Upload Z-Image Turbo OrbitQuant W2A4

README.md

@@ -77,21 +77,21 @@ The compact benchmark summary records native BF16-vs-OrbitQuant evidence for the
 | Evidence | Value |
 | --- | --- |
 | Comparison matrix | `assets/image_generation_comparison_matrix.webp` |
-| Paired prompt/seed count | `10` |
-| BF16 source generated samples | `10` |
+| Paired prompt/seed count | `1` |
+| BF16 source generated samples | `1` |
 | BF16 source generated frames | `0` |
-| BF16 source nonempty outputs | `10` |
-| OrbitQuant generated samples | `10` |
+| BF16 source nonempty outputs | `1` |
+| OrbitQuant generated samples | `1` |
 | OrbitQuant generated frames | `0` |
-| OrbitQuant nonempty outputs | `10` |
+| OrbitQuant nonempty outputs | `1` |
 
 ## Quantization
 
 - Method: `orbitquant`
 - Bits: `W2A4`
-- Runtime mode: `dequant_bf16`
+- Runtime mode: `auto_fused`
 - Activation kernel backend: `triton_cuda`
-- Activation normalization epsilon: `1e-12`
+- Activation normalization epsilon: `1e-10`
 - Quantization device: `cuda`
 - Weight quantization backend: `triton_cuda`
 - Target policy: `z_image`

SHA256SUMS

@@ -1,10 +1,10 @@
-b26ab179ce89c72f06c11b4f25ce4631395b54393db27aee969f607d2e40752a  README.md
-3d17f4a76f5625cc76d914b5fac8325e79909c82e6420317040b789d8ce0bcaf  assets/image_generation_comparison_matrix.webp
-a76152c262ad1463f103956d61deff27192410baeacfdeafe3616dda9c1e0691  benchmark/summary.json
+6dd0e7c59f36dfc530e627a9aa5f629c78403876d468e4b47841339c7cb30417  README.md
+4826107b893c41cf7517727cd2e96bf2eee4a59b1afd597624385b4d307ff83e  assets/image_generation_comparison_matrix.webp
+29926a9051050e69e5ad1dee895eb188c19b091987d6a7e470b93f7695eeeca2  benchmark/summary.json
 aea099a0445d60c272a2c45a8c4e88fd4b3cedf7181a4b3600873104eb7a7941  model.safetensors
-2bf9814dfd6b541494e150f80ef4ce1d1fc3f1563f42e9b2563879bf8034f50f  model_index.json
+e238cc9bbe81a339e7c5ba5ce4ab7f0d83a40b592714e5bc53629425ece81748  model_index.json
 d72050fd3948b189d54d1dc55020b2f2f3f505d842a178a2121d05bebdcd55d5  orbitquant_codebooks.safetensors
-711686c57ecb2343b2d792e8d6dfb4e73f62ba2ec4e27fc56633d9201b176d2c  orbitquant_manifest.json
+e3993931d417dcb9e9842490371f297944962ed423dbf6ad93da7496088743a8  orbitquant_manifest.json
 e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825  orbitquant_rotations.safetensors
 65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af  prompts.json
-59bf91760447c06d6341976af01802485b6b396ec4015a153c1e59b9b187ced7  quantization_config.json
+aba7920e16eaffbe5208d8ce2b3619cbfd55430487a83c9677a5da08352c750b  quantization_config.json

