Enum WeightFormat 

pub enum WeightFormat {
    Nvfp4,
    Fp8BlockScaled,
    Bf16Dense,
}

Checkpoint weight format, detected from safetensors metadata.

Determines how raw weight bytes are interpreted and transformed into the runtime NVFP4 format used by Atlas GEMM kernels.

Variants

Nvfp4

NVFP4 E2M1 on disk (nvidia ModelOpt or compressed-tensors). Weights load directly into QuantizedWeight with no conversion.

Fp8BlockScaled

FP8 E4M3 block-scaled on disk (e.g. quant_method: "fp8" with weight_block_size). Each weight tensor has a weight_scale_inv (BF16 per-block) companion. At load time: FP8 -> BF16 -> NVFP4 (runtime quantization).

Bf16Dense

BF16 dense on disk (unquantized, e.g. attention Q/K/V in Standard NVFP4 models). At load time: BF16 -> NVFP4 (runtime quantization).

Implementations

impl WeightFormat

pub fn detect(store: &WeightStore, config: &ModelConfig) -> Self

Detect the weight format from a WeightStore by probing key names.

pub fn is_fp8(&self) -> bool

Whether this format requires FP8 -> BF16 dequantization at load time.

Trait Implementations

impl Clone for WeightFormat

fn clone(&self) -> WeightFormat

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for WeightFormat

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

Formats the value using the given formatter.

impl PartialEq for WeightFormat

fn eq(&self, other: &WeightFormat) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for WeightFormat

impl Eq for WeightFormat

impl StructuralPartialEq for WeightFormat