__default__.yaml

seed: 12345
enable_cache: true
cache:
  root: runs
output:
  root: runs
  dirname: default
pipeline:
  dtype: torch.float16
  device: cuda
  shift_activations: false
eval:
  num_samples: 1024
  num_frames: null
  fps: null
  height: null
  width: null
  guidance_scale: null
  num_steps: null
  gt_metrics: ["clip_iqa", "clip_score", "image_reward", "fid"]
  ref_metrics: ["psnr", "lpips", "ssim", "fid"]
  gen_root: "{output}/{job}"
  ref_root: baselines/{dtype}/{model}/{protocol}
  gt_stats_root: benchmarks/stats
  num_gpus: 8
  batch_size_per_gpu: 1
  chunk_start: 0
  chunk_step: 1
  benchmarks:
    - "MJHQ"
    - "DCI"
quant:
  calib:
    data: qdiff
    path: datasets/{dtype}/{model}/{protocol}/{data}/s128
    num_samples: 128
    num_workers: 8
  wgts:
    dtype: null
    zero_point: null
    group_shapes:
    - - 1
      - -1
    scale_dtypes:
    - null
    skips: []
    enable_calib_range: true
    calib_range:
      degree: 2
      objective: OutputsError
      strategy: Manual
      granularity: Layer
      element_batch_size: 64
      sample_batch_size: 64
      element_size: 512
      sample_size: -1
      ratio: 1.0
      max_shrink: 0.2
      max_expand: 1.0
      num_grids: 80
      skips: []
    low_rank:
      rank: 32
      exclusive: false
      compensate: false
      early_stop: false
      degree: 2
      objective: OutputsError
      sample_batch_size: 64
      sample_size: -1
      num_iters: 1
      skips: []
  ipts:
    static: false
    dtype: null
    zero_point: null
    group_shapes:
    - - 1
      - -1
    scale_dtypes:
    - null
    allow_unsigned: false
    skips: []
    enable_calib_range: false
    calib_range:
      degree: 2
      objective: OutputsError
      strategy: Manual
      granularity: Layer
      element_batch_size: 64
      sample_batch_size: 64
      element_size: 512
      sample_size: -1
      ratio: 1.0
      max_shrink: 0.2
      max_expand: 1.0
      num_grids: 80
      skips: []
  enable_smooth: false
  smooth:
    enable_proj: false
    proj:
      degree: 2
      objective: OutputsError
      strategy: Manual
      granularity: Layer
      element_batch_size: -1
      sample_batch_size: 64
      element_size: -1
      sample_size: -1
      pre_reshape: true
      outputs_device: cpu
      spans:
      - - AbsMax
        - AbsMax
      alpha: 0.5
      beta: -1
      num_grids: 20
      skips: []
  develop_dtype: torch.float32