diff --git a/tests/kernels/moe/test_mxfp8_aiter_backend_selection.py b/tests/kernels/moe/test_mxfp8_aiter_backend_selection.py index 51aff4ea3b..61b26af62c 100644 --- a/tests/kernels/moe/test_mxfp8_aiter_backend_selection.py +++ b/tests/kernels/moe/test_mxfp8_aiter_backend_selection.py @@ -2,10 +2,11 @@ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project """MXFP8 MoE backend selection for the AITER FlyDSL kernel (gfx950). -GPU-free: mocks the platform (gfx950) and the ``flydsl`` package check, then +GPU-free: mocks the platform (gfx950) and the aiter fused-MoE enable flag, then exercises the oracle so the FlyDSL backend is auto-picked when usable (including under expert parallelism, since apply() forwards the expert_map as aiter's -expert_mask) and skipped (native fallback) when the device/package is missing. +expert_mask) and skipped (native fallback) when the device is unsupported or the +aiter runtime is disabled. """ import dataclasses @@ -48,6 +49,8 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import ( # noqa: kMxfp8Static, ) +from vllm._aiter_ops import rocm_aiter_ops # noqa: E402 # isort: skip + _AITER_MOD = "vllm.model_executor.layers.fused_moe.experts.aiter_mxfp8_moe" @@ -82,8 +85,8 @@ def _gfx950(): ) -def _flydsl_installed(present: bool): - return patch(f"{_AITER_MOD}.is_aiter_mxfp8_moe_available", return_value=present) +def _aiter_moe_enabled(present: bool): + return patch.object(rocm_aiter_ops, "is_fused_moe_enabled", return_value=present) def test_aiter_mxfp8_registered(): @@ -109,13 +112,18 @@ def test_ep_supported(ep_size): @pytest.mark.parametrize( "present,ep_size,supported,reason_substr", [ - (True, 1, True, None), # gfx950 + flydsl + TP -> selectable - (True, 2, True, None), # gfx950 + flydsl + EP -> selectable (expert_mask) - (False, 1, False, "flydsl package"), # package missing -> clear reason + (True, 1, True, None), # gfx950 + aiter MoE + TP -> selectable + (True, 2, True, None), # gfx950 + aiter MoE + EP -> selectable (expert_mask) + ( + False, + 1, + False, + "does not support current device", + ), # disabled -> not selected ], ) def test_is_supported_config(present, ep_size, supported, reason_substr): - with _gfx950(), _flydsl_installed(present): + with _gfx950(), _aiter_moe_enabled(present): ok, reason = AiterMxfp8Experts.is_supported_config( AiterMxfp8Experts, _config(ep_size), @@ -130,8 +138,8 @@ def test_is_supported_config(present, ep_size, supported, reason_substr): def test_explicit_moe_backend_aiter(): """--moe-backend aiter: returns FlyDSL when usable (TP or EP), else a clear - ValueError when the flydsl package is missing.""" - with _gfx950(), _flydsl_installed(True): + ValueError when the aiter MoE runtime is disabled.""" + with _gfx950(), _aiter_moe_enabled(True): assert ( _select_kernel_cls(Fp8MoeBackend.AITER_MXFP8, _config(1)) is AiterMxfp8Experts @@ -142,18 +150,17 @@ def test_explicit_moe_backend_aiter(): ) with ( _gfx950(), - _flydsl_installed(False), - pytest.raises(ValueError, match="flydsl package"), + _aiter_moe_enabled(False), + pytest.raises(ValueError, match="does not support current device"), ): _select_kernel_cls(Fp8MoeBackend.AITER_MXFP8, _config(1)) def test_gfx950_picks_aiter(): - """Auto-select on real ROCm hardware with flydsl usable -> FlyDSL wins.""" - # NOTE: Fp8MoeBackend.AITER_MXFP8 does not require VLLM_ROCM_USE_AITER=1 + """Auto-select on real ROCm hardware with aiter MoE enabled -> FlyDSL wins.""" with ( patch(f"{_AITER_MOD}.current_platform.supports_mx", return_value=True), - _flydsl_installed(True), + _aiter_moe_enabled(True), ): backend, experts_cls = select_mxfp8_moe_backend(_config()) assert backend is Fp8MoeBackend.AITER_MXFP8 diff --git a/tests/models/quantization/test_mxfp8.py b/tests/models/quantization/test_mxfp8.py index c12a72a09c..0463468eae 100644 --- a/tests/models/quantization/test_mxfp8.py +++ b/tests/models/quantization/test_mxfp8.py @@ -111,11 +111,6 @@ def