Upload Z-Image Turbo OrbitQuant W2A3

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: `W2A3`
-- 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 @@
-242565e13a27f7a39a3788401b16f9de69dacd12f1516ff3ead02fc15963a13f  README.md
-2795c8e4510a596b235bbaf070dc9595d6194e70a61a67d72702fe16eae305ef  assets/image_generation_comparison_matrix.webp
-8d0639f28dc3181e2fabfc8036de0e1d329b98fc6666fb807e03382c67d69ec3  benchmark/summary.json
+81af00da21ac4abe16e78ec01f42e2ba17178206fc2b081b89a3a3cf09a24d4f  README.md
+87f719a534ba76c19031c82ca761c2fea4e40615f0d1a23f37dd29556ee102aa  assets/image_generation_comparison_matrix.webp
+2bda4afcd7266fbe4e0d184690605dd63f9546fc3a1ec0ca88e117387bf6bdbe  benchmark/summary.json
 aea099a0445d60c272a2c45a8c4e88fd4b3cedf7181a4b3600873104eb7a7941  model.safetensors
-4d75861b1b5ea3b306e95ccc93f6657fdbc02feab693fdf2a39e7aba83f6f0b7  model_index.json
+54320a7d5779945db1e5e50b43c11dea40ec5dff74c9dd31a419bda6419160bd  model_index.json
 04da03ff11a91f322d72c88fe5686e7ebf66972de6ed5ab606e5282e682a332f  orbitquant_codebooks.safetensors
-f1ffe2b6eee1dcc054cf3752a90d00851f15d98af857b15fde4384267dec46b4  orbitquant_manifest.json
+cdddf9d2aa7c409cd59d32d14c6d037b4336706bb9f422b45e1a069bda9127fc  orbitquant_manifest.json
 e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825  orbitquant_rotations.safetensors
 65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af  prompts.json
-d0572afbae9999d041bcc4ad6445dd4b1abd344df713e9848e40a706aac0c4fc  quantization_config.json
+a0320d9f96882ee890452ea222e0d4c36db3d5ad57382fc8908a70251d91f614  quantization_config.json

benchmark/summary.json

@@ -5,166 +5,64 @@
   "weight_bits": 2,
   "activation_bits": 3,
   "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": 1.9858340960927308,
-  "orbitquant_seconds": 0.5452375756576657,
-  "adaln_seconds": 0.014681137166917324,
-  "device_transfer_seconds": 1.4173690504394472,
-  "module_device_transfer_count": 1,
+  "quantization_elapsed_seconds": 2.160409632138908,
+  "orbitquant_seconds": 0.7088581034913659,
+  "adaln_seconds": 0.011828072369098663,
+  "device_transfer_seconds": 1.4283196232281625,
+  "module_device_transfer_count": 270,
   "metrics": {
-    "original": {
-      "records": 10,
+    "orbitquant": {
+      "records": 1,
       "latest_metrics": {
         "generated_samples": 1,
-        "wall_time_seconds": 2.7144667631946504,
-        "peak_vram_bytes": 23267112448
+        "wall_time_seconds": 34.435127603821456,
+        "peak_vram_bytes": 13014530560
       }
     },
-    "orbitquant": {
-      "records": 10,
+    "original": {
+      "records": 1,
       "latest_metrics": {
         "generated_samples": 1,
-        "wall_time_seconds": 3.6933902432210743,
-        "peak_vram_bytes": 24864999936
+        "wall_time_seconds": 5.203367706853896,
+        "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": 3,
   "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": "8d0639f28dc3181e2fabfc8036de0e1d329b98fc6666fb807e03382c67d69ec3",
+    "assets/image_generation_comparison_matrix.webp": "87f719a534ba76c19031c82ca761c2fea4e40615f0d1a23f37dd29556ee102aa",
+    "benchmark/summary.json": "2bda4afcd7266fbe4e0d184690605dd63f9546fc3a1ec0ca88e117387bf6bdbe",
     "model.safetensors": "aea099a0445d60c272a2c45a8c4e88fd4b3cedf7181a4b3600873104eb7a7941",
-    "model_index.json": "4d75861b1b5ea3b306e95ccc93f6657fdbc02feab693fdf2a39e7aba83f6f0b7",
+    "model_index.json": "54320a7d5779945db1e5e50b43c11dea40ec5dff74c9dd31a419bda6419160bd",
     "orbitquant_codebooks.safetensors": "04da03ff11a91f322d72c88fe5686e7ebf66972de6ed5ab606e5282e682a332f",
     "orbitquant_rotations.safetensors": "e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825",
     "prompts.json": "65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af",
-    "quantization_config.json": "d0572afbae9999d041bcc4ad6445dd4b1abd344df713e9848e40a706aac0c4fc",
-    "assets/image_generation_comparison_matrix.webp": "2795c8e4510a596b235bbaf070dc9595d6194e70a61a67d72702fe16eae305ef"
+    "quantization_config.json": "a0320d9f96882ee890452ea222e0d4c36db3d5ad57382fc8908a70251d91f614"
   }
 }

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
 }