[Bugfix] Use latency MOE backend as default for Flashinfer and other misc fixes (#27439)

Signed-off-by: Pavani Majety <pmajety@nvidia.com>
This commit is contained in:
Pavani Majety
2025-11-07 04:18:39 -08:00
committed by GitHub
parent e0919f331d
commit 72b1c2ae2c
7 changed files with 47 additions and 12 deletions
@@ -138,6 +138,13 @@ def get_fp8_moe_backend(block_quant: bool) -> Fp8MoeBackend:
logger.info_once("Using FlashInfer FP8 MoE TRTLLM backend for SM100")
return Fp8MoeBackend.FLASHINFER_TRTLLM
else:
if block_quant:
raise ValueError(
"FlashInfer FP8 MoE throughput backend does not "
"support block quantization. Please use "
"VLLM_FLASHINFER_MOE_BACKEND=latency "
"instead."
)
logger.info_once("Using FlashInfer FP8 MoE CUTLASS backend for SM100")
return Fp8MoeBackend.FLASHINFER_CUTLASS