test_mxfp8_aiter_requires_swigluoai_activation( "_supports_current_device", staticmethod(lambda: True), ) - monkeypatch.setattr( - aiter_mxfp8_moe, - "is_aiter_mxfp8_moe_available", - lambda: True, - ) config = FusedMoEConfig( num_experts=8, @@ -164,11 +159,6 @@ def test_mxfp8_aiter_requires_swigluoai_params( "_supports_current_device", staticmethod(lambda: True), ) - monkeypatch.setattr( - aiter_mxfp8_moe, - "is_aiter_mxfp8_moe_available", - lambda: True, - ) config = FusedMoEConfig( num_experts=8, @@ -218,11 +208,6 @@ def test_mxfp8_aiter_accepts_swigluoai_params( "_supports_current_device", staticmethod(lambda: True), ) - monkeypatch.setattr( - aiter_mxfp8_moe, - "is_aiter_mxfp8_moe_available", - lambda: True, - ) config = FusedMoEConfig( num_experts=8, diff --git a/vllm/model_executor/layers/fused_moe/experts/aiter_mxfp8_moe.py b/vllm/model_executor/layers/fused_moe/experts/aiter_mxfp8_moe.py index 93a4088fc4..c15a1a13d4 100644 --- a/vllm/model_executor/layers/fused_moe/experts/aiter_mxfp8_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/aiter_mxfp8_moe.py @@ -11,6 +11,7 @@ import math import torch import vllm.model_executor.layers.fused_moe.modular_kernel as mk +from vllm._aiter_ops import rocm_aiter_ops from vllm.logger import init_logger from vllm.model_executor.layers.fused_moe.activation import MoEActivation from vllm.model_executor.layers.fused_moe.experts.mxfp8_emulation_moe import ( @@ -24,41 +25,6 @@ _AITER_SWIGLU_ALPHA = 1.702 _AITER_SWIGLU_BETA = 1.0 -def is_aiter_mxfp8_moe_available() -> bool: - """True when the FlyDSL MXFP8 MoE can run here: gfx950, the ``flydsl`` - package is importable, AND the installed aiter carries the mxfp8 FlyDSL - 2-stage support from ROCm/aiter#3811. - - ``flydsl`` and ``aiter`` are separate packages, so ``is_flydsl_available()`` - (flydsl pkg + arch) is necessary but not sufficient: an older aiter without - #3811 still ships the flydsl pkg and the ``aiter.ops.flydsl`` module but a - broken/missing ``per_1x32 + fp8`` 2-stage path. Without this extra gate a - nightly lacking #3811 would wrongly select FlyDSL instead of falling back to - the native Triton dot_scaled path. #3811 added no probe-able public symbol, - so detect the ``minimax_m3_mxfp8`` tuned config it shipped. Every check fails - closed (returns False -> triton dot_scaled), which is always safe.""" - if not (current_platform.is_rocm() and current_platform.supports_mx()): - return False - try: - import os - - import aiter - from aiter.ops.flydsl.utils import is_flydsl_available - - if not is_flydsl_available(): - return False - return os.path.exists( - os.path.join( - os.path.dirname(aiter.__file__), - "configs", - "model_configs", - "minimax_m3_mxfp8_tuned_fmoe.csv", - ) - ) - except Exception: - return False - - class AiterMxfp8Experts(Mxfp8TritonExpertsBase): """MXFP8 MoE through AITER's FlyDSL two-stage grouped GEMM (gfx950).""" @@ -79,10 +45,7 @@ class AiterMxfp8Experts(Mxfp8TritonExpertsBase): @staticmethod def _supports_current_device() -> bool: - # Device capability only (gfx950 / MX-capable ROCm). The flydsl package - # check lives in is_supported_config so a missing package is reported - # distinctly from an unsupported device. - return current_platform.is_rocm() and current_platform.supports_mx() + return current_platform.supports_mx() and rocm_aiter_ops.is_fused_moe_enabled() @staticmethod def _supports_parallel_config(moe_parallel_config) -> bool: @@ -95,12 +58,6 @@ class AiterMxfp8Experts(Mxfp8TritonExpertsBase): is_supported, reason = super().is_supported_config( cls, moe_config, weight_key, activation_key, activation_format ) - # _supports_current_device() only gates on the device; surface a clear - # reason when the device is fine but the flydsl package is missing. - if is_supported and not is_aiter_mxfp8_moe_available(): - return False, ( - "kernel requires the aiter flydsl package, which is not installed" - ) if ( is_supported and moe_config.activation != MoEActivation.SWIGLUOAI_UNINTERLEAVE