WaveCut commited on
Commit
ad06b4d
·
verified ·
1 Parent(s): cc35a7e

Replace artifact with paper-correct codebook v2 and native proof

Browse files
README.md CHANGED
@@ -90,7 +90,7 @@ The compact benchmark summary records native BF16-vs-OrbitQuant evidence for the
90
  - Method: `orbitquant`
91
  - Bits: `W2A3`
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 +103,7 @@ 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: `238`
108
  - AdaLN INT4 modules: `32`
109
  - Skipped modules: `6`
 
90
  - Method: `orbitquant`
91
  - Bits: `W2A3`
92
  - Runtime mode: `auto_fused`
93
+ - Activation kernel backend: `auto`
94
  - Activation normalization epsilon: `1e-10`
95
  - Quantization device: `cuda`
96
  - Weight quantization backend: `triton_cuda`
 
103
  - Block size: `paper`
104
  - Block size policy: `largest_power_of_two_dividing_dim`
105
  - Codebook: `lloyd_max`
106
+ - Codebook version: `2`
107
  - Quantized transformer modules: `238`
108
  - AdaLN INT4 modules: `32`
109
  - Skipped modules: `6`
SHA256SUMS CHANGED
@@ -1,10 +1,10 @@
1
- 0ccb2cbee8bf87fff9b604ae7db0eb9e9a821d0cfc21ceeef0ded4bb6107b374 README.md
2
- 87f719a534ba76c19031c82ca761c2fea4e40615f0d1a23f37dd29556ee102aa assets/image_generation_comparison_matrix.webp
3
- 2bda4afcd7266fbe4e0d184690605dd63f9546fc3a1ec0ca88e117387bf6bdbe benchmark/summary.json
4
- aea099a0445d60c272a2c45a8c4e88fd4b3cedf7181a4b3600873104eb7a7941 model.safetensors
5
- 54320a7d5779945db1e5e50b43c11dea40ec5dff74c9dd31a419bda6419160bd model_index.json
6
- 04da03ff11a91f322d72c88fe5686e7ebf66972de6ed5ab606e5282e682a332f orbitquant_codebooks.safetensors
7
- cdddf9d2aa7c409cd59d32d14c6d037b4336706bb9f422b45e1a069bda9127fc orbitquant_manifest.json
8
  e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825 orbitquant_rotations.safetensors
9
  65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af prompts.json
10
- a0320d9f96882ee890452ea222e0d4c36db3d5ad57382fc8908a70251d91f614 quantization_config.json
 
1
+ 18ac533ceb7b92c0e1b6ae51376dc405ee7a20981143cc257083bb09bd86bdfc README.md
2
+ f7e324ad238c63d0d25568acb0d4cee76bf9ddfd3c43f22845ad2138728daef7 assets/image_generation_comparison_matrix.webp
3
+ 59f01ffd627235229bc83868ac999ce280ccca656a138603baea85ea32929075 benchmark/summary.json
4
+ aa8e04810221025dc262f66632975bdbf8466b341ad19db61ba4ad4f174f402f model.safetensors
5
+ 61be70974b031f683d3af31c38322999f93c0788909aa7e33b91499962db225a model_index.json
6
+ 3c06a807bddc6af0246586ded2276ec8dca0e309543a9701de1b6872fb8b6b31 orbitquant_codebooks.safetensors
7
+ d0aa18e68101aa217b3bcba80382ebca0ef1a7919b000da5f6f9512e535c41b9 orbitquant_manifest.json
8
  e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825 orbitquant_rotations.safetensors
9
  65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af prompts.json
10
+ d5fbd286ff541dc629e0a2b5a216580674faaf7a3814679b0fb7e4a2122cc67b quantization_config.json
assets/image_generation_comparison_matrix.webp CHANGED

Git LFS Details

  • SHA256: 87f719a534ba76c19031c82ca761c2fea4e40615f0d1a23f37dd29556ee102aa
  • Pointer size: 130 Bytes
  • Size of remote file: 73.5 kB

Git LFS Details

  • SHA256: f7e324ad238c63d0d25568acb0d4cee76bf9ddfd3c43f22845ad2138728daef7
  • Pointer size: 130 Bytes
  • Size of remote file: 65.4 kB
benchmark/summary.json CHANGED
@@ -1,42 +1,27 @@
1
  {
2
- "status": "metrics_recorded",
3
  "source_model_id": "Tongyi-MAI/Z-Image-Turbo",
4
  "source_revision": "f332072aa78be7aecdf3ee76d5c247082da564a6",
5
  "weight_bits": 2,
6
  "activation_bits": 3,
 
