A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 20:00 UTC
[moe] Allow disabling DP chunking (#29936)
Signed-off-by: Ming Yang <minos.future@gmail.com>
Ming Yang committed
Dec 8, 2025 at 16:29 UTC
9d6235ca9a36e76911045999ed72e3c8aad66b8a
2 files changed
+5
-1
vllm/envs.py
+4
@@ -144,6 +144,7 @@ if TYPE_CHECKING:
144
VLLM_DP_MASTER_IP: str = ""
145
VLLM_DP_MASTER_PORT: int = 0
146
VLLM_MOE_DP_CHUNK_SIZE: int = 256
147
+ VLLM_ENABLE_MOE_DP_CHUNK: bool = True
148
VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
149
VLLM_RAY_DP_PACK_STRATEGY: Literal["strict", "fill", "span"] = "strict"
150
VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
@@ -1101,6 +1102,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
1102
# rank. All DP ranks process the activations in VLLM_MOE_DP_CHUNK_SIZE
1103
# units.
1104
"VLLM_MOE_DP_CHUNK_SIZE": lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),
1105
+ "VLLM_ENABLE_MOE_DP_CHUNK": lambda: bool(
1106
+ int(os.getenv("VLLM_ENABLE_MOE_DP_CHUNK", "1"))
1107
+ ),
1108
# Randomize inputs during dummy runs when using Data Parallel
1109
"VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: os.environ.get(
1110
"VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
vllm/model_executor/layers/fused_moe/layer.py
+1
-1
@@ -753,7 +753,7 @@ class FusedMoE(CustomOp):
753
self.moe_parallel_config.use_pplx_kernels
754
or self.moe_parallel_config.use_deepep_ll_kernels
755
or (self.dp_size > 1 and self.use_flashinfer_cutlass_kernels)
756
- )
756
+ ) and envs.VLLM_ENABLE_MOE_DP_CHUNK
757
758
@property
759
def is_internal_router(self) -> bool: