biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Feature]: Remove Chunking From FusedMoE (#34086)

Browse Source 
Signed-off-by: SouthWest7 <am1ao@qq.com>
Signed-off-by: Southwest <1403572259@qq.com>
Signed-off-by: southwest <am1ao@qq.com>
Signed-off-by: Xinan Miao <1403572259@qq.com>
Co-authored-by: SouthWest7 <am1ao@qq.com>
This commit is contained in:
Xinan Miao
2026-03-13 02:24:38 +08:00
committed by GitHub
parent c973ecdead
commit 2cdf92228c
28 changed files with 152 additions and 523 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/envs.py
View File

@@ -53,8 +53,6 @@ if TYPE_CHECKING:
VLLM_CPU_SGL_KERNEL: bool = False
VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
VLLM_XLA_CHECK_RECOMPILATION: bool = False
VLLM_FUSED_MOE_CHUNK_SIZE: int = 16 * 1024
VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
@@ -822,15 +820,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
),
# Enable SPMD mode for TPU backend.
"VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
"VLLM_FUSED_MOE_CHUNK_SIZE": lambda: int(
os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(16 * 1024))
),
# Control whether to use fused MoE activation chunking. Current chunking
# logic is incompatible with torch.compile and causes IMA. See issue
# https://github.com/vllm-project/vllm/issues/19631.
"VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING": lambda: bool(
int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))
),
# If set, the OpenAI API server will stay alive even after the underlying
# AsyncLLMEngine errors and stops serving requests
"VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(