benchmark/orbitquant.summary.json

{
  "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
}