A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 09:45 UTC
[Bugfix] [ROCm]: Remove assertion logic when using AITER fused moe in unquantizedMethod to reenable LLama4 BF16 (#18205)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
TJian committed
May 16, 2025 at 00:53 UTC
92540529c051fe6e8f111d7688ffac84ac561a0d
1 file changed
-1
vllm/model_executor/layers/fused_moe/layer.py
-1
@@ -503,7 +503,6 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
503
indices_type=torch.uint32 if self.moe.use_pplx_kernels else None)
504
505
if self.rocm_aiter_moe_enabled:
506
- assert not apply_router_weight_on_input
506
assert expert_map is None
507
return self.rocm_aiter_fused_experts(
508
hidden_states=x,