Replace artifact with paper-correct codebook v2 and native proof

.gitattributes

@@ -52,3 +52,4 @@ assets/wan-native_seed0_original_subject-consistency.mp4 filter=lfs diff=lfs mer
 assets/wan-native_seed0_original_subject-consistency_contact_sheet.webp filter=lfs diff=lfs merge=lfs -text
 assets/wan-native_seed0_original_text-rendering.mp4 filter=lfs diff=lfs merge=lfs -text
 assets/wan-native_seed0_original_text-rendering_contact_sheet.webp filter=lfs diff=lfs merge=lfs -text
+assets/video_generation_comparison_matrix.webp filter=lfs diff=lfs merge=lfs -text

README.md

@@ -81,21 +81,21 @@ The compact benchmark summary records native BF16-vs-OrbitQuant evidence for the
 | Evidence | Value |
 | --- | --- |
 | Comparison matrix | `assets/video_generation_comparison_matrix.webp` |
-| Paired prompt/seed count | `5` |
-| BF16 source generated samples | `5` |
-| BF16 source generated frames | `405` |
-| BF16 source nonempty outputs | `5` |
-| OrbitQuant generated samples | `5` |
-| OrbitQuant generated frames | `405` |
-| OrbitQuant nonempty outputs | `5` |
+| Paired prompt/seed count | `1` |
+| BF16 source generated samples | `1` |
+| BF16 source generated frames | `81` |
+| BF16 source nonempty outputs | `1` |
+| OrbitQuant generated samples | `1` |
+| OrbitQuant generated frames | `81` |
+| OrbitQuant nonempty outputs | `1` |
 
 ## Quantization
 
 - Method: `orbitquant`
 - Bits: `W4A4`
-- Runtime mode: `dequant_bf16`
-- Activation kernel backend: `triton_cuda`
-- Activation normalization epsilon: `1e-12`
+- Runtime mode: `auto_fused`
+- Activation kernel backend: `auto`
+- Activation normalization epsilon: `1e-10`
 - Quantization device: `cuda`
 - Weight quantization backend: `triton_cuda`
 - Target policy: `wan`
@@ -107,7 +107,7 @@ The compact benchmark summary records native BF16-vs-OrbitQuant evidence for the
 - Block size: `paper`
 - Block size policy: `largest_power_of_two_dividing_dim`
 - Codebook: `lloyd_max`
-- Codebook version: `1`
+- Codebook version: `2`
 - Quantized transformer modules: `300`
 - AdaLN INT4 modules: `0`
 - Skipped modules: `6`

SHA256SUMS

@@ -1,10 +1,10 @@
-d3986399318d7adf58135c4f9e9ba3fa7f92b26ff2aed25a754b77159717f5b5  README.md
-c7d0a00ee273efd217df7a3d4c38aae28065af27a6d9179d8b8cbfae762f2551  assets/video_generation_comparison_matrix.webp
-ccf87334e02eee391e4ced63f056c1cc5efeb2ec4d32a7050ab089463e8bc1cd  benchmark/summary.json
-2fe65bda9f0f318516c36ac28d22f76ee645409bc220692926a4311547ca9c4a  model.safetensors
-766fa2af9e7f6413bad62c7352cc1d982d91fff01d14bea291991df33229885b  model_index.json
-cf10bbd01f3ad7f81fc2d18467c299627611d61ea251e4a166db7b86d9c38358  orbitquant_codebooks.safetensors
-9927193fdf827014d1957a31ee548d5bf0dec99f3aaa753b14bde3d097a35b0c  orbitquant_manifest.json
+b2ef7cab1c38d8ebb729fd757cf27a7e4ec404925d4fae00ebfbd6d94d02d7ed  README.md
+de523743fa4c42852524cd7635d6ace1d11149305aa8615ffd788a4bf59aeb46  assets/video_generation_comparison_matrix.webp
+14966a818ef2b61424bf02154a0414bdad5c4583d444dcbafaf87b056cde6dfe  benchmark/summary.json
+5778da463b35c5c68761c7307e05634f3c818e2a05d481eeb50799cceb376889  model.safetensors
+923430fc4a686c5d0a5f1a82d5b3b40eeeececfc3773eb26a31cf4668b3f70a3  model_index.json
+2b534248778a59626b9c21e32ccf4760d4061a06b6aa7fcfdacb5a37a660605e  orbitquant_codebooks.safetensors
+4d619eba696d82dbe9725c016cd40307324250e59e3c885637bbcedb474f3171  orbitquant_manifest.json
 b7b2d1e28dd66554c2539d9152b63938a86f1fae0d65664a0df786313dee5dd9  orbitquant_rotations.safetensors
 b01eb3abccf0db3cb1699c09033db918ceb7886cd9b6b6d4f03304ed330ddbc8  prompts.json
-450100d23073d338526cc049c3149240a3b2fb4f1fc326c98862d78f6743a2bf  quantization_config.json
+f8a5a7dc5e808adade3f064fa556d535d4bdc5c13a778552929347939f2fc97c  quantization_config.json

