Calibration Data Collector

Collect per-layer Hessians (and optionally activation amax) for external quantization. No weights are modified.

By ThunderFun·Created 2 months ago·Updated about a month ago· 0

model
conditioning

calibration_path

◄num_steps4►

◄num_samples16►

◄seed0►

◄hessian_block_size128►

◄hessian_formatdlr_nystrom►

◄dlr_rank128►

◄collect_amaxtrue►

◄output_path/tmp/ComfyUI/output/calibration.pt►

◄latent_height64►

◄latent_width64►

◄convrotfalse►

◄rot_size256►

◄rpbhfalse►

◄rpbh_seed42►

◄rpbh_block_size256►

◄permuquantfalse►

◄pisofalse►

◄sigma_min0.00►

◄sigma_max1.00►

◄force_cpu_hookfalse►

◄timestep_channel_amaxfalse►

◄timestep_buckets4►

Categorymodel/quantization

Inputs (24)

Name	Type	Default	Description
model	MODEL		Loaded diffusion model (FP16/BF16/FP32).
conditioning	CONDITIONING		Pre-encoded conditioning from CLIPTextEncode or similar.
num_steps	INT	41–50	Denoising steps per sample.
num_samples	INT	161–4096	Independent samples to accumulate over. 16-128 recommended.
seed	INT	00–4294967295	Seed for noise and timestep sampling.
hessian_block_size	INT	1280–1024	0 = full H (paper-accurate, auto memory-mapped to disk). 128 = diagonal blocks (default, saves RAM). Ignored when hessian_format='dlr'.
hessian_format	COMBO	dlr_nystrom	Hessian storage format. 'block' = diagonal blocks (use hessian_block_size). 'full' = full Hessian (memory-mapped). 'dlr' = Diagonal + Low-Rank via FrequentDirections (streaming, exact diagonal, repeated SVD). 'dlr_nystrom' = DLR via randomized Nyström (streaming, exact diagonal, ~100× faster construction — recommended default).
dlr_rank	INT	1281–4096	Rank for DLR Hessian (only used when hessian_format is 'dlr' or 'dlr_nystrom'). Same memory budget as block_size=rank. Recommended: 64-256.
collect_amax	BOOLEAN	true	Also collect max(abs(x)) per layer. Required for activation quantization; not used by weight-only GPTQ.
output_path	STRING	/tmp/ComfyUI/output/calibration.pt	Where to save the calibration .pt file.
latent_heightopt	INT	648–1024	Latent spatial height. 128 for 1024px, 64 for 512px.
latent_widthopt	INT	648–1024	Latent spatial width. 128 for 1024px, 64 for 512px.
convrotopt	BOOLEAN	false	Enable ConvRot Hadamard rotation. Collects Hessians in rotated space for better block-diagonal approximation.
rot_sizeopt	INT	25616–4096	Hadamard group size (must be power of 2). 256 recommended for ConvRot.
rpbhopt	BOOLEAN	false	Enable RPBH (Randomized Permuted Block-Hadamard) rotation. Mutually exclusive with ConvRot. Spreads outliers via random permutation + block Hadamard + sign flip. Uses per-layer auto-detected block size (largest power-of-2 dividing in_features). Best paired with DLR Hessian format — block format is lossy because RPBH's permutation destroys block-diagonal structure. Not recommended with PermuQuant.
rpbh_seedopt	INT	420–2147483647	Random seed for RPBH permutation and signs. The converter must use the same seed to regenerate matching rotation parameters for weight folding.
rpbh_block_sizeopt	INT	2560–65536	RPBH Hadamard block size. 0 = auto-detect (largest power-of-2 dividing in_features, default). Set to 256 for uniform block size across all layers — smaller blocks give finer-grained outlier spreading. Must be a power of 2 if non-zero. The converter must use the same block size.
permuquantopt	BOOLEAN	false	Enable PermuQuant channel reordering. Runs a second calibration pass with channels sorted by second moment for better quantization.
pisoopt	BOOLEAN	false	Collect Hessian diagonal for PiSO data-aware scale optimization. Adds a small overhead to store diag(X^T X) per layer, which the converter uses to compute optimal per-row scales instead of absmax.
sigma_minopt	FLOAT	0.000–1	Lower bound of the sigma range to sample. Set to 0.875 with Wan 2.2 high-noise expert, or 0.0 for full range (default).
sigma_maxopt	FLOAT	1.000–1	Upper bound of the sigma range to sample. Set to 0.875 with Wan 2.2 low-noise expert, or 1.0 for full range (default).
force_cpu_hookopt	BOOLEAN	false	Force hook-side processing to CPU. Enable only if you hit GPU OOM during calibration — the GPU-fast path uses ~20-50 MB of transient VRAM per layer.
timestep_channel_amaxopt	BOOLEAN	false	Collect per-timestep-bucket per-channel amax in addition to overall per-channel amax. Enables the converter to analyse time-varying channel imbalance for static-dynamic decomposition (ViDiT-Q Section 4.2). Automatically enables per-channel amax collection. Adds 'channel_amax_by_timestep' to the output.
timestep_bucketsopt	INT	41–64	Number of timestep buckets for per-timestep channel amax. 4 = divide denoising into 4 equal ranges (ViDiT-Q paper default).

Outputs (1)

Name	Type	Description
calibration_path	STRING	—