  - Runtime mode: `auto_fused`
93
- - Activation kernel backend: `triton_cuda`
94
  - Activation normalization epsilon: `1e-10`
95
  - Quantization device: `cuda`
96
  - Weight quantization backend: `triton_cuda`
@@ -103,7 +88,7 @@ The compact benchmark summary records native BF16-vs-OrbitQuant evidence for the
103
  - Block size: `paper`
104
  - Block size policy: `largest_power_of_two_dividing_dim`
105
  - Codebook: `lloyd_max`
106
- - Codebook version: `1`
107
  - Quantized transformer modules: `100`
108
  - AdaLN INT4 modules: `3`
109
  - Skipped modules: `6`
@@ -112,9 +97,7 @@ The compact benchmark summary records native BF16-vs-OrbitQuant evidence for the
112
 
113
  ## Visual Comparison
114
 
115
- The following assets are stored in this artifact and compare the BF16 base generation against the OrbitQuant generation with the same prompt and seed.
116
-
117
- ![assets/image_generation_comparison_matrix.webp](assets/image_generation_comparison_matrix.webp)
118
 
119
  ## Source
120
 
 
70
  - Release-grade paper metrics: not applicable to this extra target.
71
  - The model card reports artifact-level validation status only.
72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
73
  ## Quantization
74
 
75
  - Method: `orbitquant`
76
  - Bits: `W4A4`
77
  - Runtime mode: `auto_fused`
78
+ - Activation kernel backend: `auto`
79
  - Activation normalization epsilon: `1e-10`
80
  - Quantization device: `cuda`
81
  - Weight quantization backend: `triton_cuda`
 
88
  - Block size: `paper`
89
  - Block size policy: `largest_power_of_two_dividing_dim`
90
  - Codebook: `lloyd_max`
91
+ - Codebook version: `2`
92
  - Quantized transformer modules: `100`
93
  - AdaLN INT4 modules: `3`
94
  - Skipped modules: `6`
 
97
 
98
  ## Visual Comparison
99
 
100
+ Validation status: comparison asset missing. This artifact does not include a generation comparison matrix.
 
 
101
 
102
  ## Source
103
 
SHA256SUMS CHANGED
@@ -1,10 +1,9 @@
1
- dd5da716ec3fbf61ba336e02399bbd6d57da2d27cc6f549179ee417dd46a19e6 README.md
2
- 4983dedd382a6ca16289cac20ceb4a54d0eef84d5195673bddde99c8c6f32012 assets/image_generation_comparison_matrix.webp
3
- d39ce2afa1fb562244a9acb8712e2645aca4b5da60ecc3115d5e299d6c8ec1ad benchmark/summary.json
4
- f03eaef37e04ff47af0935607a334bb80d26a3cf1bd9dfdc11a675b56bedfedc model.safetensors
5
- f99a4281801bd8d56a9e0eea0ab2f6c450d3d4cf425b5db238e49f6be15b6228 model_index.json
6
- 818e4de6a927335c47bb7d148f23a05dda61a3abb6e8f67cd8b9510960bdcf4c orbitquant_codebooks.safetensors
7
- fd5327032b7e41dcd6abcd5fb6cffa467c4fca2d33e170e1cbb94f8e38098bba orbitquant_manifest.json
8
  3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17 orbitquant_rotations.safetensors
9
  9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd prompts.json
10
- d9207a741f4c353143ce5038f06349679c4c89e38e334b3f0773ed1e4c5d356a quantization_config.json
 
1
+ 80c3fccb41c682f2dda12edbce8459ba1029486e27cd4a1dc63c18a5eb62bb45 README.md
2
+ 6b9671bc41754292ac58aa84fbb263f2c82e10f57d7d34c04dfa63859709ffaf benchmark/summary.json
3
+ bd85d0a739b4f5ddf2e3a109c8163ab019ba4a7f68ed07b4ff4d3362b41c62a7 model.safetensors
4
+ 4cf21005259004e390579286648418945378ef08022e396d14e165ae13672f2e model_index.json
5
+ e4f2fff365e0b0b6019117abb39ba2817c92173521be726e02919d6456929411 orbitquant_codebooks.safetensors
6
+ 68b747b9002db6f058826a86996c5508f8526fcd1a7e6091fec1e24e9689e9ee orbitquant_manifest.json
 
7
  3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17 orbitquant_rotations.safetensors
8
  9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd prompts.json
9
+ d90798b18112e3a3b6af9e17913426ee6b500a3ab57e04b6df4cedb306a4cdd8 quantization_config.json
assets/image_generation_comparison_matrix.webp DELETED

Git LFS Details

  • SHA256: 4983dedd382a6ca16289cac20ceb4a54d0eef84d5195673bddde99c8c6f32012
  • Pointer size: 130 Bytes
  • Size of remote file: 55.7 kB
benchmark/summary.json CHANGED
@@ -1,112 +1,24 @@
1
  {
2
- "status": "metrics_recorded",
3
  "source_model_id": "black-forest-labs/FLUX.2-klein-4B",
4
  "source_revision": "e7b7dc27f91deacad38e78976d1f2b499d76a294",
5
  "weight_bits": 4,
6
  "activation_bits": 4,
 