benchmark/summary.json

@@ -5,166 +5,64 @@
   "weight_bits": 2,
   "activation_bits": 4,
   "target_policy": "z_image",
-  "runtime_mode": "dequant_bf16",
+  "runtime_mode": "auto_fused",
   "activation_kernel_backend": "triton_cuda",
+  "activation_eps": 1e-10,
   "quantization_device": "cuda",
   "weight_quantization_backend": "triton_cuda",
-  "quantization_staging_mode": "component",
+  "quantization_staging_mode": "streaming",
   "synchronize_per_module": false,
-  "quantization_elapsed_seconds": 2.9623005548492074,
-  "orbitquant_seconds": 1.0405256771482527,
-  "adaln_seconds": 0.01457158662378788,
-  "device_transfer_seconds": 1.898455814924091,
-  "module_device_transfer_count": 1,
+  "quantization_elapsed_seconds": 2.39361240202561,
+  "orbitquant_seconds": 0.9436650313436985,
+  "adaln_seconds": 0.011932591442018747,
+  "device_transfer_seconds": 1.4291342729702592,
+  "module_device_transfer_count": 270,
   "metrics": {
-    "original": {
-      "records": 10,
+    "orbitquant": {
+      "records": 1,
       "latest_metrics": {
         "generated_samples": 1,
-        "wall_time_seconds": 2.7209880319423974,
-        "peak_vram_bytes": 23267112448
+        "wall_time_seconds": 35.89397249696776,
+        "peak_vram_bytes": 13017512448
       }
     },
-    "orbitquant": {
-      "records": 10,
+    "original": {
+      "records": 1,
       "latest_metrics": {
         "generated_samples": 1,
-        "wall_time_seconds": 3.7248233389109373,
-        "peak_vram_bytes": 24860445184
+        "wall_time_seconds": 5.200292671099305,
+        "peak_vram_bytes": 23267081728
       }
     }
   },
-  "published_summary": "compact",
-  "raw_generation_records": "local-only",
-  "activation_eps": 1e-12,
   "native_smoke": {
     "proof_format": "orbitquant-native-smoke-v1",
-    "proof_source": "local_paired_native_smoke_backup",
     "comparison_asset_path": "assets/image_generation_comparison_matrix.webp",
-    "paired_prompt_seed_count": 10,
+    "paired_prompt_seed_count": 1,
     "paired_prompt_seed_keys": [
-      [
-        "z-image-native",
-        "0",
-        "color-binding"
-      ],
-      [
-        "z-image-native",
-        "0",
-        "counting"
-      ],
-      [
-        "z-image-native",
-        "0",
-        "cyrillic-text-rendering"
-      ],
-      [
-        "z-image-native",
-        "0",
-        "english-text-rendering"
-      ],
-      [
-        "z-image-native",
-        "0",
-        "long-prompt"
-      ],
-      [
-        "z-image-native",
-        "0",
-        "occlusion-reflection"
-      ],
       [
         "z-image-native",
         "0",
         "simple-object"
-      ],
-      [
-        "z-image-native",
-        "0",
-        "spatial-relationship"
-      ],
-      [
-        "z-image-native",
-        "0",
-        "style-heavy"
-      ],
-      [
-        "z-image-native",
-        "0",
-        "two-object-composition"
       ]
     ],
     "splits": {
       "original": {
-        "records": 10,
-        "generated_samples": 10,
+        "records": 1,
+        "generated_samples": 1,
         "generated_frames": 0,
-        "nonempty_output_count": 10,
+        "nonempty_output_count": 1,
         "seeds": [
           "0"
         ],
         "prompt_ids": [
-          "color-binding",
-          "counting",
-          "cyrillic-text-rendering",
-          "english-text-rendering",
-          "long-prompt",
-          "occlusion-reflection",
-          "simple-object",
-          "spatial-relationship",
-          "style-heavy",
-          "two-object-composition"
+          "simple-object"
         ],
         "pair_keys": [
-          [
-            "z-image-native",
-            "0",
-            "color-binding"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "counting"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "cyrillic-text-rendering"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "english-text-rendering"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "long-prompt"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "occlusion-reflection"
-          ],
           [
             "z-image-native",
             "0",
             "simple-object"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "spatial-relationship"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "style-heavy"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "two-object-composition"
           ]
         ],
         "native_settings": [
@@ -179,75 +77,21 @@
         ]
       },
       "orbitquant": {
-        "records": 10,
-        "generated_samples": 10,
+        "records": 1,
+        "generated_samples": 1,
         "generated_frames": 0,
-        "nonempty_output_count": 10,
+        "nonempty_output_count": 1,
         "seeds": [
           "0"
         ],
         "prompt_ids": [
-          "color-binding",
-          "counting",
-          "cyrillic-text-rendering",
-          "english-text-rendering",
-          "long-prompt",
-          "occlusion-reflection",
-          "simple-object",
-          "spatial-relationship",
-          "style-heavy",
-          "two-object-composition"
+          "simple-object"
         ],
         "pair_keys": [
-          [
-            "z-image-native",
-            "0",
-            "color-binding"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "counting"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "cyrillic-text-rendering"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "english-text-rendering"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "long-prompt"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "occlusion-reflection"
-          ],
           [
             "z-image-native",
             "0",
             "simple-object"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "spatial-relationship"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "style-heavy"
-          ],
-          [
-            "z-image-native",
-            "0",
-            "two-object-composition"
           ]
         ],
         "native_settings": [
@@ -262,5 +106,7 @@
         ]
       }
     }
-  }
+  },
+  "published_summary": "compact",
+  "raw_generation_records": "local-only"
 }

