Merge branch 'v0.9.2-dev-ds-deepep-yql' into 'v0.9.2-dev-ds'

add ALLOW_MNNV default falase, use VLLM_ALLOW_MNNVL=1

See merge request dcutoolkit/deeplearing/vllm!286

vllm/distributed/device_communicators/all2all.py

@@ -8,7 +8,7 @@ import torch.distributed as dist
 from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
 from vllm.utils import has_deep_ep, has_pplx
-
+import vllm.envs as envs
 from .base_device_communicator import All2AllManagerBase, Cache
 
 logger = init_logger(__name__)
@@ -253,7 +253,9 @@ class DeepEPLLAll2AllManager(DeepEPAll2AllManagerBase):
                     num_nvl_bytes=num_nvl_bytes,
                     num_rdma_bytes=num_rdma_bytes,
                     low_latency_mode=True,
-                    num_qps_per_rank=num_qps_per_rank)
+                    num_qps_per_rank=num_qps_per_rank,
+                    allow_mnnvl=envs.VLLM_ALLOW_MNNVL,
+                    )
 
     def get_handle(self, kwargs):
         """

vllm/envs.py

@@ -128,6 +128,7 @@ if TYPE_CHECKING:
     VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
     VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
     VLLM_ALL2ALL_BACKEND: str = "naive"
+    VLLM_ALLOW_MNNVL: bool = False
     VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
     VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
     VLLM_SLEEP_WHEN_IDLE: bool = False
@@ -952,6 +953,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_ALL2ALL_BACKEND":
     lambda: os.getenv("VLLM_ALL2ALL_BACKEND", "naive"),
     
+    # use ALLOW_MNNVL
+    "VLLM_ALLOW_MNNVL":
+    lambda: (os.environ.get("VLLM_ALLOW_MNNVL", "False").lower() in
+             ("true", "1")),
+
     # Control the maximum number of tokens per expert supported by the
     # NVFP4 MoE CUTLASS Kernel. This value is used to create a buffer for
     # the blockscale tensor of activations NVFP4 Quantization.