Struct Fp8BlockScaledFormat 

pub struct Fp8BlockScaledFormat {
    pub ignore: IgnoreList,
}

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

Fields

ignore: IgnoreList

2026-09-30: The config’s ignore / exclude_modules (QuantizationConfig::ignore_modules; an HF modules_to_not_convert list does not reach it). The FP8 checkpoints that carry one are ModelOpt exports (nvidia/DeepSeek-V4-Flash-NVFP4: *.attn.*, mtp.*), so it matches as ModelOpt does: exact names or globs.

Implementations

impl Fp8BlockScaledFormat

pub fn new(ignore_modules: &[String]) -> Result<Self>

2026-09-30: A malformed ignore entry is refused.

Trait Implementations

impl Debug for Fp8BlockScaledFormat

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl QuantFormat for Fp8BlockScaledFormat

fn name(&self) -> &'static str

2026-09-25: Name for logs.

fn base_variant(&self) -> Nvfp4Variant

2026-09-25: The Nvfp4Variant this layout maps to.

fn is_ignored(&self, module_path: &str) -> bool

2026-09-30: Whether the ignore list names the module module_path (a module path, not a tensor name: no .weight), under the format’s own matching (IgnoreList).

fn variant_for(&self, module_path: &str) -> Nvfp4Variant

2026-09-25: Bf16Raw for an ignored module, else Self::base_variant.