Skip to content

vllm.model_executor.layers.fused_moe.oracle.mxfp8

Classes:

Functions:

Mxfp8MoEKernelOracle

Bases: MoEKernelOracle[Fp8MoeBackend]

MXFP8 MoE backend selection for one deployment configuration.

Source code in vllm/model_executor/layers/fused_moe/oracle/mxfp8.py 
79
80
81
82
83
84
85
86
87
class Mxfp8MoEKernelOracle(MoEKernelOracle[Fp8MoeBackend]):
    """MXFP8 MoE backend selection for one deployment configuration."""

    name = "Mxfp8"

    def backend_to_kernel_cls(
        self, backend: Fp8MoeBackend
    ) -> list[type[mk.FusedMoEExperts]]:
        return _mxfp8_backend_to_kernel_cls(backend)

_mxfp8_backend_to_kernel_cls(backend)

Resolve the MXFP8 expert classes for a backend.

DeepGEMM resolves directly to DeepGemmExperts (not the TritonOrDeepGemmExperts wrapper, whose Triton fallback cannot handle the MXFP8 1x32 scheme); all other backends defer to the FP8 resolver.

Source code in vllm/model_executor/layers/fused_moe/oracle/mxfp8.py 
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
def _mxfp8_backend_to_kernel_cls(
    backend: Fp8MoeBackend,
) -> list[type[mk.FusedMoEExperts]]:
    """Resolve the MXFP8 expert classes for a backend.

    DeepGEMM resolves directly to ``DeepGemmExperts`` (not the
    ``TritonOrDeepGemmExperts`` wrapper, whose Triton fallback cannot handle the
    MXFP8 1x32 scheme); all other backends defer to the FP8 resolver.
    """
    if backend == Fp8MoeBackend.DEEPGEMM:
        from vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe import (
            DeepGemmExperts,
        )

        return [DeepGemmExperts]
    if backend == Fp8MoeBackend.AITER_MXFP8:
        from vllm.model_executor.layers.fused_moe.experts.aiter_mxfp8_moe import (
            AiterMxfp8Experts,
        )

        return [AiterMxfp8Experts]
    if backend == Fp8MoeBackend.TRITON_MXFP8:
        from vllm.model_executor.layers.fused_moe.experts.mxfp8_native_moe import (
            Mxfp8NativeTritonExperts,
        )

        return [Mxfp8NativeTritonExperts]
    if backend == Fp8MoeBackend.EMULATION:
        from vllm.model_executor.layers.fused_moe.experts.mxfp8_emulation_moe import (
            Mxfp8EmulationTritonExperts,
        )

        return [Mxfp8EmulationTritonExperts]
    return backend_to_kernel_cls(backend)

_select_kernel_cls(backend, config)

Select the first supported expert class for the MXFP8 config.

Source code in vllm/model_executor/layers/fused_moe/oracle/mxfp8.py 
90
91
92
93
94
95
96
97
98
def _select_kernel_cls(
    backend: Fp8MoeBackend,
    config: FusedMoEConfig,
) -> type[mk.FusedMoEExperts]:
    """Select the first supported expert class for the MXFP8 config."""
    _, experts_cls = Mxfp8MoEKernelOracle(
        config, kMxfp8Static, kMxfp8Dynamic
    ).return_or_raise(backend)
    return experts_cls

select_mxfp8_moe_backend(config)

Select the MXFP8 MoE backend and the best expert class.

Returns:

Source code in vllm/model_executor/layers/fused_moe/oracle/mxfp8.py 
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
def select_mxfp8_moe_backend(
    config: FusedMoEConfig,
) -> tuple[Fp8MoeBackend, type[mk.FusedMoEExperts]]:
    """Select the MXFP8 MoE backend and the best expert class.

    Returns:
        A tuple of (fp8_backend, experts_cls).

    """
    runner_backend = config.moe_backend
    if runner_backend != "auto":
        backend = _BACKEND_NAME_MAP.get(runner_backend)
        if backend is None:
            raise ValueError(
                f"moe_backend='{runner_backend}' is not supported for "
                f"MXFP8 MoE. Expected one of "
                f"{list(_BACKEND_NAME_MAP.keys())}."
            )
        return backend, _select_kernel_cls(backend, config)

    # Auto-select: pick the first supported backend.
    backends = prioritize_humming(list(_SUPPORTED_BACKENDS))
    for backend in backends:
        try:
            experts_cls = _select_kernel_cls(backend, config)
        except ValueError:
            continue
        return backend, experts_cls

    raise ValueError("No MXFP8 MoE backends available.")