Upload OrbitQuant artifact

README.md

@@ -35,9 +35,9 @@ artifact_dir = snapshot_download(artifact_id, repo_type="model")
 pipe = load_quantized_pipeline_from_artifact(
     artifact_dir,
     torch_dtype=torch.bfloat16,
-    device="cuda",
     runtime_mode="auto_fused",
 )
+pipe.enable_model_cpu_offload(device="cuda")
 
 image = pipe(
     prompt="A precise product photo of a red ceramic mug on a wooden desk",
@@ -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: `W3A3`
-- 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: `flux`

SHA256SUMS

@@ -1,10 +1,10 @@
-ee3cd9dc900c67b07558cde34b9a7903acab68514f2242cdba58109339bf7a0d  README.md
-f09f310429c85d4bd7c9cd08fe4dfdb66907bfc0a4a0a5ada574b0dc9fadd64c  assets/image_generation_comparison_matrix.webp
-6229ebbb5c33047cb97ce1a5ff7ba9f8d8601ca4cf38d7aae3359131887328b0  benchmark/summary.json
+3c9e8a61d8cd46d6dfee94c078012c7c3659d2d51bc2dffb4c2ea0261bba5b6b  README.md
+37bffbabc93e9fce907623210ef45e7a8f7b931c40b0806f50292f6fab4c4aaf  assets/image_generation_comparison_matrix.webp
+8e05e5b3f4e747d84608b5dac99380fbe9415c134144f0cb127e51554162b416  benchmark/summary.json
 4db8791c3442e15657ca6d1f04090051be511b89f0c5c83fa60f2d55d5d6e703  model.safetensors
-d49dae0b24df0661af5c737fb49742fde8c213fd62abf746825fd92b13c041cf  model_index.json
+012732bc630cf7a58a92a440732061c3f3086d1ace8f7f26a6830e656eb66e0c  model_index.json
 ad19d28b6ee1a57cdc6079142e4e6f43e3b2bb5db79dcf563466646023aa40f0  orbitquant_codebooks.safetensors
-ad215d961fb47a37b7fefb43365d56bc5d35e698b64945bc7d0b64b08464a804  orbitquant_manifest.json
+a7145e3ac3c0e6cd74156e599638066d4cee398c41987a0219a14a4c442cf504  orbitquant_manifest.json
 ac6045b8b9ea4f802bc0aa394caea3300e527af02bb7743517b8dcc7cbf5d31a  orbitquant_rotations.safetensors
 2beb04fc153a129a8e3a2c95645720b80e6eeb4cad99280e9ac5c5b9e7836a07  prompts.json
-ba2aafa412d48711a72d07c5778689052560bb29e50be2f19a8e91029eba55c2  quantization_config.json
+c8bfb282b023c5775be8cc5826d2f9eef426831248c7feec97b1a516a90dcd32  quantization_config.json

