vllm.model_executor.layers.fused_moe.oracle.mxfp8 ¶
_select_kernel_cls ¶
_select_kernel_cls(
backend: Fp8MoeBackend, config: FusedMoEConfig
) -> type[FusedMoEExperts]
Select the first supported expert class for the MXFP8 config.
Source code in vllm/model_executor/layers/fused_moe/oracle/mxfp8.py
select_mxfp8_moe_backend ¶
select_mxfp8_moe_backend(
config: FusedMoEConfig,
) -> tuple[Fp8MoeBackend, type[FusedMoEExperts]]
Select the MXFP8 MoE backend and the best expert class.
Returns:
| Type | Description |
|---|---|
tuple[Fp8MoeBackend, type[FusedMoEExperts]] | A tuple of (fp8_backend, experts_cls). |