[ROCm][Perf] Use flydsl moe with Minimax-M3 mxfp8 weights on gfx950 and implemented moe-backend selection (#46184)
Signed-off-by: Hongxia Yang <hongxia.yang@amd.com> Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: TJian <tunjian.tan@embeddedllm.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com> Co-authored-by: Tan Pin Siang <tanpinsiang@gmail.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
TJian
Andreas Karatzas
Tan Pin Siang
parent
9fd00ee006
commit
867fd5e8ed
@@ -0,0 +1,135 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# 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
|
||||
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.
|
||||
"""
|
||||
|
||||
import dataclasses
|
||||
from unittest.mock import patch
|
||||
|
||||
import pytest
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
if not current_platform.is_rocm():
|
||||
pytest.skip("This test can only run on ROCm.", allow_module_level=True)
|
||||
|
||||
from tests.kernels.moe.utils import make_dummy_moe_config # noqa: E402
|
||||
from vllm.model_executor.layers.fused_moe.experts.aiter_mxfp8_moe import ( # noqa: E402
|
||||
AiterMxfp8Experts,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.modular_kernel import ( # noqa: E402
|
||||
FusedMoEActivationFormat,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.oracle.fp8 import ( # noqa: E402
|
||||
Fp8MoeBackend,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.oracle.mxfp8 import ( # noqa: E402
|
||||
_BACKEND_NAME_MAP,
|
||||
_SUPPORTED_BACKENDS,
|
||||
_mxfp8_backend_to_kernel_cls,
|
||||
_select_kernel_cls,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import ( # noqa: E402
|
||||
kMxfp8Dynamic,
|
||||
kMxfp8Static,
|
||||
)
|
||||
|
||||
_AITER_MOD = "vllm.model_executor.layers.fused_moe.experts.aiter_mxfp8_moe"
|
||||
|
||||
|
||||
def _config(ep_size: int = 1):
|
||||
cfg = make_dummy_moe_config(num_experts=128, experts_per_token=4, hidden_dim=6144)
|
||||
if ep_size != 1:
|
||||
cfg = dataclasses.replace(
|
||||
cfg,
|
||||
moe_parallel_config=dataclasses.replace(
|
||||
cfg.moe_parallel_config, ep_size=ep_size, use_ep=True
|
||||
),
|
||||
)
|
||||
return cfg
|
||||
|
||||
|
||||
def _gfx950():
|
||||
"""Patch the platform so the device gate (gfx950 / MX) passes off-ROCm."""
|
||||
return patch.multiple(
|
||||
f"{_AITER_MOD}.current_platform",
|
||||
is_rocm=lambda: True,
|
||||
supports_mx=lambda: True,
|
||||
)
|
||||
|
||||
|
||||
def _flydsl_installed(present: bool):
|
||||
return patch(f"{_AITER_MOD}.is_aiter_mxfp8_moe_available", return_value=present)
|
||||
|
||||
|
||||
def test_aiter_mxfp8_registered():
|
||||
"""The FlyDSL backend is auto-selectable and reachable via --moe-backend aiter."""
|
||||
assert Fp8MoeBackend.AITER_MXFP8 in _SUPPORTED_BACKENDS
|
||||
assert _BACKEND_NAME_MAP["aiter"] is Fp8MoeBackend.AITER_MXFP8
|
||||
assert _mxfp8_backend_to_kernel_cls(Fp8MoeBackend.AITER_MXFP8) == [
|
||||
AiterMxfp8Experts
|
||||
]
|
||||
|
||||
|
||||
def test_triton_selectable():
|
||||
assert _BACKEND_NAME_MAP["triton"] is Fp8MoeBackend.TRITON_MXFP8
|
||||
# Not auto-selected (only reachable explicitly), so FlyDSL still wins auto.
|
||||
assert Fp8MoeBackend.TRITON_MXFP8 not in _SUPPORTED_BACKENDS
|
||||
|
||||
|
||||
@pytest.mark.parametrize("ep_size", [1, 2])
|
||||
def test_ep_supported(ep_size):
|
||||
"""FlyDSL accepts both TP and EP: apply() forwards expert_map as expert_mask."""
|
||||
assert (
|
||||
AiterMxfp8Experts._supports_parallel_config(
|
||||
_config(ep_size).moe_parallel_config
|
||||
)
|
||||
is True
|
||||
)
|
||||
|
||||
|
||||
@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
|
||||
],
|
||||
)
|
||||
def test_is_supported_config(present, ep_size, supported, reason_substr):
|
||||
with _gfx950(), _flydsl_installed(present):
|
||||
ok, reason = AiterMxfp8Experts.is_supported_config(
|
||||
AiterMxfp8Experts,
|
||||
_config(ep_size),
|
||||
kMxfp8Static,
|
||||
kMxfp8Dynamic,
|
||||
FusedMoEActivationFormat.Standard,
|
||||
)
|
||||
assert ok is supported
|
||||
if reason_substr is not None:
|
||||
assert reason_substr in reason
|
||||
|
||||
|
||||
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):
|
||||
assert (
|
||||
_select_kernel_cls(Fp8MoeBackend.AITER_MXFP8, _config(1))
|
||||
is AiterMxfp8Experts
|
||||
)
|
||||
assert (
|
||||
_select_kernel_cls(Fp8MoeBackend.AITER_MXFP8, _config(2))
|
||||
is AiterMxfp8Experts
|
||||
)
|
||||
with (
|
||||
_gfx950(),
|
||||
_flydsl_installed(False),
|
||||
pytest.raises(ValueError, match="flydsl package"),
|
||||
):
|
||||
_select_kernel_cls(Fp8MoeBackend.AITER_MXFP8, _config(1))
|
||||
Reference in New Issue
Block a user