7
  "target_policy": "flux2",
8
  "runtime_mode": "auto_fused",
9
- "activation_kernel_backend": "triton_cuda",
10
  "activation_eps": 1e-10,
11
  "quantization_device": "cuda",
12
  "weight_quantization_backend": "triton_cuda",
13
  "quantization_staging_mode": "streaming",
14
  "synchronize_per_module": false,
15
- "quantization_elapsed_seconds": 3.4752471749670804,
16
- "orbitquant_seconds": 2.135687416419387,
17
- "adaln_seconds": 0.4508278868161142,
18
- "device_transfer_seconds": 0.882997278124094,
19
  "module_device_transfer_count": 103,
20
- "metrics": {
21
- "orbitquant": {
22
- "records": 1,
23
- "latest_metrics": {
24
- "generated_samples": 1,
25
- "wall_time_seconds": 8.579359667841345,
26
- "peak_vram_bytes": 13604859904
27
- }
28
- },
29
- "original": {
30
- "records": 1,
31
- "latest_metrics": {
32
- "generated_samples": 1,
33
- "wall_time_seconds": 1.9006238640286028,
34
- "peak_vram_bytes": 18602007552
35
- }
36
- }
37
- },
38
- "native_smoke": {
39
- "proof_format": "orbitquant-native-smoke-v1",
40
- "comparison_asset_path": "assets/image_generation_comparison_matrix.webp",
41
- "paired_prompt_seed_count": 1,
42
- "paired_prompt_seed_keys": [
43
- [
44
- "flux2-native",
45
- "0",
46
- "simple-object"
47
- ]
48
- ],
49
- "splits": {
50
- "original": {
51
- "records": 1,
52
- "generated_samples": 1,
53
- "generated_frames": 0,
54
- "nonempty_output_count": 1,
55
- "seeds": [
56
- "0"
57
- ],
58
- "prompt_ids": [
59
- "simple-object"
60
- ],
61
- "pair_keys": [
62
- [
63
- "flux2-native",
64
- "0",
65
- "simple-object"
66
- ]
67
- ],
68
- "native_settings": [
69
- {
70
- "frames": null,
71
- "guidance": 1.0,
72
- "height": 1024,
73
- "steps": 4,
74
- "suite": "flux2-native",
75
- "width": 1024
76
- }
77
- ]
78
- },
79
- "orbitquant": {
80
- "records": 1,
81
- "generated_samples": 1,
82
- "generated_frames": 0,
83
- "nonempty_output_count": 1,
84
- "seeds": [
85
- "0"
86
- ],
87
- "prompt_ids": [
88
- "simple-object"
89
- ],
90
- "pair_keys": [
91
- [
92
- "flux2-native",
93
- "0",
94
- "simple-object"
95
- ]
96
- ],
97
- "native_settings": [
98
- {
99
- "frames": null,
100
- "guidance": 1.0,
101
- "height": 1024,
102
- "steps": 4,
103
- "suite": "flux2-native",
104
- "width": 1024
105
- }
106
- ]
107
- }
108
- }
109
- },
110
  "published_summary": "compact",
111
  "raw_generation_records": "local-only"
112
  }
 
1
  {
2
+ "status": "not_run",
3
  "source_model_id": "black-forest-labs/FLUX.2-klein-4B",
4
  "source_revision": "e7b7dc27f91deacad38e78976d1f2b499d76a294",
5
  "weight_bits": 4,
6
  "activation_bits": 4,
7
+ "codebook_version": 2,
8
  "target_policy": "flux2",
9
  "runtime_mode": "auto_fused",
10
+ "activation_kernel_backend": "auto",
11
  "activation_eps": 1e-10,
12
  "quantization_device": "cuda",
13
  "weight_quantization_backend": "triton_cuda",
14
  "quantization_staging_mode": "streaming",
15
  "synchronize_per_module": false,
16
+ "quantization_elapsed_seconds": 6.446110541932285,
17
+ "orbitquant_seconds": 3.5517221023328602,
18
+ "adaln_seconds": 2.1711772345006466,
19
+ "device_transfer_seconds": 0.717645316850394,
20
  "module_device_transfer_count": 103,
21
+ "metrics": {},
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
22
  "published_summary": "compact",
23
  "raw_generation_records": "local-only"
24
  }
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f03eaef37e04ff47af0935607a334bb80d26a3cf1bd9dfdc11a675b56bedfedc
3
  size 2024361976
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bd85d0a739b4f5ddf2e3a109c8163ab019ba4a7f68ed07b4ff4d3362b41c62a7
3
  size 2024361976
model_index.json CHANGED
@@ -13,9 +13,10 @@
13
  "rotations": "orbitquant_rotations.safetensors",
14
  "weight_bits": 4,
15
  "activation_bits": 4,
 
