Rename env

Signed-off-by: ilmarkov <imarkov@redhat.com>

Author: ilmarkov

vllm/distributed/device_communicators/cuda_communicator.py

@@ -72,7 +72,7 @@ def __init__(self,
                 # currently be an MI300 series.
                 self.qr_comm = QuickAllReduce(group=self.cpu_group,
                                               device=self.device)
-        if envs.VLLM_USE_SYMM_MEM and current_platform.is_cuda():
+        if envs.VLLM_ALLREDUCE_USE_SYMM_MEM and current_platform.is_cuda():
             self.symm_mem_comm = SymmMemCommunicator(
                 group=self.cpu_group,
                 device=self.device,

vllm/distributed/device_communicators/custom_all_reduce.py

@@ -117,7 +117,7 @@ def __init__(self,
         # now `device` is a `torch.device` object
         assert isinstance(device, torch.device)
         self.device = device
-        if current_platform.is_cuda() and envs.VLLM_USE_SYMM_MEM:
+        if current_platform.is_cuda() and envs.VLLM_ALLREDUCE_USE_SYMM_MEM:
             max_size = CustomAllreduce._MAX_SIZES[world_size]
 
         cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES

vllm/envs.py

@@ -139,7 +139,7 @@
     VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
     VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: Optional[int] = None
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 120
-    VLLM_USE_SYMM_MEM: bool = False
+    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = False
 
 
 def get_default_cache_root():
@@ -965,8 +965,8 @@ def get_vllm_port() -> Optional[int]:
     lambda: int(os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "120")),
 
     # Whether to use pytorch symmetric memory for allreduce
-    "VLLM_USE_SYMM_MEM":
-    lambda: bool(int(os.getenv("VLLM_USE_SYMM_MEM", "0"))),
+    "VLLM_ALLREDUCE_USE_SYMM_MEM":
+    lambda: bool(int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "0"))),
 }
 
 # --8<-- [end:env-vars-definition]