7
  "target_policy": "z_image",
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": 2.160409632138908,
16
- "orbitquant_seconds": 0.7088581034913659,
17
- "adaln_seconds": 0.011828072369098663,
18
- "device_transfer_seconds": 1.4283196232281625,
19
  "module_device_transfer_count": 270,
20
- "metrics": {
21
- "orbitquant": {
22
- "records": 1,
23
- "latest_metrics": {
24
- "generated_samples": 1,
25
- "wall_time_seconds": 34.435127603821456,
26
- "peak_vram_bytes": 13014530560
27
- }
28
- },
29
- "original": {
30
- "records": 1,
31
- "latest_metrics": {
32
- "generated_samples": 1,
33
- "wall_time_seconds": 5.203367706853896,
34
- "peak_vram_bytes": 23267081728
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": [
 
1
  {
2
+ "status": "not_run",
3
  "source_model_id": "Tongyi-MAI/Z-Image-Turbo",
4
  "source_revision": "f332072aa78be7aecdf3ee76d5c247082da564a6",
5
  "weight_bits": 2,
6
  "activation_bits": 3,
7
+ "codebook_version": 2,
8
  "target_policy": "z_image",
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": 3.5582019980065525,
17
+ "orbitquant_seconds": 2.1125443908385932,
18
+ "adaln_seconds": 0.015195954591035843,
19
+ "device_transfer_seconds": 1.4178410354070365,
20
  "module_device_transfer_count": 270,
21
+ "metrics": {},
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
22
  "native_smoke": {
23
  "proof_format": "orbitquant-native-smoke-v1",
24
+ "proof_source": "local_compare_native_bundle",
25
  "comparison_asset_path": "assets/image_generation_comparison_matrix.webp",
26
  "paired_prompt_seed_count": 1,
27
  "paired_prompt_seed_keys": [
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:aea099a0445d60c272a2c45a8c4e88fd4b3cedf7181a4b3600873104eb7a7941
3
  size 1599432256
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa8e04810221025dc262f66632975bdbf8466b341ad19db61ba4ad4f174f402f
3
  size 1599432256
model_index.json CHANGED
@@ -13,9 +13,10 @@
13
  "rotations": "orbitquant_rotations.safetensors",
14
  "weight_bits": 2,
15
  "activation_bits": 3,
 
16
  "target_policy": "z_image",
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": 2,
15
  "activation_bits": 3,
16
+ "codebook_version": 2,
17
  "target_policy": "z_image",
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 sha256:04da03ff11a91f322d72c88fe5686e7ebf66972de6ed5ab606e5282e682a332f
3
  size 816
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c06a807bddc6af0246586ded2276ec8dca0e309543a9701de1b6872fb8b6b31
3
  size 816
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",
@@ -2721,13 +2721,13 @@
2721
  ]
2722
  },
2723
  "checksums": {
2724
- "assets/image_generation_comparison_matrix.webp": "87f719a534ba76c19031c82ca761c2fea4e40615f0d1a23f37dd29556ee102aa",
2725
- "benchmark/summary.json": "2bda4afcd7266fbe4e0d184690605dd63f9546fc3a1ec0ca88e117387bf6bdbe",
2726
- "model.safetensors": "aea099a0445d60c272a2c45a8c4e88fd4b3cedf7181a4b3600873104eb7a7941",
2727
- "model_index.json": "54320a7d5779945db1e5e50b43c11dea40ec5dff74c9dd31a419bda6419160bd",
2728
- "orbitquant_codebooks.safetensors": "04da03ff11a91f322d72c88fe5686e7ebf66972de6ed5ab606e5282e682a332f",
2729
  "orbitquant_rotations.safetensors": "e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825",
2730
  "prompts.json": "65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af",
2731
- "quantization_config.json": "a0320d9f96882ee890452ea222e0d4c36db3d5ad57382fc8908a70251d91f614"
2732
  }
2733
  }
 
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",
 
2721
  ]
2722
  },
2723
  "checksums": {
2724
+ "assets/image_generation_comparison_matrix.webp": "f7e324ad238c63d0d25568acb0d4cee76bf9ddfd3c43f22845ad2138728daef7",
2725
+ "benchmark/summary.json": "59f01ffd627235229bc83868ac999ce280ccca656a138603baea85ea32929075",
2726
+ "model.safetensors": "aa8e04810221025dc262f66632975bdbf8466b341ad19db61ba4ad4f174f402f",
2727
+ "model_index.json": "61be70974b031f683d3af31c38322999f93c0788909aa7e33b91499962db225a",
2728
+ "orbitquant_codebooks.safetensors": "3c06a807bddc6af0246586ded2276ec8dca0e309543a9701de1b6872fb8b6b31",
2729
  "orbitquant_rotations.safetensors": "e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825",
2730
  "prompts.json": "65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af",
2731
+ "quantization_config.json": "d5fbd286ff541dc629e0a2b5a216580674faaf7a3814679b0fb7e4a2122cc67b"
2732
  }
2733
  }
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
  }