{
  "protocol": "sequential components, lossless valid-token compaction, strict BF16 Flash SDPA, no callbacks in timed path, image save excluded",
  "source_model": "krea/Krea-2-Turbo",
  "source_revision": "98e0fe118d17c9e3547fbb2e25acdbae2cadf7c7",
  "original": {
    "variant": "original",
    "hardware": "NVIDIA A40",
    "source_revision": "98e0fe118d17c9e3547fbb2e25acdbae2cadf7c7",
    "quantized_revision": null,
    "candidate_root": null,
    "settings": {
      "width": 2048,
      "height": 2048,
      "num_inference_steps": 8,
      "guidance_scale": 0.0,
      "seed_base": 61000,
      "dtype": "bfloat16",
      "attention_backend": "strict Flash SDPA",
      "prompt_compaction": "lossless valid-token compaction",
      "vae_tiling": [
        1024,
        896
      ],
      "component_policy": "Qwen encode then release; DiT and tiled VAE render"
    },
    "load_seconds": 16.22555137425661,
    "load_gpu_peak_mb": 25103,
    "load_torch_peak_mb": 24706.955078125,
    "qwen_inventory": {
      "class_name": "Qwen3VLModel",
      "orbitquant_linear_count": 0,
      "source_precision_linear_count": 356,
      "executed_orbitquant_linear_count": 0,
      "effective_runtime_modes": [],
      "activation_kernel_backends": [],
      "last_activation_cache_hits": 0,
      "dequantized_cache_bytes": 0,
      "int8_weight_cache_bytes": 0
    },
    "dit_inventory": {
      "class_name": "Krea2Transformer2DModel",
      "orbitquant_linear_count": 0,
      "source_precision_linear_count": 264,
      "executed_orbitquant_linear_count": 0,
      "effective_runtime_modes": [],
      "activation_kernel_backends": [],
      "last_activation_cache_hits": 0,
      "dequantized_cache_bytes": 0,
      "int8_weight_cache_bytes": 0
    },
    "generated_samples": 10,
    "first_generation_seconds": 54.746916238218546,
    "hot_generation_mean_seconds": 54.64761028728552,
    "hot_generation_median_seconds": 54.57881709188223,
    "generation_mean_seconds": 54.65754088237882,
    "generation_median_seconds": 54.648766504600644,
    "gpu_peak_mb": 31441,
    "torch_peak_mb": 28982.806640625
  },
  "orbitquant": {
    "variant": "orbitquant",
    "hardware": "NVIDIA A40",
    "source_revision": "98e0fe118d17c9e3547fbb2e25acdbae2cadf7c7",
    "quantized_revision": "49e89c37f51cb394b4ee199196d9df8bcd042475",
    "candidate_root": "/workspace/krea2-final-transformer-imgin",
    "settings": {
      "width": 2048,
      "height": 2048,
      "num_inference_steps": 8,
      "guidance_scale": 0.0,
      "seed_base": 61000,
      "dtype": "bfloat16",
      "attention_backend": "strict Flash SDPA",
      "prompt_compaction": "lossless valid-token compaction",
      "vae_tiling": [
        1024,
        896
      ],
      "component_policy": "Qwen encode then release; DiT and tiled VAE render"
    },
    "load_seconds": 5.343890082091093,
    "load_gpu_peak_mb": 12635,
    "load_torch_peak_mb": 12431.72021484375,
    "qwen_inventory": {
      "class_name": "Qwen3VLModel",
      "orbitquant_linear_count": 356,
      "source_precision_linear_count": 0,
      "executed_orbitquant_linear_count": 252,
      "effective_runtime_modes": [
        "native_packed_matmul"
      ],
      "activation_kernel_backends": [
        "native_cuda_int8_surrogate",
        "triton_cuda_int8_surrogate"
      ],
      "last_activation_cache_hits": 108,
      "dequantized_cache_bytes": 0,
      "int8_weight_cache_bytes": 0
    },
    "dit_inventory": {
      "class_name": "Krea2Transformer2DModel",
      "orbitquant_linear_count": 232,
      "source_precision_linear_count": 32,
