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

23
vllm/model_executor/layers/fused_moe/flashinfer_trtllm_moe.py
View File

@@ -129,25 +129,28 @@ def is_supported_config_trtllm_fp8(
return f"kernel does not support {reason}"
if not _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 _supports_no_act_and_mul()):
return False, _make_reason("no act_and_mul MLP layer")
elif not _supports_activation(moe_config.activation):
return False, _make_reason(f"{moe_config.activation} activation")
elif not _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 _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 not _supports_routing_method(
weight_key, activation_key, moe_config.routing_method
):
return False, _make_reason("routing method")
return False, _make_reason(f"routing method {moe_config.routing_method}")
elif activation_format != mk.FusedMoEActivationFormat.Standard:
return False, _make_reason("activation format")
return False, _make_reason(f"activation format {activation_format}")
elif not _supports_router_logits_dtype(
moe_config.router_logits_dtype, moe_config.routing_method
):
return False, _make_reason("float32 router_logits with non-DeepSeekV3 routing")
return False, _make_reason(
"float32 router_logits with non-DeepSeekV3 routing "
f"{moe_config.router_logits_dtype}x{moe_config.routing_method}"
)
return True, None
@@ -165,17 +168,17 @@ def is_supported_config_trtllm_bf16(
return f"kernel does not support {reason}"
if not _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 _supports_no_act_and_mul()):
return False, _make_reason("no act_and_mul MLP layer")
elif not _supports_activation(moe_config.activation):
return False, _make_reason(f"{moe_config.activation} activation")
elif not _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 not _supports_routing_method_bf16(moe_config.routing_method):
return False, _make_reason("routing method")
return False, _make_reason(f"routing method {moe_config.routing_method}")
elif activation_format != mk.FusedMoEActivationFormat.Standard:
return False, _make_reason("activation format")
return False, _make_reason(f"activation format {activation_format}")
return True, None

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

14
vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
View File

@@ -109,21 +109,23 @@ def is_supported_config_trtllm(
return f"kernel does not support {reason}"
if not _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 _supports_no_act_and_mul()):
return False, _make_reason("no act_and_mul MLP layer")
elif not _supports_activation(moe_config.activation):
return False, _make_reason(f"{moe_config.activation} activation")
elif not _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 _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 not _supports_routing_method(moe_config.routing_method):
return False, _make_reason("routing method")
return False, _make_reason(f"routing method {moe_config.routing_method}")
elif activation_format != mk.FusedMoEActivationFormat.Standard:
return False, _make_reason("activation format")
return False, _make_reason(f"activation format {activation_format}")
elif moe_config.hidden_dim % 512 != 0:
return False, _make_reason("hidden_dim must be divisible by 512")
return False, _make_reason(
f"hidden_dim must be divisible by 512, found {moe_config.hidden_dim}"
)
return True, None