[MoE] Remove inplace fused experts mechanism (#43727)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Yongye Zhu
2026-05-27 20:00:19 -07:00
committed by GitHub
parent 05eec7120e
commit 2d2c660104
44 changed files with 11 additions and 237 deletions
@@ -1,6 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import functools
from math import prod
import torch
@@ -32,7 +31,6 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils.math_utils import cdiv
from vllm.utils.torch_utils import is_torch_equal_or_newer
@triton.jit
@@ -368,14 +366,6 @@ def normalize_batched_scales_shape(
return scales
# Torch custom ops can't deal with outputs aliasing inputs so we need to
# disable inplace for torch >= 2.9.
# See https://github.com/vllm-project/vllm/issues/26378
@functools.cache
def disable_inplace() -> bool:
return is_torch_equal_or_newer("2.9")
@triton.jit
def _pack_topk_ids_weights_kernel(
topk_ids_ptr,