      "executed_orbitquant_linear_count": 232,
      "effective_runtime_modes": [
        "native_packed_matmul"
      ],
      "activation_kernel_backends": [
        "triton_cuda",
        "triton_cuda_int8_surrogate",
        "triton_cuda_packed_w4"
      ],
      "last_activation_cache_hits": 128,
      "dequantized_cache_bytes": 0,
      "int8_weight_cache_bytes": 0
    },
    "generated_samples": 10,
    "first_generation_seconds": 51.562784906476736,
    "hot_generation_mean_seconds": 50.08951893697182,
    "hot_generation_median_seconds": 49.98552576825023,
    "generation_mean_seconds": 50.23684553392231,
    "generation_median_seconds": 49.98872868530452,
    "gpu_peak_mb": 19177,
    "torch_peak_mb": 16798.73583984375
  },
  "orbitquant_empty_triton_cache": {
    "variant": "orbitquant",
    "hardware": "NVIDIA A40",
    "source_revision": "98e0fe118d17c9e3547fbb2e25acdbae2cadf7c7",
    "quantized_revision": "49e89c37f51cb394b4ee199196d9df8bcd042475",
    "candidate_root": "/workspace/krea2-final-transformer-imgin",
    "settings": {
      "width": 2048,
      "height": 2048,
      "num_inference_steps": 8,
      "guidance_scale": 0.0,
      "seed_base": 61000,
      "dtype": "bfloat16",
      "attention_backend": "strict Flash SDPA",
      "prompt_compaction": "lossless valid-token compaction",
      "vae_tiling": [
        1024,
        896
      ],
      "component_policy": "Qwen encode then release; DiT and tiled VAE render"
    },
    "load_seconds": 8.288543954491615,
    "load_gpu_peak_mb": 12827,
    "load_torch_peak_mb": 12432.95947265625,
    "qwen_inventory": {
      "class_name": "Qwen3VLModel",
      "orbitquant_linear_count": 356,
      "source_precision_linear_count": 0,
      "executed_orbitquant_linear_count": 252,
      "effective_runtime_modes": [
        "native_packed_matmul"
      ],
      "activation_kernel_backends": [
        "native_cuda_int8_surrogate",
        "triton_cuda_int8_surrogate"
      ],
      "last_activation_cache_hits": 108,
      "dequantized_cache_bytes": 0,
      "int8_weight_cache_bytes": 0
    },
    "dit_inventory": {
      "class_name": "Krea2Transformer2DModel",
      "orbitquant_linear_count": 232,
      "source_precision_linear_count": 32,
      "executed_orbitquant_linear_count": 232,
      "effective_runtime_modes": [
        "native_packed_matmul"
      ],
      "activation_kernel_backends": [
        "triton_cuda",
        "triton_cuda_int8_surrogate",
        "triton_cuda_packed_w4"
      ],
      "last_activation_cache_hits": 128,
      "dequantized_cache_bytes": 0,
      "int8_weight_cache_bytes": 0
    },
    "generated_samples": 1,
    "first_generation_seconds": 68.87463613599539,
    "hot_generation_mean_seconds": 68.87463613599539,
    "hot_generation_median_seconds": 68.87463613599539,
    "generation_mean_seconds": 68.87463613599539,
    "generation_median_seconds": 68.87463613599539,
    "gpu_peak_mb": 18657,
    "torch_peak_mb": 16792.41357421875
  },
  "comparison": {
    "hot_end_to_end_speedup": 1.0918924279185949,
    "hot_end_to_end_elapsed_change_percent": -8.415886544223373,
    "nvml_peak_change_percent": -39.00639292643364,
    "torch_allocated_peak_change_percent": -42.03896107046763
  },
  "matrix": {
    "matrix_path": "assets/original_vs_orbitquant_w4a4.webp",
    "matrix_size": [
      8192,
      10720
    ],
    "tile_size": [
      2048,
      2048
    ],
    "prompt_count": 10,
    "prompt_pairs_per_row": 2,
    "label_height": 96,
    "resized": false,
    "preview_path": "assets/original_vs_orbitquant_w4a4_preview.webp",
    "preview_size": [
      2048,
      2680
    ]
  }
}