[Model] Add support for openai/privacy-filter (#41026)

Signed-off-by: Fabian Joswig <fjosw@users.noreply.github.com>
Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
This commit is contained in:
Fabian Joswig
2026-06-28 16:18:22 -07:00
committed by GitHub
co-authored by wang.yuqi Tyler Michael Smith
parent 95528527ea
commit 4dfbf1503b
8 changed files with 225 additions and 5 deletions
@@ -11,6 +11,7 @@ import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm._aiter_ops import rocm_aiter_ops
from vllm.config.kernel import MoEBackend
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
from vllm.model_executor.layers.fused_moe.all2all_utils import (
maybe_make_prepare_finalize,
)
@@ -83,6 +84,14 @@ def _get_priority_backends(moe_config: FusedMoEConfig) -> list[UnquantizedMoeBac
if moe_config.moe_parallel_config.dp_size > 1:
_move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_CUTLASS)
# HACK: unquantized FlashInfer aliases SWIGLUOAI to plain Swiglu
# (swiglu_alpha/limit only set on the MXFP4 branch). Route to
# Triton's swigluoai_and_mul until that's plumbed through. Same
# demotion pattern as the Qwen3.5/dp_size hack above.
if moe_config.activation == MoEActivation.SWIGLUOAI:
_move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_TRTLLM)
_move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_CUTLASS)
elif current_platform.is_xpu():
_AVAILABLE_BACKENDS = [UnquantizedMoeBackend.XPU]
elif current_platform.is_cpu():