[Bugfix][ROCm][MoE] Fall back instead of crashing when AITER MoE is requested for a non-gated (is_act_and_mul=False) model (#56590)
Signed-off-by: Shantipriya Parida <shantipriya.parida@amd.com>
This commit is contained in:
@@ -6,6 +6,7 @@ import pytest
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
from tests.kernels.moe.utils import make_dummy_moe_config
|
from tests.kernels.moe.utils import make_dummy_moe_config
|
||||||
|
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
|
||||||
from vllm.model_executor.layers.fused_moe.config import RoutingMethodType
|
from vllm.model_executor.layers.fused_moe.config import RoutingMethodType
|
||||||
from vllm.model_executor.layers.fused_moe.oracle.unquantized import (
|
from vllm.model_executor.layers.fused_moe.oracle.unquantized import (
|
||||||
UnquantizedMoeBackend,
|
UnquantizedMoeBackend,
|
||||||
@@ -228,6 +229,65 @@ def test_select_rocm_aiter_backend(mock_aiter_enabled, mock_has_flashinfer):
|
|||||||
assert expert_cls is not None
|
assert expert_cls is not None
|
||||||
|
|
||||||
|
|
||||||
|
@patch(
|
||||||
|
"vllm.utils.flashinfer.has_flashinfer",
|
||||||
|
return_value=False,
|
||||||
|
)
|
||||||
|
@patch(
|
||||||
|
"vllm.model_executor.layers.fused_moe.oracle.unquantized.rocm_aiter_ops."
|
||||||
|
"is_fused_moe_enabled",
|
||||||
|
return_value=True,
|
||||||
|
)
|
||||||
|
@patch(
|
||||||
|
"vllm.model_executor.layers.fused_moe.oracle.unquantized.rocm_aiter_ops."
|
||||||
|
"is_rdna_aiter_enabled",
|
||||||
|
return_value=False,
|
||||||
|
)
|
||||||
|
@pytest.mark.skipif(
|
||||||
|
not current_platform.is_rocm(), reason="ROCm-specific backend selection test"
|
||||||
|
)
|
||||||
|
def test_select_rocm_aiter_backend_non_gated_activation_falls_back(
|
||||||
|
mock_rdna_disabled, mock_aiter_enabled, mock_has_flashinfer, monkeypatch
|
||||||
|
):
|
||||||
|
"""Test ROCm backend selection falls back (not raises) for non-gated MoE."""
|
||||||
|
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
|
||||||
|
monkeypatch.setenv("VLLM_ROCM_USE_AITER_MOE", "1")
|
||||||
|
|
||||||
|
with patch(
|
||||||
|
"vllm.model_executor.layers.fused_moe.oracle.unquantized.current_platform"
|
||||||
|
) as mock_platform:
|
||||||
|
mock_platform.is_cuda.return_value = False
|
||||||
|
mock_platform.is_rocm.return_value = True
|
||||||
|
mock_platform.is_cpu.return_value = False
|
||||||
|
mock_platform.is_xpu.return_value = False
|
||||||
|
mock_platform.is_tpu.return_value = False
|
||||||
|
mock_platform.is_out_of_tree.return_value = False
|
||||||
|
|
||||||
|
moe_config = make_dummy_moe_config(activation=MoEActivation.SILU_NO_MUL)
|
||||||
|
assert moe_config.is_act_and_mul is False
|
||||||
|
|
||||||
|
selected_backend, expert_cls = select_unquantized_moe_backend(
|
||||||
|
moe_config=moe_config,
|
||||||
|
)
|
||||||
|
|
||||||
|
assert selected_backend != UnquantizedMoeBackend.AITER
|
||||||
|
assert expert_cls is not None
|
||||||
|
|
||||||
|
|
||||||
|
@pytest.mark.skipif(
|
||||||
|
not current_platform.is_rocm(), reason="ROCm-specific backend selection test"
|
||||||
|
)
|
||||||
|
def test_explicit_aiter_backend_non_gated_activation_still_raises():
|
||||||
|
"""Explicit `--moe-backend aiter` still raises for non-gated MoE;
|
||||||
|
only the env-var opt-in path falls back."""
|
||||||
|
moe_config = make_dummy_moe_config(activation=MoEActivation.SILU_NO_MUL)
|
||||||
|
moe_config.moe_backend = "aiter" # explicit pin, not "auto"
|
||||||
|
assert moe_config.is_act_and_mul is False
|
||||||
|
|
||||||
|
with pytest.raises(ValueError):
|
||||||
|
select_unquantized_moe_backend(moe_config=moe_config)
|
||||||
|
|
||||||
|
|
||||||
@patch(
|
@patch(
|
||||||
"vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsMonolithic.is_supported_config",
|
"vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsMonolithic.is_supported_config",
|
||||||
return_value=(True, None),
|
return_value=(True, None),
|
||||||
|
|||||||
@@ -305,7 +305,33 @@ def select_unquantized_moe_backend(
|
|||||||
AVAILABLE_BACKENDS.remove(UnquantizedMoeBackend.AITER)
|
AVAILABLE_BACKENDS.remove(UnquantizedMoeBackend.AITER)
|
||||||
else:
|
else:
|
||||||
backend = UnquantizedMoeBackend.AITER
|
backend = UnquantizedMoeBackend.AITER
|
||||||
return _return_or_raise(backend, moe_config, activation_format)
|
reason = None
|
||||||
|
for k_cls in backend_to_kernel_cls(backend):
|
||||||
|
supported, reason = k_cls.is_supported_config(
|
||||||
|
k_cls, moe_config, None, None, activation_format
|
||||||
|
)
|
||||||
|
if supported:
|
||||||
|
logger.info_once(_make_log_backend(backend))
|
||||||
|
return backend, k_cls
|
||||||
|
# AITER was explicitly requested but does not support this
|
||||||
|
# deployment configuration (e.g. a non-gated MoE activation,
|
||||||
|
# which AiterExperts._supports_no_act_and_mul() rejects
|
||||||
|
# unconditionally). Rather than hard-crashing at engine init,
|
||||||
|
# warn — explicitly stating that this is falling back, and
|
||||||
|
# which backends will be tried next — and fall back to the
|
||||||
|
# remaining backends in priority order below.
|
||||||
|
logger.warning_once(
|
||||||
|
"VLLM_ROCM_USE_AITER_MOE=1 was requested, but %s "
|
||||||
|
"Falling back to try the remaining available MoE backends: %s.",
|
||||||
|
_make_log_unsupported(backend, reason),
|
||||||
|
", ".join(
|
||||||
|
b.value
|
||||||
|
for b in AVAILABLE_BACKENDS
|
||||||
|
if b != UnquantizedMoeBackend.AITER
|
||||||
|
),
|
||||||
|
)
|
||||||
|
if UnquantizedMoeBackend.AITER in AVAILABLE_BACKENDS:
|
||||||
|
AVAILABLE_BACKENDS.remove(UnquantizedMoeBackend.AITER)
|
||||||
|
|
||||||
for backend in AVAILABLE_BACKENDS:
|
for backend in AVAILABLE_BACKENDS:
|
||||||
for k_cls in backend_to_kernel_cls(backend):
|
for k_cls in backend_to_kernel_cls(backend):
|
||||||
|
|||||||
Reference in New Issue
Block a user