[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
+3 -3
View File
@@ -155,7 +155,7 @@ if TYPE_CHECKING:
VLLM_USE_FLASHINFER_MOE_FP16: bool = False
VLLM_USE_FLASHINFER_MOE_FP8: bool = False
VLLM_USE_FLASHINFER_MOE_FP4: bool = False
VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency"] = "throughput"
VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency"] = "latency"
VLLM_XGRAMMAR_CACHE_MB: int = 0
VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
@@ -1218,7 +1218,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
# - "latency":
# Uses TensorRT-LLM kernels optimized for low-latency inference.
"VLLM_FLASHINFER_MOE_BACKEND": env_with_choices(
"VLLM_FLASHINFER_MOE_BACKEND", "throughput", ["throughput", "latency"]
"VLLM_FLASHINFER_MOE_BACKEND", "latency", ["throughput", "latency"]
),
# Control the maximum number of tokens per expert supported by the
# NVFP4 MoE CUTLASS Kernel. This value is used to create a buffer for
@@ -1325,7 +1325,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_NVFP4_GEMM_BACKEND": env_with_choices(
"VLLM_NVFP4_GEMM_BACKEND",
None,
["flashinfer-cudnn", "flashinfer-trtllm", "flashinfer-cutlass"],
["flashinfer-cudnn", "flashinfer-trtllm", "flashinfer-cutlass", "cutlass"],
),
# Controls garbage collection during CUDA graph capture.
# If set to 0 (default), enables GC freezing to speed up capture time.