biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[UX] More descriptive reasons in is_supported_config for MoE (#34908)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2026-02-19 18:20:52 -05:00
committed by GitHub
parent 662205d34e
commit f72061a19a
3 changed files with 29 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/model_executor/layers/fused_moe/modular_kernel.py
View File

@@ -29,6 +29,7 @@ from vllm.model_executor.layers.fused_moe.utils import (
from vllm.model_executor.layers.quantization.utils.quant_utils import (
QuantKey,
)
from vllm.platforms import current_platform
from vllm.utils.math_utils import cdiv
from vllm.v1.worker.ubatching import (
dbo_enabled,
@@ -498,15 +499,19 @@ class FusedMoEPermuteExpertsUnpermute(ABC):
return f"kernel does not support {reason}"
if not cls._supports_current_device():
return False, _make_reason("current device")
return False, _make_reason(f"current device {current_platform.device_name}")
elif not (moe_config.is_act_and_mul or cls._supports_no_act_and_mul()):
return False, _make_reason("no act_and_mul MLP layer")
elif not cls._supports_activation(moe_config.activation):
return False, _make_reason(f"{moe_config.activation} activation")
elif not cls._supports_quant_scheme(weight_key, activation_key):
return False, _make_reason("quantization scheme")
return False, _make_reason(
f"quantization scheme {weight_key}x{activation_key}"
)
elif not cls._supports_parallel_config(moe_config.moe_parallel_config):
return False, _make_reason("parallel config")
return False, _make_reason(
f"parallel config {moe_config.moe_parallel_config}"
)
elif activation_format != cls.activation_format():
return False, _make_reason(f"{activation_format.value} activation format")
return True, None