Upload FLUX.2 Klein OrbitQuant W4A4

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: `W4A4`
-- 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: `flux2`

SHA256SUMS

@@ -1,10 +1,10 @@
-88da4ec4377d590e016e7a0a0dcea0dc2c371bcdc629e620036aa9866d11b460  README.md
-7f1b987d3068bd488fcc9f38fd742521b5c2a879d569153f9cee503020230eaf  assets/image_generation_comparison_matrix.webp
-836adbe76bb4af829c8392dd5bc9aa16096209051d5dbb61b9201d4c8e2bdbf3  benchmark/summary.json
+43e85b8b735147b297753dde56df6d61cb42cdac8321b8c6137f898d2a2e93e2  README.md
+4983dedd382a6ca16289cac20ceb4a54d0eef84d5195673bddde99c8c6f32012  assets/image_generation_comparison_matrix.webp
+d39ce2afa1fb562244a9acb8712e2645aca4b5da60ecc3115d5e299d6c8ec1ad  benchmark/summary.json
 f03eaef37e04ff47af0935607a334bb80d26a3cf1bd9dfdc11a675b56bedfedc  model.safetensors
-3ae62e3d093928f912bac2fb77cb6c8e60125de2964e7a9054729b8e2b1dbf50  model_index.json
+f99a4281801bd8d56a9e0eea0ab2f6c450d3d4cf425b5db238e49f6be15b6228  model_index.json
 818e4de6a927335c47bb7d148f23a05dda61a3abb6e8f67cd8b9510960bdcf4c  orbitquant_codebooks.safetensors
-3214dae8bd2af5219716258eece9eb987c20d35a55f0915c0cfdf50ba281e831  orbitquant_manifest.json
+fd5327032b7e41dcd6abcd5fb6cffa467c4fca2d33e170e1cbb94f8e38098bba  orbitquant_manifest.json
 3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17  orbitquant_rotations.safetensors
 9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd  prompts.json
-5be7e81b42cf0dce679c49138c1560e77bb0e5f4942ae25a701af6f54db7676d  quantization_config.json
+d9207a741f4c353143ce5038f06349679c4c89e38e334b3f0773ed1e4c5d356a  quantization_config.json

benchmark/summary.json

@@ -5,166 +5,64 @@
   "weight_bits": 4,
   "activation_bits": 4,
   "target_policy": "flux2",
-  "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": 2.2789748371578753,
-  "orbitquant_seconds": 0.8833848647773266,
-  "adaln_seconds": 0.33560978434979916,
-  "device_transfer_seconds": 1.0559081598185003,
-  "module_device_transfer_count": 1,
+  "quantization_elapsed_seconds": 3.4752471749670804,
+  "orbitquant_seconds": 2.135687416419387,
+  "adaln_seconds": 0.4508278868161142,
+  "device_transfer_seconds": 0.882997278124094,
+  "module_device_transfer_count": 103,
   "metrics": {
-    "original": {
-      "records": 10,
+    "orbitquant": {
+      "records": 1,
       "latest_metrics": {
         "generated_samples": 1,
-        "wall_time_seconds": 1.1371133890934289,
-        "peak_vram_bytes": 18602007552
+        "wall_time_seconds": 8.579359667841345,
+        "peak_vram_bytes": 13604859904
       }
     },
-    "orbitquant": {
-      "records": 10,
+    "original": {
+      "records": 1,
       "latest_metrics": {
         "generated_samples": 1,
-        "wall_time_seconds": 1.4097541328519583,
-        "peak_vram_bytes": 20526148608
+        "wall_time_seconds": 1.9006238640286028,
+        "peak_vram_bytes": 18602007552
       }
     }
   },
-  "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": [
-      [
-        "flux2-native",
-        "0",
-        "color-binding"
-      ],
-      [
-        "flux2-native",
-        "0",
-        "counting"
-      ],
-      [
-        "flux2-native",
-        "0",
-        "cyrillic-text-rendering"
-      ],
-      [
-        "flux2-native",
-        "0",
-        "english-text-rendering"
-      ],
-      [
-        "flux2-native",
-        "0",
-        "long-prompt"
-      ],
-      [
-        "flux2-native",
-        "0",
-        "occlusion-reflection"
-      ],
       [
         "flux2-native",
         "0",
         "simple-object"
-      ],
-      [
-        "flux2-native",
-        "0",
-        "spatial-relationship"
-      ],
-      [
-        "flux2-native",
-        "0",
-        "style-heavy"
-      ],
-      [
-        "flux2-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": [
-          [
-            "flux2-native",
-            "0",
-            "color-binding"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "counting"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "cyrillic-text-rendering"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "english-text-rendering"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "long-prompt"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "occlusion-reflection"
-          ],
           [
             "flux2-native",
             "0",
             "simple-object"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "spatial-relationship"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "style-heavy"
-          ],
-          [
-            "flux2-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": [
-          [
-            "flux2-native",
-            "0",
-            "color-binding"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "counting"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "cyrillic-text-rendering"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "english-text-rendering"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "long-prompt"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "occlusion-reflection"
-          ],
           [
             "flux2-native",
             "0",
             "simple-object"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "spatial-relationship"
-          ],
-          [
-            "flux2-native",
-            "0",
-            "style-heavy"
-          ],
-          [
-            "flux2-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": 4,
   "activation_bits": 4,
   "target_policy": "flux2",
-  "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": "flux2",
   "adaln_policy": "int4_rtn_group64_bf16_activation",
   "quantized_modules": [
@@ -966,13 +966,13 @@
     ]
   },
   "checksums": {
-    "benchmark/summary.json": "836adbe76bb4af829c8392dd5bc9aa16096209051d5dbb61b9201d4c8e2bdbf3",
+    "assets/image_generation_comparison_matrix.webp": "4983dedd382a6ca16289cac20ceb4a54d0eef84d5195673bddde99c8c6f32012",
+    "benchmark/summary.json": "d39ce2afa1fb562244a9acb8712e2645aca4b5da60ecc3115d5e299d6c8ec1ad",
     "model.safetensors": "f03eaef37e04ff47af0935607a334bb80d26a3cf1bd9dfdc11a675b56bedfedc",
-    "model_index.json": "3ae62e3d093928f912bac2fb77cb6c8e60125de2964e7a9054729b8e2b1dbf50",
+    "model_index.json": "f99a4281801bd8d56a9e0eea0ab2f6c450d3d4cf425b5db238e49f6be15b6228",
     "orbitquant_codebooks.safetensors": "818e4de6a927335c47bb7d148f23a05dda61a3abb6e8f67cd8b9510960bdcf4c",
     "orbitquant_rotations.safetensors": "3a4014dbeee3988e275fcae07cc9312aceab5d14fb06c55f9a6548d1d7e2fa17",
     "prompts.json": "9d0a30e526f2840b3c68f61565e7345602d23175f2765dec60a6319a9bfcaedd",
-    "quantization_config.json": "5be7e81b42cf0dce679c49138c1560e77bb0e5f4942ae25a701af6f54db7676d",
-    "assets/image_generation_comparison_matrix.webp": "7f1b987d3068bd488fcc9f38fd742521b5c2a879d569153f9cee503020230eaf"
+    "quantization_config.json": "d9207a741f4c353143ce5038f06349679c4c89e38e334b3f0773ed1e4c5d356a"
   }
 }

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": "flux2",
   "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
 }