Replace artifact with paper-correct codebook v2 and native proof

README.md

@@ -90,7 +90,7 @@ The compact benchmark summary records native BF16-vs-OrbitQuant evidence for the
 - Method: `orbitquant`
 - Bits: `W3A3`
 - Runtime mode: `auto_fused`
-- Activation kernel backend: `triton_cuda`
+- Activation kernel backend: `auto`
 - Activation normalization epsilon: `1e-10`
 - Quantization device: `cuda`
 - Weight quantization backend: `triton_cuda`
@@ -103,7 +103,7 @@ 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: `238`
 - AdaLN INT4 modules: `32`
 - Skipped modules: `6`

SHA256SUMS

@@ -1,10 +1,10 @@
-e38366473940aab2237685e8b14a557b0a72a7f80e4c3f34f99b635299a0de71  README.md
-10bd69eb323fd3a1e59d94e19f0a13539ed7bceb66553689f4e2e7101740a050  assets/image_generation_comparison_matrix.webp
-09dff01693b1de8c8d43b6804135123917c82b2fd5d6b1447a1f210a4758d85b  benchmark/summary.json
-53a4a4444b217130187c72ccc1ea23011c0b20576d4b88ae96a0448865b9c9e1  model.safetensors
-7231b61b747c373ebdcc0dcb415c4a2dd1cae95c7b2c7087b860af37ef1f70fa  model_index.json
-0b74017deb25b79989c9fc90f5b190f456b7fa38c30a62df250221aa13be8767  orbitquant_codebooks.safetensors
-37dec7f755e4f4f2f4854b7f60f7132031686de840ecb101344f12c9d7d7b460  orbitquant_manifest.json
+b9d113adcc4d9732f91f0311d3b8ad9570025422691c80e44f634acd5828299b  README.md
+09e7a50d7db741bf65bd014e8e20e69dacb7d6c753f76c360ba6182b117a68ba  assets/image_generation_comparison_matrix.webp
+4d991cd3c6e93dac9b30b4962e62d1f8e248bad6a8f773d2813833e69332dc5d  benchmark/summary.json
+13e6cfb9a08f889f2b09acd8a467af246819742d72c653d2b66098b2fcca6da8  model.safetensors
+e4cecb0d1691cb9fac9f23b751d929c12f5ec21d47b23a0634511ab86cabf799  model_index.json
+1eee6cfd0c387f363ed56a297d5e0c91c3a8768cc1a02d7c167ae9083e78f354  orbitquant_codebooks.safetensors
+2a5427e7c81b2758c409101cd28dbad66081f18d1a897adc91c239127c9aea64  orbitquant_manifest.json
 e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825  orbitquant_rotations.safetensors
 65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af  prompts.json
-39429d49ecdb79af28a16af38c9a42c28d5d4ba2c9f59b113053624cd36db833  quantization_config.json
+4815c8ce48d538f46ca226a92b956c6fc45578b2fa04b8868759c789e7128fd5  quantization_config.json

benchmark/summary.json

@@ -1,42 +1,27 @@
 {
-  "status": "metrics_recorded",
+  "status": "not_run",
   "source_model_id": "Tongyi-MAI/Z-Image-Turbo",
   "source_revision": "f332072aa78be7aecdf3ee76d5c247082da564a6",
   "weight_bits": 3,
   "activation_bits": 3,
+  "codebook_version": 2,
   "target_policy": "z_image",
   "runtime_mode": "auto_fused",
-  "activation_kernel_backend": "triton_cuda",
+  "activation_kernel_backend": "auto",
   "activation_eps": 1e-10,
   "quantization_device": "cuda",
   "weight_quantization_backend": "triton_cuda",
   "quantization_staging_mode": "streaming",
   "synchronize_per_module": false,
-  "quantization_elapsed_seconds": 2.528441976290196,
-  "orbitquant_seconds": 1.0835723914206028,
-  "adaln_seconds": 0.01221933914348483,
-  "device_transfer_seconds": 1.423280114773661,
+  "quantization_elapsed_seconds": 5.273913557175547,
+  "orbitquant_seconds": 3.8487836881540716,
+  "adaln_seconds": 0.02463266858831048,
+  "device_transfer_seconds": 1.3874301291070879,
   "module_device_transfer_count": 270,
-  "metrics": {
-    "orbitquant": {
-      "records": 1,
-      "latest_metrics": {
-        "generated_samples": 1,
-        "wall_time_seconds": 32.681167694274336,
-        "peak_vram_bytes": 13758003712
-      }
-    },
-    "original": {
-      "records": 1,
-      "latest_metrics": {
-        "generated_samples": 1,
-        "wall_time_seconds": 5.207647128030658,
-        "peak_vram_bytes": 23267081728
-      }
-    }
-  },
+  "metrics": {},
   "native_smoke": {
     "proof_format": "orbitquant-native-smoke-v1",
+    "proof_source": "local_compare_native_bundle",
     "comparison_asset_path": "assets/image_generation_comparison_matrix.webp",
     "paired_prompt_seed_count": 1,
     "paired_prompt_seed_keys": [