benchmark/summary.json

@@ -5,166 +5,64 @@
   "weight_bits": 3,
   "activation_bits": 3,
   "target_policy": "flux",
-  "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": 6.238366776611656,
-  "orbitquant_seconds": 2.8115671570412815,
-  "adaln_seconds": 0.5923770787194371,
-  "device_transfer_seconds": 2.816385400015861,
-  "module_device_transfer_count": 1,
+  "quantization_elapsed_seconds": 5.941415391396731,
+  "orbitquant_seconds": 0.8518197890371084,
+  "adaln_seconds": 0.3548668515868485,
+  "device_transfer_seconds": 4.711370585486293,
+  "module_device_transfer_count": 494,
   "metrics": {
-    "original": {
-      "records": 10,
+    "orbitquant": {
+      "records": 1,
       "latest_metrics": {
         "generated_samples": 1,
-        "wall_time_seconds": 1.6293397820554674,
-        "peak_vram_bytes": 36301890048
+        "wall_time_seconds": 34.16444475017488,
+        "peak_vram_bytes": 12919665152
       }
     },
-    "orbitquant": {
-      "records": 10,
+    "original": {
+      "records": 1,
       "latest_metrics": {
         "generated_samples": 1,
-        "wall_time_seconds": 2.4774683727882802,
-        "peak_vram_bytes": 41268513280
+        "wall_time_seconds": 24.054778529331088,
+        "peak_vram_bytes": 24287500800
       }
     }
   },
-  "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": [
-      [
-        "flux1-schnell-native",
-        "0",
-        "color-binding"
-      ],
-      [
-        "flux1-schnell-native",
-        "0",
-        "counting"
-      ],
-      [
-        "flux1-schnell-native",
-        "0",
-        "cyrillic-text-rendering"
-      ],
-      [
-        "flux1-schnell-native",
-        "0",
-        "english-text-rendering"
-      ],
-      [
-        "flux1-schnell-native",
-        "0",
-        "long-prompt"
-      ],
-      [
-        "flux1-schnell-native",
-        "0",
-        "occlusion-reflection"
-      ],
       [
         "flux1-schnell-native",
         "0",
         "simple-object"
-      ],
-      [
-        "flux1-schnell-native",
-        "0",
-        "spatial-relationship"
-      ],
-      [
-        "flux1-schnell-native",
-        "0",
-        "style-heavy"
-      ],
-      [
-        "flux1-schnell-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": [
-          [
-            "flux1-schnell-native",
-            "0",
-            "color-binding"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "counting"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "cyrillic-text-rendering"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "english-text-rendering"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "long-prompt"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "occlusion-reflection"
-          ],
           [
             "flux1-schnell-native",
             "0",
             "simple-object"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "spatial-relationship"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "style-heavy"
-          ],
-          [
-            "flux1-schnell-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": [
-          [
-            "flux1-schnell-native",
-            "0",
-            "color-binding"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "counting"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "cyrillic-text-rendering"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "english-text-rendering"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "long-prompt"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "occlusion-reflection"
-          ],
           [
             "flux1-schnell-native",
             "0",
             "simple-object"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "spatial-relationship"
-          ],
-          [
-            "flux1-schnell-native",
-            "0",
-            "style-heavy"
-          ],
-          [
-            "flux1-schnell-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": 3,
   "activation_bits": 3,
   "target_policy": "flux",
-  "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": "flux",
   "adaln_policy": "int4_rtn_group64_bf16_activation",
   "quantized_modules": [
@@ -5568,13 +5568,13 @@
     ]
   },
   "checksums": {
-    "benchmark/summary.json": "6229ebbb5c33047cb97ce1a5ff7ba9f8d8601ca4cf38d7aae3359131887328b0",
+    "assets/image_generation_comparison_matrix.webp": "37bffbabc93e9fce907623210ef45e7a8f7b931c40b0806f50292f6fab4c4aaf",
+    "benchmark/summary.json": "8e05e5b3f4e747d84608b5dac99380fbe9415c134144f0cb127e51554162b416",
     "model.safetensors": "4db8791c3442e15657ca6d1f04090051be511b89f0c5c83fa60f2d55d5d6e703",
-    "model_index.json": "d49dae0b24df0661af5c737fb49742fde8c213fd62abf746825fd92b13c041cf",
+    "model_index.json": "012732bc630cf7a58a92a440732061c3f3086d1ace8f7f26a6830e656eb66e0c",
     "orbitquant_codebooks.safetensors": "ad19d28b6ee1a57cdc6079142e4e6f43e3b2bb5db79dcf563466646023aa40f0",
     "orbitquant_rotations.safetensors": "ac6045b8b9ea4f802bc0aa394caea3300e527af02bb7743517b8dcc7cbf5d31a",
     "prompts.json": "2beb04fc153a129a8e3a2c95645720b80e6eeb4cad99280e9ac5c5b9e7836a07",
-    "quantization_config.json": "ba2aafa412d48711a72d07c5778689052560bb29e50be2f19a8e91029eba55c2",
-    "assets/image_generation_comparison_matrix.webp": "f09f310429c85d4bd7c9cd08fe4dfdb66907bfc0a4a0a5ada574b0dc9fadd64c"
+    "quantization_config.json": "c8bfb282b023c5775be8cc5826d2f9eef426831248c7feec97b1a516a90dcd32"
   }
 }

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": "flux",
   "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
 }