WaveCut commited on
Commit
9bf7a15
·
verified ·
1 Parent(s): f5838f8

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: `W3A3`
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: `W3A3`
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
- e38366473940aab2237685e8b14a557b0a72a7f80e4c3f34f99b635299a0de71 README.md
2
- 10bd69eb323fd3a1e59d94e19f0a13539ed7bceb66553689f4e2e7101740a050 assets/image_generation_comparison_matrix.webp
3
- 09dff01693b1de8c8d43b6804135123917c82b2fd5d6b1447a1f210a4758d85b benchmark/summary.json
4
- 53a4a4444b217130187c72ccc1ea23011c0b20576d4b88ae96a0448865b9c9e1 model.safetensors
5
- 7231b61b747c373ebdcc0dcb415c4a2dd1cae95c7b2c7087b860af37ef1f70fa model_index.json
6
- 0b74017deb25b79989c9fc90f5b190f456b7fa38c30a62df250221aa13be8767 orbitquant_codebooks.safetensors
7
- 37dec7f755e4f4f2f4854b7f60f7132031686de840ecb101344f12c9d7d7b460 orbitquant_manifest.json
8
  e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825 orbitquant_rotations.safetensors
9
  65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af prompts.json
10
- 39429d49ecdb79af28a16af38c9a42c28d5d4ba2c9f59b113053624cd36db833 quantization_config.json
 
1
+ b9d113adcc4d9732f91f0311d3b8ad9570025422691c80e44f634acd5828299b README.md
2
+ 09e7a50d7db741bf65bd014e8e20e69dacb7d6c753f76c360ba6182b117a68ba assets/image_generation_comparison_matrix.webp
3
+ 4d991cd3c6e93dac9b30b4962e62d1f8e248bad6a8f773d2813833e69332dc5d benchmark/summary.json
4
+ 13e6cfb9a08f889f2b09acd8a467af246819742d72c653d2b66098b2fcca6da8 model.safetensors
5
+ e4cecb0d1691cb9fac9f23b751d929c12f5ec21d47b23a0634511ab86cabf799 model_index.json
6
+ 1eee6cfd0c387f363ed56a297d5e0c91c3a8768cc1a02d7c167ae9083e78f354 orbitquant_codebooks.safetensors
7
+ 2a5427e7c81b2758c409101cd28dbad66081f18d1a897adc91c239127c9aea64 orbitquant_manifest.json
8
  e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825 orbitquant_rotations.safetensors
9
  65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af prompts.json
10
+ 4815c8ce48d538f46ca226a92b956c6fc45578b2fa04b8868759c789e7128fd5 quantization_config.json
assets/image_generation_comparison_matrix.webp CHANGED

Git LFS Details

  • SHA256: 10bd69eb323fd3a1e59d94e19f0a13539ed7bceb66553689f4e2e7101740a050
  • Pointer size: 130 Bytes
  • Size of remote file: 58 kB

Git LFS Details

  • SHA256: 09e7a50d7db741bf65bd014e8e20e69dacb7d6c753f76c360ba6182b117a68ba
  • Pointer size: 130 Bytes
  • Size of remote file: 60.3 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": 3,
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": 2.528441976290196,
16
- "orbitquant_seconds": 1.0835723914206028,
17
- "adaln_seconds": 0.01221933914348483,
18
- "device_transfer_seconds": 1.423280114773661,
19
  "module_device_transfer_count": 270,
20
- "metrics": {
21
- "orbitquant": {
22
- "records": 1,
23
- "latest_metrics": {
24
- "generated_samples": 1,
25
- "wall_time_seconds": 32.681167694274336,
26
- "peak_vram_bytes": 13758003712
27
- }
28
- },
29
- "original": {
30
- "records": 1,
31
- "latest_metrics": {
32
- "generated_samples": 1,
33
- "wall_time_seconds": 5.207647128030658,
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": 3,
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": 5.273913557175547,
17
+ "orbitquant_seconds": 3.8487836881540716,
18
+ "adaln_seconds": 0.02463266858831048,
19
+ "device_transfer_seconds": 1.3874301291070879,
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 sha256:53a4a4444b217130187c72ccc1ea23011c0b20576d4b88ae96a0448865b9c9e1
3
  size 2351458072
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:13e6cfb9a08f889f2b09acd8a467af246819742d72c653d2b66098b2fcca6da8
3
  size 2351458072
model_index.json CHANGED
@@ -13,9 +13,10 @@
13
  "rotations": "orbitquant_rotations.safetensors",
14
  "weight_bits": 3,
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": 3,
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 sha256:0b74017deb25b79989c9fc90f5b190f456b7fa38c30a62df250221aa13be8767
3
  size 448
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1eee6cfd0c387f363ed56a297d5e0c91c3a8768cc1a02d7c167ae9083e78f354
3
  size 448
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": "10bd69eb323fd3a1e59d94e19f0a13539ed7bceb66553689f4e2e7101740a050",
2725
- "benchmark/summary.json": "09dff01693b1de8c8d43b6804135123917c82b2fd5d6b1447a1f210a4758d85b",
2726
- "model.safetensors": "53a4a4444b217130187c72ccc1ea23011c0b20576d4b88ae96a0448865b9c9e1",
2727
- "model_index.json": "7231b61b747c373ebdcc0dcb415c4a2dd1cae95c7b2c7087b860af37ef1f70fa",
2728
- "orbitquant_codebooks.safetensors": "0b74017deb25b79989c9fc90f5b190f456b7fa38c30a62df250221aa13be8767",
2729
  "orbitquant_rotations.safetensors": "e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825",
2730
  "prompts.json": "65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af",
2731
- "quantization_config.json": "39429d49ecdb79af28a16af38c9a42c28d5d4ba2c9f59b113053624cd36db833"
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": "09e7a50d7db741bf65bd014e8e20e69dacb7d6c753f76c360ba6182b117a68ba",
2725
+ "benchmark/summary.json": "4d991cd3c6e93dac9b30b4962e62d1f8e248bad6a8f773d2813833e69332dc5d",
2726
+ "model.safetensors": "13e6cfb9a08f889f2b09acd8a467af246819742d72c653d2b66098b2fcca6da8",
2727
+ "model_index.json": "e4cecb0d1691cb9fac9f23b751d929c12f5ec21d47b23a0634511ab86cabf799",
2728
+ "orbitquant_codebooks.safetensors": "1eee6cfd0c387f363ed56a297d5e0c91c3a8768cc1a02d7c167ae9083e78f354",
2729
  "orbitquant_rotations.safetensors": "e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825",
2730
  "prompts.json": "65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af",
2731
+ "quantization_config.json": "4815c8ce48d538f46ca226a92b956c6fc45578b2fa04b8868759c789e7128fd5"
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
  }