vllm.model_executor.layers.fused_moe.oracle.fp8 ¶

def _get_priority_backends(
    moe_config: FusedMoEConfig,
    weight_key: QuantKey | None,
    activation_key: QuantKey | None,
) -> list[Fp8MoeBackend]:
    """
    Get available backends in priority order based on platform and config.

    This function can be extended to become more complex as needed.
    """

    _AVAILABLE_BACKENDS = [
        Fp8MoeBackend.AITER,
        Fp8MoeBackend.FLASHINFER_TRTLLM,
        Fp8MoeBackend.FLASHINFER_CUTLASS,
        Fp8MoeBackend.DEEPGEMM,
        Fp8MoeBackend.VLLM_CUTLASS,
        Fp8MoeBackend.TRITON,
        Fp8MoeBackend.MARLIN,
        Fp8MoeBackend.BATCHED_DEEPGEMM,
        Fp8MoeBackend.BATCHED_VLLM_CUTLASS,
        Fp8MoeBackend.BATCHED_TRITON,
        Fp8MoeBackend.XPU,
        Fp8MoeBackend.CPU,
    ]

    def _move_to_front(backends: list[Fp8MoeBackend], backend: Fp8MoeBackend) -> None:
        backends.insert(0, backends.pop(backends.index(backend)))

    # With DeepEP v2 contiguous layout (do_expand=False), tensors are
    # worst-case allocated with padding. TrtLLM's tile-level skipping
    # avoids wasted compute on padding rows; other backends process all rows.
    if (
        current_platform.is_cuda()
        and current_platform.is_device_capability_family(100)
        and moe_config.moe_parallel_config.use_deepep_v2_kernels
        and activation_key == kFp8Dynamic128Sym
        and weight_key == kFp8Static128BlockSym
    ):
        _move_to_front(_AVAILABLE_BACKENDS, Fp8MoeBackend.FLASHINFER_TRTLLM)

    # On Hopper for Block Fp8, prefer Triton for TP and FI CUTLASS for EP.
    if (
        current_platform.is_cuda()
        and current_platform.is_device_capability(90)
        and activation_key == kFp8Dynamic128Sym
        and weight_key == kFp8Static128BlockSym
    ):
        if moe_config.moe_parallel_config.ep_size > 1:
            _move_to_front(_AVAILABLE_BACKENDS, Fp8MoeBackend.FLASHINFER_CUTLASS)
        else:
            _move_to_front(_AVAILABLE_BACKENDS, Fp8MoeBackend.TRITON)

    if current_platform.is_xpu():
        # XPU platform supports TritonExperts and XPUExpertsFp8,
        # move XPU backend to the front.
        _move_to_front(_AVAILABLE_BACKENDS, Fp8MoeBackend.XPU)

    if current_platform.is_cpu():
        # CPU platform uses FP8 W8A16 fused MoE kernel.
        _move_to_front(_AVAILABLE_BACKENDS, Fp8MoeBackend.CPU)

    return _AVAILABLE_BACKENDS

def make_fp8_moe_quant_config(
    fp8_backend: Fp8MoeBackend,
    w1_scale: torch.Tensor,
    w2_scale: torch.Tensor,
    a1_scale: torch.Tensor | None,
    a2_scale: torch.Tensor | None,
    w1_bias: torch.Tensor | None = None,
    w2_bias: torch.Tensor | None = None,
    block_shape: list[int] | None = None,
    per_act_token_quant: bool = False,
    per_out_ch_quant: bool = False,
    swiglu_limit: float | None = None,
) -> FusedMoEQuantConfig:
    """
    Create FusedMoEQuantConfig for the specified FP8 Backend.
    The FusedMoEQuantConfig holds the scales that are used
    at runtime by the Modular Kernel abstraction.

    Note that certain kernels (e.g. Flashinfer CUTLASS) need
    special Quant configs to handle non-standard inputs to
    their kernel interfaces.

    In a future PR, we will have this function should be
    a method of the modular kernel itself.
    """

    # MARLIN and CPU are mixed precision W8A16 config.
    if fp8_backend == Fp8MoeBackend.MARLIN or fp8_backend == Fp8MoeBackend.CPU:
        return fp8_w8a16_moe_quant_config(
            w1_scale=w1_scale,
            w2_scale=w2_scale,
            w1_bias=w1_bias,
            w2_bias=w2_bias,
            block_shape=block_shape,
        )