benchmark/summary.json

@@ -1,185 +1,99 @@
 {
-  "status": "metrics_recorded",
+  "status": "not_run",
   "source_model_id": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
   "source_revision": "0fad780a534b6463e45facd96134c9f345acfa5b",
   "weight_bits": 4,
   "activation_bits": 4,
+  "codebook_version": 2,
   "target_policy": "wan",
-  "runtime_mode": "dequant_bf16",
-  "activation_kernel_backend": "triton_cuda",
+  "runtime_mode": "auto_fused",
+  "activation_kernel_backend": "auto",
+  "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": 1.3253011349588633,
-  "orbitquant_seconds": 0.44857147336006165,
+  "quantization_elapsed_seconds": 2.125622207298875,
+  "orbitquant_seconds": 1.8142682267352939,
   "adaln_seconds": 0.0,
-  "device_transfer_seconds": 0.8674816680140793,
-  "module_device_transfer_count": 1,
-  "metrics": {
-    "original": {
-      "records": 5,
-      "latest_metrics": {
-        "generated_samples": 1,
-        "wall_time_seconds": 122.07954734982923,
-        "generated_frames": 81,
-        "peak_vram_bytes": 26752849408
-      }
-    },
-    "orbitquant": {
-      "records": 5,
-      "latest_metrics": {
-        "generated_samples": 1,
-        "wall_time_seconds": 189.13376781577244,
-        "generated_frames": 81,
-        "peak_vram_bytes": 27537771008
-      }
-    }
-  },
-  "published_summary": "compact",
-  "raw_generation_records": "local-only",
-  "activation_eps": 1e-12,
+  "device_transfer_seconds": 0.2972148791886866,
+  "module_device_transfer_count": 300,
+  "metrics": {},
   "native_smoke": {
     "proof_format": "orbitquant-native-smoke-v1",
-    "proof_source": "local_paired_native_smoke_backup",
+    "proof_source": "local_compare_native_bundle",
     "comparison_asset_path": "assets/video_generation_comparison_matrix.webp",
-    "paired_prompt_seed_count": 5,
+    "paired_prompt_seed_count": 1,
     "paired_prompt_seed_keys": [
-      [
-        "wan-native",
-        "0",
-        "camera-movement"
-      ],
-      [
-        "wan-native",
-        "0",
-        "scene-consistency"
-      ],
       [
         "wan-native",
         "0",
         "simple-motion"
-      ],
-      [
-        "wan-native",
-        "0",
-        "subject-consistency"
-      ],
-      [
-        "wan-native",
-        "0",
-        "text-rendering"
       ]
     ],
     "splits": {
       "original": {
-        "records": 5,
-        "generated_samples": 5,
-        "generated_frames": 405,
-        "nonempty_output_count": 5,
+        "records": 1,
+        "generated_samples": 1,
+        "generated_frames": 81,
+        "nonempty_output_count": 1,
         "seeds": [
           "0"
         ],
         "prompt_ids": [
-          "camera-movement",
-          "scene-consistency",
-          "simple-motion",
-          "subject-consistency",
-          "text-rendering"
+          "simple-motion"
         ],
         "pair_keys": [
-          [
-            "wan-native",
-            "0",
-            "camera-movement"
-          ],
-          [
-            "wan-native",
-            "0",
-            "scene-consistency"
-          ],
           [
             "wan-native",
             "0",
             "simple-motion"
-          ],
-          [
-            "wan-native",
-            "0",
-            "subject-consistency"
-          ],
-          [
-            "wan-native",
-            "0",
-            "text-rendering"
           ]
         ],
         "native_settings": [
           {
-            "suite": "wan-native",
-            "height": 480,
-            "width": 832,
+            "export_fps": 16,
             "frames": 81,
-            "steps": 50,
             "guidance": 5.0,
-            "export_fps": 16
+            "height": 480,
+            "steps": 50,
+            "suite": "wan-native",
+            "width": 832
           }
         ]
       },
       "orbitquant": {
-        "records": 5,
-        "generated_samples": 5,
-        "generated_frames": 405,
-        "nonempty_output_count": 5,
+        "records": 1,
+        "generated_samples": 1,
+        "generated_frames": 81,
+        "nonempty_output_count": 1,
         "seeds": [
           "0"
         ],
         "prompt_ids": [
-          "camera-movement",
-          "scene-consistency",
-          "simple-motion",
-          "subject-consistency",
-          "text-rendering"
+          "simple-motion"
         ],
         "pair_keys": [
-          [
-            "wan-native",
-            "0",
-            "camera-movement"
-          ],
-          [
-            "wan-native",
-            "0",
-            "scene-consistency"
-          ],
           [
             "wan-native",
             "0",
             "simple-motion"
-          ],
-          [
-            "wan-native",
-            "0",
-            "subject-consistency"
-          ],
-          [
-            "wan-native",
-            "0",
-            "text-rendering"
           ]
         ],
         "native_settings": [
           {
-            "suite": "wan-native",
-            "height": 480,
-            "width": 832,
+            "export_fps": 16,
             "frames": 81,
-            "steps": 50,
             "guidance": 5.0,
-            "export_fps": 16
+            "height": 480,
+            "steps": 50,
+            "suite": "wan-native",
+            "width": 832
           }
         ]
       }
     }
-  }
+  },
+  "published_summary": "compact",
+  "raw_generation_records": "local-only"
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2fe65bda9f0f318516c36ac28d22f76ee645409bc220692926a4311547ca9c4a
+oid sha256:5778da463b35c5c68761c7307e05634f3c818e2a05d481eeb50799cceb376889
 size 757753432

