runtime_config.json

{
  "runtime": {
    "recommended_torch_cuda_alloc_conf": null,
    "avoid_torch_cuda_alloc_conf": [
      "expandable_segments:True,max_split_size_mb:32"
    ],
    "keep_model_resident": true,
    "avoid_empty_cache_between_generations": true,
    "use_pe_for_image_benchmarks": false
  },
  "sdnq": {
    "requires_explicit_apply_quantized_matmul": true,
    "apply_quantized_matmul_components": [
      "pe",
      "text_encoder",
      "transformer"
    ],
    "apply_quantized_matmul_function": "sdnq.loader.apply_sdnq_options_to_model(component, use_quantized_matmul=True)"
  },
  "validated": {
    "device": "NVIDIA RTX 6000 Ada Generation",
    "torch": "2.8.0+cu128",
    "sdnq": "0.1.9",
    "num_inference_steps": 8,
    "guidance_scale": 1.0,
    "use_pe": false
  },
  "metrics": {
    "explicit_quantized_matmul_default_allocator": "metrics/ernie_uint4_qmm_explicit_default_allocator_8step_metrics.json",
    "allocator_debug": "metrics/runtime_allocator_debug_metrics.json"
  }
}