[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:
Varun Sundar Rabindranath
2025-10-14 02:22:16 +00:00
committed by GitHub
co-authored by Varun Sundar Rabindranath
parent 8a0af6a561
commit 8ae169286f
10 changed files with 22 additions and 52 deletions
@@ -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,