flashinfer.quantization.resolve_nvfp4_4over6

flashinfer.quantization.resolve_nvfp4_4over6(setting: ~flashinfer.quantization.nvfp4_quantization_utils.NVFP44Over6Config | None = <nvfp4_4over6 unset: read FLASHINFER_NVFP4_4OVER6*>) → NVFP44Over6Config | None

Resolve a public nvfp4_4over6= argument to the recipe the kernels use.

This is the only place the FLASHINFER_NVFP4_4OVER6* environment variables are consulted on the Python side, so precedence is decided exactly once.

Parameters:

setting (NVFP44Over6Config or None) –

  • omitted (the default): read the environment. A FutureWarning is emitted when it enables 4over6, since the variables are a compatibility shim.

  • None: 4over6 off. The environment is ignored.

  • NVFP44Over6Config: on with exactly that recipe. The environment is ignored, with no per-field merge.

Returns:

The resolved recipe, where None means 4over6 is off. This is the value every kernel driver below the public boundary expects.

Return type:

NVFP44Over6Config or None