[torch.compile] Unwrap fused_marlin_moe custom op (#26739)
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com> Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
This commit is contained in:
co-authored by
Varun Sundar Rabindranath
parent
8a0af6a561
commit
8ae169286f
@@ -26,6 +26,7 @@ from vllm.model_executor.layers.fused_moe.config import (
|
||||
FusedMoEQuantConfig,
|
||||
fp8_w8a8_moe_quant_config,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_moe
|
||||
from vllm.model_executor.layers.fused_moe.layer import UnquantizedFusedMoEMethod
|
||||
from vllm.model_executor.layers.linear import (
|
||||
LinearBase,
|
||||
@@ -1196,7 +1197,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
|
||||
elif self.use_marlin:
|
||||
assert activation == "silu", f"{activation} not supported for Marlin MoE."
|
||||
assert self.fused_experts is None
|
||||
result = torch.ops.vllm.fused_marlin_moe(
|
||||
result = fused_marlin_moe(
|
||||
x,
|
||||
layer.w13_weight,
|
||||
layer.w2_weight,
|
||||
|
||||
Reference in New Issue
Block a user