Set is_distributed false by default in vllm (#266)

huggingface · Sep 20, 2024 · 99a90c9 · 99a90c9
1 parent 1992de3
commit 99a90c9
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/optimum_benchmark/trackers/latency.py b/optimum_benchmark/trackers/latency.py
@@ -121,7 +121,8 @@ def __init__(self, device: str, backend: str):
         self.device = device
         self.backend = backend
         self.is_asynchronous = self.backend == "pytorch" and self.device == "cuda"
-        self.is_distributed = is_torch_distributed_available() and torch.distributed.is_initialized()
+        self.is_distributed = (self.backend != "vllm" and
+                               is_torch_distributed_available() and torch.distributed.is_initialized())
 
         if self.is_asynchronous:
             LOGGER.info("\t+ Tracking latency using Pytorch CUDA events")