16
  "target_policy": "flux2",
17
  "runtime_mode": "auto_fused",
18
- "activation_kernel_backend": "triton_cuda",
19
  "activation_eps": 1e-10,
20
  "quantization_device": "cuda",
21
  "weight_quantization_backend": "triton_cuda",
 
13
  "rotations": "orbitquant_rotations.safetensors",
14
  "weight_bits": 4,
15
  "activation_bits": 4,
16
+ "codebook_version": 2,
17
  "target_policy": "flux2",
18
  "runtime_mode": "auto_fused",
19
+ "activation_kernel_backend": "auto",
20
  "activation_eps": 1e-10,
21
  "quantization_device": "cuda",
22
  "weight_quantization_backend": "triton_cuda",
orbitquant_codebooks.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:818e4de6a927335c47bb7d148f23a05dda61a3abb6e8f67cd8b9510960bdcf4c
3
  size 868
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e4f2fff365e0b0b6019117abb39ba2817c92173521be726e02919d6456929411
3
  size 868
orbitquant_manifest.json CHANGED
@@ -12,10 +12,10 @@
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": "auto_fused",
18
- "activation_kernel_backend": "triton_cuda",
19
  "activation_eps": 1e-10,
20
  "adaln_group_size": 64,
21
  "quantization_device": "cuda",
@@ -966,13 +966,12 @@
966
  ]
967
  },
968
  "checksums": {
969
- "assets/image_generation_comparison_matrix.webp": "4983dedd382a6ca16289cac20ceb4a54d0eef84d5195673bddde99c8c6f32012",
970
- "benchmark/summary.json": "d39ce2afa1fb562244a9acb8712e2645aca4b5da60ecc3115d5e299d6c8ec1ad",
971
- "model.safetensors": "f03eaef37e04ff47af0935607a334bb80d26a3cf1bd9dfdc11a675b56bedfedc",
972
- "model_index.json": "f99a4281801bd8d56a9e0eea0ab2f6c450d3d4cf425b5db238e49f6be15b6228",
973
- "orbitquant_codebooks.safetensors": "818e4de6a927335c47bb7d148f23a05dda61a3abb6e8f67cd8b9510960bdcf4c",
974
  "orbitquant_rotations.safetensors": "3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17",
975
  "prompts.json": "9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd",
976
- "quantization_config.json": "d9207a741f4c353143ce5038f06349679c4c89e38e334b3f0773ed1e4c5d356a"
977
  }
978
  }
 
12
  "block_size": "paper",
13
  "block_size_policy": "largest_power_of_two_dividing_dim",
14
  "codebook": "lloyd_max",
15
+ "codebook_version": 2,
16
  "row_norm_dtype": "bfloat16",
17
  "runtime_mode": "auto_fused",
18
+ "activation_kernel_backend": "auto",
19
  "activation_eps": 1e-10,
20
  "adaln_group_size": 64,
21
  "quantization_device": "cuda",
 
966
  ]
967
  },
968
  "checksums": {
969
+ "benchmark/summary.json": "6b9671bc41754292ac58aa84fbb263f2c82e10f57d7d34c04dfa63859709ffaf",
970
+ "model.safetensors": "bd85d0a739b4f5ddf2e3a109c8163ab019ba4a7f68ed07b4ff4d3362b41c62a7",
971
+ "model_index.json": "4cf21005259004e390579286648418945378ef08022e396d14e165ae13672f2e",
972
+ "orbitquant_codebooks.safetensors": "e4f2fff365e0b0b6019117abb39ba2817c92173521be726e02919d6456929411",
 
973
  "orbitquant_rotations.safetensors": "3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17",
974
  "prompts.json": "9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd",
975
+ "quantization_config.json": "d90798b18112e3a3b6af9e17913426ee6b500a3ab57e04b6df4cedb306a4cdd8"
976
  }
977
  }
quantization_config.json CHANGED
@@ -6,6 +6,7 @@
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",
@@ -18,9 +19,9 @@
18
  "modules_dtype_dict": {},
19
  "artifact_format_version": 1,
20
  "runtime_mode": "auto_fused",
21
- "activation_kernel_backend": "triton_cuda",
22
- "packed_matmul_block_m": 32,
23
  "packed_matmul_block_n": 64,
24
- "packed_matmul_block_k": 64,
25
- "packed_matmul_num_warps": 8
26
  }
 
6
  "rotation_seed": 0,
7
  "block_size": "paper",
8
  "codebook": "lloyd_max",
9
+ "codebook_version": 2,
10
  "codebook_dtype": "float32",
11
  "row_norm_dtype": "bfloat16",
12
  "activation_norm_dtype": "float32",
 
19
  "modules_dtype_dict": {},
20
  "artifact_format_version": 1,
21
  "runtime_mode": "auto_fused",
22
+ "activation_kernel_backend": "auto",
23
+ "packed_matmul_block_m": 64,
24
  "packed_matmul_block_n": 64,
25
+ "packed_matmul_block_k": 128,
26
+ "packed_matmul_num_warps": 4
27
  }