model_index.json

@@ -14,10 +14,10 @@
   "weight_bits": 2,
   "activation_bits": 4,
   "target_policy": "z_image",
-  "runtime_mode": "dequant_bf16",
+  "runtime_mode": "auto_fused",
   "activation_kernel_backend": "triton_cuda",
+  "activation_eps": 1e-10,
   "quantization_device": "cuda",
   "weight_quantization_backend": "triton_cuda",
-  "quantization_staging_mode": "component",
-  "activation_eps": 1e-12
+  "quantization_staging_mode": "streaming"
 }

orbitquant_manifest.json

@@ -14,13 +14,13 @@
   "codebook": "lloyd_max",
   "codebook_version": 1,
   "row_norm_dtype": "bfloat16",
-  "runtime_mode": "dequant_bf16",
+  "runtime_mode": "auto_fused",
   "activation_kernel_backend": "triton_cuda",
-  "activation_eps": 1e-12,
+  "activation_eps": 1e-10,
   "adaln_group_size": 64,
   "quantization_device": "cuda",
   "weight_quantization_backend": "triton_cuda",
-  "quantization_staging_mode": "component",
+  "quantization_staging_mode": "streaming",
   "target_policy": "z_image",
   "adaln_policy": "int4_rtn_group64_bf16_activation",
   "quantized_modules": [
@@ -2721,13 +2721,13 @@
     ]
   },
   "checksums": {
-    "benchmark/summary.json": "a76152c262ad1463f103956d61deff27192410baeacfdeafe3616dda9c1e0691",
+    "assets/image_generation_comparison_matrix.webp": "4826107b893c41cf7517727cd2e96bf2eee4a59b1afd597624385b4d307ff83e",
+    "benchmark/summary.json": "29926a9051050e69e5ad1dee895eb188c19b091987d6a7e470b93f7695eeeca2",
     "model.safetensors": "aea099a0445d60c272a2c45a8c4e88fd4b3cedf7181a4b3600873104eb7a7941",
-    "model_index.json": "2bf9814dfd6b541494e150f80ef4ce1d1fc3f1563f42e9b2563879bf8034f50f",
+    "model_index.json": "e238cc9bbe81a339e7c5ba5ce4ab7f0d83a40b592714e5bc53629425ece81748",
     "orbitquant_codebooks.safetensors": "d72050fd3948b189d54d1dc55020b2f2f3f505d842a178a2121d05bebdcd55d5",
     "orbitquant_rotations.safetensors": "e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825",
     "prompts.json": "65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af",
-    "quantization_config.json": "59bf91760447c06d6341976af01802485b6b396ec4015a153c1e59b9b187ced7",
-    "assets/image_generation_comparison_matrix.webp": "3d17f4a76f5625cc76d914b5fac8325e79909c82e6420317040b789d8ce0bcaf"
+    "quantization_config.json": "aba7920e16eaffbe5208d8ce2b3619cbfd55430487a83c9677a5da08352c750b"
   }
 }

quantization_config.json

@@ -9,7 +9,7 @@
   "codebook_dtype": "float32",
   "row_norm_dtype": "bfloat16",
   "activation_norm_dtype": "float32",
-  "activation_eps": 1e-12,
+  "activation_eps": 1e-10,
   "weight_pack_dtype": "uint8",
   "target_policy": "z_image",
   "adaln_policy": "int4_rtn",
@@ -17,6 +17,10 @@
   "modules_to_not_convert": [],
   "modules_dtype_dict": {},
   "artifact_format_version": 1,
-  "runtime_mode": "dequant_bf16",
-  "activation_kernel_backend": "triton_cuda"
+  "runtime_mode": "auto_fused",
+  "activation_kernel_backend": "triton_cuda",
+  "packed_matmul_block_m": 32,
+  "packed_matmul_block_n": 64,
+  "packed_matmul_block_k": 64,
+  "packed_matmul_num_warps": 8
 }