model_index.json

@@ -13,11 +13,12 @@
   "rotations": "orbitquant_rotations.safetensors",
   "weight_bits": 4,
   "activation_bits": 4,
+  "codebook_version": 2,
   "target_policy": "wan",
-  "runtime_mode": "dequant_bf16",
-  "activation_kernel_backend": "triton_cuda",
+  "runtime_mode": "auto_fused",
+  "activation_kernel_backend": "auto",
+  "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_codebooks.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf10bbd01f3ad7f81fc2d18467c299627611d61ea251e4a166db7b86d9c38358
+oid sha256:2b534248778a59626b9c21e32ccf4760d4061a06b6aa7fcfdacb5a37a660605e
 size 576

orbitquant_manifest.json

@@ -12,15 +12,15 @@
   "block_size": "paper",
   "block_size_policy": "largest_power_of_two_dividing_dim",
   "codebook": "lloyd_max",
-  "codebook_version": 1,
+  "codebook_version": 2,
   "row_norm_dtype": "bfloat16",
-  "runtime_mode": "dequant_bf16",
-  "activation_kernel_backend": "triton_cuda",
-  "activation_eps": 1e-12,
+  "runtime_mode": "auto_fused",
+  "activation_kernel_backend": "auto",
+  "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": "wan",
   "adaln_policy": "int4_rtn_group64_bf16_activation",
   "quantized_modules": [
@@ -3784,13 +3784,13 @@
     ]
   },
   "checksums": {
-    "benchmark/summary.json": "ccf87334e02eee391e4ced63f056c1cc5efeb2ec4d32a7050ab089463e8bc1cd",
-    "model.safetensors": "2fe65bda9f0f318516c36ac28d22f76ee645409bc220692926a4311547ca9c4a",
-    "model_index.json": "766fa2af9e7f6413bad62c7352cc1d982d91fff01d14bea291991df33229885b",
-    "orbitquant_codebooks.safetensors": "cf10bbd01f3ad7f81fc2d18467c299627611d61ea251e4a166db7b86d9c38358",
+    "assets/video_generation_comparison_matrix.webp": "de523743fa4c42852524cd7635d6ace1d11149305aa8615ffd788a4bf59aeb46",
+    "benchmark/summary.json": "14966a818ef2b61424bf02154a0414bdad5c4583d444dcbafaf87b056cde6dfe",
+    "model.safetensors": "5778da463b35c5c68761c7307e05634f3c818e2a05d481eeb50799cceb376889",
+    "model_index.json": "923430fc4a686c5d0a5f1a82d5b3b40eeeececfc3773eb26a31cf4668b3f70a3",
+    "orbitquant_codebooks.safetensors": "2b534248778a59626b9c21e32ccf4760d4061a06b6aa7fcfdacb5a37a660605e",
     "orbitquant_rotations.safetensors": "b7b2d1e28dd66554c2539d9152b63938a86f1fae0d65664a0df786313dee5dd9",
     "prompts.json": "b01eb3abccf0db3cb1699c09033db918ceb7886cd9b6b6d4f03304ed330ddbc8",
-    "quantization_config.json": "450100d23073d338526cc049c3149240a3b2fb4f1fc326c98862d78f6743a2bf",
-    "assets/video_generation_comparison_matrix.webp": "c7d0a00ee273efd217df7a3d4c38aae28065af27a6d9179d8b8cbfae762f2551"
+    "quantization_config.json": "f8a5a7dc5e808adade3f064fa556d535d4bdc5c13a778552929347939f2fc97c"
   }
 }

quantization_config.json

@@ -6,10 +6,11 @@
   "rotation_seed": 0,
   "block_size": "paper",
   "codebook": "lloyd_max",
+  "codebook_version": 2,
   "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": "wan",
   "adaln_policy": "int4_rtn",
@@ -17,6 +18,10 @@
   "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": "auto",
+  "packed_matmul_block_m": 64,
+  "packed_matmul_block_n": 64,
+  "packed_matmul_block_k": 128,
+  "packed_matmul_num_warps": 4
 }