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 14, 05:30 UTC
[Bugfix]: Assertion error when using FlashInfer backend (#25933)
Signed-off-by: simondanielsson <simon.danielsson99@hotmail.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Simon Danielsson committed
Oct 5, 2025 at 09:46 UTC
432e1cbc23240dd4a002c93c5513c6b353d00b4a
1 file changed
+2
-2
vllm/model_executor/layers/quantization/fp8.py
+2
-2
@@ -508,7 +508,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
508
self.layer = layer
509
self.quant_config = quant_config
510
self.weight_block_size = self.quant_config.weight_block_size
511
- self.block_quant = self.weight_block_size is not None
511
+ self.block_quant: bool = self.weight_block_size is not None
512
513
self.fused_experts: Optional[
514
mk.FusedMoEModularKernel] = None # type: ignore
@@ -1094,7 +1094,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
1094
expert_map=expert_map,
1095
)
1096
elif self.flashinfer_moe_backend == FlashinferMoeBackend.CUTLASS:
1097
- assert self.block_quant is None
1097
+ assert not self.block_quant
1098
assert (not renormalize and custom_routing_function is not None)
1099
assert activation == 'silu', (
1100
f"Expected 'silu' activation but got {activation}")