    # Flashinfer CUTLASS per-tensor uses single dq scale
    # (alpha = w_scale * a_scale) and inverse a2 scale.
    if fp8_backend == Fp8MoeBackend.FLASHINFER_CUTLASS and block_shape is None:
        assert a1_scale is not None and a2_scale is not None
        return fp8_w8a8_moe_quant_config(
            w1_scale=w1_scale,
            w2_scale=w2_scale,
            w1_bias=w1_bias,
            w2_bias=w2_bias,
            a1_scale=a1_scale,
            a2_scale=a2_scale,
            a1_gscale=(1.0 / a1_scale),
            a2_gscale=(1.0 / a2_scale),
            g1_alphas=(w1_scale * a1_scale).squeeze(),
            g2_alphas=(w2_scale * a2_scale).squeeze(),
            gemm1_clamp_limit=swiglu_limit,
        )
    # MXFP8 uses "mxfp8" quant_dtype so the prepare step dispatches to
    # _mxfp8_e4m3_quantize rather than standard FP8 block quantization.
    # Non-swizzled layout is required since the TRTLLM kernel expects
    # scales in (num_tokens, hidden_dim // 32) format.
    if block_shape == [1, 32]:
        return FusedMoEQuantConfig.make(
            "mxfp8",
            w1_scale=w1_scale,
            w2_scale=w2_scale,
            w1_bias=w1_bias,
            w2_bias=w2_bias,
            a1_scale=a1_scale,
            a2_scale=a2_scale,
            block_shape=block_shape,
            is_scale_swizzled=False,
            gemm1_clamp_limit=swiglu_limit,
        )

    # All other backends use normal config.
    return fp8_w8a8_moe_quant_config(
        w1_scale=w1_scale,
        w2_scale=w2_scale,
        w1_bias=w1_bias,
        w2_bias=w2_bias,
        a1_scale=a1_scale,
        a2_scale=a2_scale,
        block_shape=block_shape,
        per_act_token_quant=per_act_token_quant,
        per_out_ch_quant=per_out_ch_quant,
        gemm1_clamp_limit=swiglu_limit,
    )

def map_fp8_backend(runner_backend: MoEBackend) -> Fp8MoeBackend:
    """Map user's MoEBackend to Fp8MoeBackend."""
    mapping = {
        "triton": Fp8MoeBackend.TRITON,
        "deep_gemm": Fp8MoeBackend.DEEPGEMM,
        "cutlass": Fp8MoeBackend.VLLM_CUTLASS,
        "flashinfer_trtllm": Fp8MoeBackend.FLASHINFER_TRTLLM,
        "flashinfer_cutlass": Fp8MoeBackend.FLASHINFER_CUTLASS,
        "marlin": Fp8MoeBackend.MARLIN,
        "aiter": Fp8MoeBackend.AITER,
    }
    if backend := mapping.get(runner_backend):
        return backend
    raise ValueError(
        f"moe_backend='{runner_backend}' is not supported for FP8 MoE. "
        f"Expected one of {list(mapping.keys())}."
    )

def select_fp8_moe_backend(
    config: FusedMoEConfig,
    weight_key: QuantKey | None,
    activation_key: QuantKey | None,
    allow_vllm_cutlass: bool = False,
) -> tuple[Fp8MoeBackend, type[mk.FusedMoEExperts] | None]:
    """
    Select the primary FP8 MoE backend
    Note: Shape-specific fallbacks may still occur at runtime.
    """

    # NOTE: the kernels are selected in the following order.
    AVAILABLE_BACKENDS = _get_priority_backends(config, weight_key, activation_key)

    # NOTE(rob): We need to peak into the P/F selection to determine
    # if we are using the batched or standard expert format, which
    # if not ideal. Once we unify TP + DP/EP, we can select P/F first.
    activation_format = (
        mk.FusedMoEActivationFormat.BatchedExperts
        if config.moe_parallel_config.use_batched_activation_format
        else mk.FusedMoEActivationFormat.Standard
    )

    def _make_log_backend(backend: Fp8MoeBackend):
        available_backend_strs = [b.value for b in AVAILABLE_BACKENDS]
        return (
            f"Using {backend.value} Fp8 MoE backend out "
            f"of potential backends: {available_backend_strs}."
        )

    def _make_log_unsupported(backend: Fp8MoeBackend, reason: str | None) -> str:
        if reason:
            return (
                f"FP8 MoE backend {backend.value} does not support the "
                f"deployment configuration since {reason}."
            )
        else:
            return (
                f"FP8 MoE backend '{backend.value}' does not support the "
                "deployment configuration."
            )

    def _return_or_raise(
        backend: Fp8MoeBackend,
        config: FusedMoEConfig,
        weight_key: QuantKey | None,
        activation_key: QuantKey | None,
        activation_format: mk.FusedMoEActivationFormat,
    ) -> tuple[Fp8MoeBackend, type[mk.FusedMoEExperts]]:
        for k_cls in backend_to_kernel_cls(backend):
            supported, reason = k_cls.is_supported_config(
                k_cls, config, weight_key, activation_key, activation_format
            )
            if supported:
                logger.info_once(_make_log_backend(backend))
                return backend, k_cls
        raise ValueError(_make_log_unsupported(backend, reason))