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:53a4a4444b217130187c72ccc1ea23011c0b20576d4b88ae96a0448865b9c9e1
+oid sha256:13e6cfb9a08f889f2b09acd8a467af246819742d72c653d2b66098b2fcca6da8
 size 2351458072

model_index.json

@@ -13,9 +13,10 @@
   "rotations": "orbitquant_rotations.safetensors",
   "weight_bits": 3,
   "activation_bits": 3,
+  "codebook_version": 2,
   "target_policy": "z_image",
   "runtime_mode": "auto_fused",
-  "activation_kernel_backend": "triton_cuda",
+  "activation_kernel_backend": "auto",
   "activation_eps": 1e-10,
   "quantization_device": "cuda",
   "weight_quantization_backend": "triton_cuda",

orbitquant_codebooks.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0b74017deb25b79989c9fc90f5b190f456b7fa38c30a62df250221aa13be8767
+oid sha256:1eee6cfd0c387f363ed56a297d5e0c91c3a8768cc1a02d7c167ae9083e78f354
 size 448

orbitquant_manifest.json

@@ -12,10 +12,10 @@
   "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": "auto_fused",
-  "activation_kernel_backend": "triton_cuda",
+  "activation_kernel_backend": "auto",
   "activation_eps": 1e-10,
   "adaln_group_size": 64,
   "quantization_device": "cuda",
@@ -2721,13 +2721,13 @@
     ]
   },
   "checksums": {
-    "assets/image_generation_comparison_matrix.webp": "10bd69eb323fd3a1e59d94e19f0a13539ed7bceb66553689f4e2e7101740a050",
-    "benchmark/summary.json": "09dff01693b1de8c8d43b6804135123917c82b2fd5d6b1447a1f210a4758d85b",
-    "model.safetensors": "53a4a4444b217130187c72ccc1ea23011c0b20576d4b88ae96a0448865b9c9e1",
-    "model_index.json": "7231b61b747c373ebdcc0dcb415c4a2dd1cae95c7b2c7087b860af37ef1f70fa",
-    "orbitquant_codebooks.safetensors": "0b74017deb25b79989c9fc90f5b190f456b7fa38c30a62df250221aa13be8767",
+    "assets/image_generation_comparison_matrix.webp": "09e7a50d7db741bf65bd014e8e20e69dacb7d6c753f76c360ba6182b117a68ba",
+    "benchmark/summary.json": "4d991cd3c6e93dac9b30b4962e62d1f8e248bad6a8f773d2813833e69332dc5d",
+    "model.safetensors": "13e6cfb9a08f889f2b09acd8a467af246819742d72c653d2b66098b2fcca6da8",
+    "model_index.json": "e4cecb0d1691cb9fac9f23b751d929c12f5ec21d47b23a0634511ab86cabf799",
+    "orbitquant_codebooks.safetensors": "1eee6cfd0c387f363ed56a297d5e0c91c3a8768cc1a02d7c167ae9083e78f354",
     "orbitquant_rotations.safetensors": "e1724c9394bf83ea05da6b457b8072d2bcb3d5a533dafce916da8deb517d3825",
     "prompts.json": "65ec8afff39a6cf80feee2eac2f06fa66e51cd471a8382f4355a04fd74eda0af",
-    "quantization_config.json": "39429d49ecdb79af28a16af38c9a42c28d5d4ba2c9f59b113053624cd36db833"
+    "quantization_config.json": "4815c8ce48d538f46ca226a92b956c6fc45578b2fa04b8868759c789e7128fd5"
   }
 }

quantization_config.json

@@ -6,6 +6,7 @@
   "rotation_seed": 0,
   "block_size": "paper",
   "codebook": "lloyd_max",
+  "codebook_version": 2,
   "codebook_dtype": "float32",
   "row_norm_dtype": "bfloat16",
   "activation_norm_dtype": "float32",
@@ -18,9 +19,9 @@
   "modules_dtype_dict": {},
   "artifact_format_version": 1,
   "runtime_mode": "auto_fused",
-  "activation_kernel_backend": "triton_cuda",
-  "packed_matmul_block_m": 32,
+  "activation_kernel_backend": "auto",
+  "packed_matmul_block_m": 64,
   "packed_matmul_block_n": 64,
-  "packed_matmul_block_k": 64,
-  "packed_matmul_num_warps": 8
+  "packed_matmul_block_k": 128,
+  "packed_matmul_num_warps": 4
 }