Module quant_format 

2026-09-25: The quantized-weight layout of a checkpoint, and its ignore list: CompressedTensorsFormat, ModeloptFormat or Fp8BlockScaledFormat, chosen by detect_quant_format.

Owner: model-layers (weight loading). Invariants: none beyond the types.

Structs

CompressedTensorsFormat

2026-09-25: A compressed-tensors checkpoint.

Fp8BlockScaledFormat

2026-09-25: An FP8 block-scaled checkpoint.

IgnoreList

2026-09-30: A checkpoint’s ignore list, parsed with the matching rules of its format (metrale_config::precision_plan::Target, the matcher the declared precision plan uses): compressed-tensors re: regexes or exact names, ModelOpt exact names or globs, HF fp8 module paths on whole dotted segments.

ModeloptFormat

2026-09-25: A ModelOpt checkpoint, and the fallback of super::detect_quant_format when no quantized weights are found.

Traits

QuantFormat

2026-09-25: A quantized-weight layout: its log name, the Nvfp4Variant it maps to, and the module globs its ignore list keeps unquantized. The weight loaders pick their variant through weight_map::detect_nvfp4_variant, not through this trait.

Functions

detect_quant_format

2026-09-25: Pick the QuantFormat for a checkpoint.