Module fp8_dequant 

2026-09-25: FP8 E4M3 → BF16 dequantization for weights with one per-tensor scale: a GPU path for an FP32 scale, and a host path for an FP32 or BF16 scale.

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

Functions

dequant_fp8_to_bf16_into

2026-09-25: Dequantize FP8 E4M3 + per-tensor scale → BF16 into dest, a caller-provided buffer of at least 2 × elements bytes, and return it as a DenseWeight. The GPU path does not synchronise, so the caller must order any reuse of dest after the launch (e.g. after quantize_to_nvfp4, which synchronises its stream).