    # Handle explicit moe_backend from user.
    runner_backend = config.moe_backend
    if runner_backend != "auto":
        requested_backend = map_fp8_backend(runner_backend)
        # For batched activation format, use batched variants if available.
        if activation_format == mk.FusedMoEActivationFormat.BatchedExperts:
            if requested_backend == Fp8MoeBackend.DEEPGEMM:
                requested_backend = Fp8MoeBackend.BATCHED_DEEPGEMM
            elif requested_backend == Fp8MoeBackend.TRITON:
                requested_backend = Fp8MoeBackend.BATCHED_TRITON
            elif requested_backend == Fp8MoeBackend.VLLM_CUTLASS:
                requested_backend = Fp8MoeBackend.BATCHED_VLLM_CUTLASS

        if (
            requested_backend
            in [
                Fp8MoeBackend.VLLM_CUTLASS,
                Fp8MoeBackend.BATCHED_VLLM_CUTLASS,
            ]
            and not allow_vllm_cutlass
        ):
            raise ValueError(
                "vLLM CUTLASS FP8 MoE backend is disabled for this configuration."
            )

        return _return_or_raise(
            requested_backend, config, weight_key, activation_key, activation_format
        )

    # Handle explicit DeepGEMM FP8 configuration.
    if envs.is_set("VLLM_USE_DEEP_GEMM") or envs.is_set("VLLM_MOE_USE_DEEP_GEMM"):
        if not envs.VLLM_USE_DEEP_GEMM or not envs.VLLM_MOE_USE_DEEP_GEMM:
            AVAILABLE_BACKENDS.remove(Fp8MoeBackend.DEEPGEMM)
            AVAILABLE_BACKENDS.remove(Fp8MoeBackend.BATCHED_DEEPGEMM)
        else:
            backend = (
                Fp8MoeBackend.DEEPGEMM
                if activation_format == mk.FusedMoEActivationFormat.Standard
                else Fp8MoeBackend.BATCHED_DEEPGEMM
            )
            return _return_or_raise(
                backend, config, weight_key, activation_key, activation_format
            )

    # Handle explicit MARLIN FP8 configuration.
    if envs.VLLM_TEST_FORCE_FP8_MARLIN:
        backend = Fp8MoeBackend.MARLIN
        return _return_or_raise(
            backend, config, weight_key, activation_key, activation_format
        )

    # Handle explicit AITER FP8 configuration.
    if envs.is_set("VLLM_ROCM_USE_AITER") or envs.is_set("VLLM_ROCM_USE_AITER_MOE"):
        if not envs.VLLM_ROCM_USE_AITER or not envs.VLLM_ROCM_USE_AITER_MOE:
            AVAILABLE_BACKENDS.remove(Fp8MoeBackend.AITER)
        else:
            backend = Fp8MoeBackend.AITER
            return _return_or_raise(
                backend, config, weight_key, activation_key, activation_format
            )

    if not allow_vllm_cutlass:
        AVAILABLE_BACKENDS.remove(Fp8MoeBackend.VLLM_CUTLASS)
        AVAILABLE_BACKENDS.remove(Fp8MoeBackend.BATCHED_VLLM_CUTLASS)

    # Select kernels in order of backend.
    for backend in AVAILABLE_BACKENDS:
        for k_cls in backend_to_kernel_cls(backend):
            supported, reason = k_cls.is_supported_config(
                k_cls,
                config,
                weight_key,
                activation_key,
                activation_format,
            )
            if supported:
                logger.info_once(_make_log_backend(backend))
                return backend, k_cls
            else:
                logger.debug_once(_make_log_unsupported(backend, reason))

    # TODO(rob): per discussion with TPU team, we need a way to register
    # MoE backends by OOT plugins, rather than having an explicit list
    # of AVAILABLE_BACKENDS. Enabling returning `Fp8MoeBackend.NONE` is
    # a temporary measure until these register APIs are complete.
    if current_platform.is_cuda() or current_platform.is_rocm():
        raise NotImplementedError(
            "No FP8 MoE backend supports the deployment configuration."
        )

    return Fp8MoeBackend.NONE, None