Diffusers
Safetensors
OrbitQuantComponentArtifact
orbitquant
quantized
diffusion-transformer
8-bit precision
Instructions to use WaveCut/FLUX.2-klein-4B-OrbitQuant-W4A4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use WaveCut/FLUX.2-klein-4B-OrbitQuant-W4A4 with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("WaveCut/FLUX.2-klein-4B-OrbitQuant-W4A4", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
Replace with paper-conformant codebook v2 artifact
Browse files- README.md +3 -20
- SHA256SUMS +7 -8
- assets/image_generation_comparison_matrix.webp +0 -3
- benchmark/summary.json +8 -96
- model.safetensors +1 -1
- model_index.json +2 -1
- orbitquant_codebooks.safetensors +1 -1
- orbitquant_manifest.json +7 -8
- quantization_config.json +5 -4
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: `
|
| 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: `
|
| 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 |
-
|
| 116 |
-
|
| 117 |
-

|
| 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 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
fd5327032b7e41dcd6abcd5fb6cffa467c4fca2d33e170e1cbb94f8e38098bba orbitquant_manifest.json
|
| 8 |
3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17 orbitquant_rotations.safetensors
|
| 9 |
9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd prompts.json
|
| 10 |
-
|
|
|
|
| 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
|
benchmark/summary.json
CHANGED
|
@@ -1,112 +1,24 @@
|
|
| 1 |
{
|
| 2 |
-
"status": "
|
| 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": "
|
| 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":
|
| 16 |
-
"orbitquant_seconds":
|
| 17 |
-
"adaln_seconds":
|
| 18 |
-
"device_transfer_seconds": 0.
|
| 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:
|
| 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": "
|
| 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:
|
| 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":
|
| 16 |
"row_norm_dtype": "bfloat16",
|
| 17 |
"runtime_mode": "auto_fused",
|
| 18 |
-
"activation_kernel_backend": "
|
| 19 |
"activation_eps": 1e-10,
|
| 20 |
"adaln_group_size": 64,
|
| 21 |
"quantization_device": "cuda",
|
|
@@ -966,13 +966,12 @@
|
|
| 966 |
]
|
| 967 |
},
|
| 968 |
"checksums": {
|
| 969 |
-
"
|
| 970 |
-
"
|
| 971 |
-
"
|
| 972 |
-
"
|
| 973 |
-
"orbitquant_codebooks.safetensors": "818e4de6a927335c47bb7d148f23a05dda61a3abb6e8f67cd8b9510960bdcf4c",
|
| 974 |
"orbitquant_rotations.safetensors": "3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17",
|
| 975 |
"prompts.json": "9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd",
|
| 976 |
-
"quantization_config.json": "
|
| 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": "
|
| 22 |
-
"packed_matmul_block_m":
|
| 23 |
"packed_matmul_block_n": 64,
|
| 24 |
-
"packed_matmul_block_k":
|
| 25 |
-
"packed_matmul_num_warps":
|
| 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 |
}
|