[BugFix] Fix memory spike in workspace allocation (#30744)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

.buildkite/test-pipeline.yaml

@@ -1223,6 +1223,8 @@ steps:
     # FIXIT: find out which code initialize cuda before running the test
     # before the fix, we need to use spawn to test it
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    # Alot of these tests are on the edge of OOMing
+    - export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
     # There is some Tensor Parallelism related processing logic in LoRA that
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py

vllm/v1/worker/workspace.py

@@ -145,12 +145,20 @@ class WorkspaceManager:
 
             for ubatch_id in range(self._num_ubatches):
                 current_workspace = self._current_workspaces[ubatch_id]
-                if current_workspace is None:
+                if (
+                    current_workspace is None
+                    or self._workspace_size_bytes(current_workspace) < required_bytes
+                ):
+                    # Delete old tensor before allocating new one to avoid
+                    # memory spike from resize_(). resize_() allocates new
+                    # memory before freeing old, which can cause OOM.
+                    # Must clear the list reference first since local var
+                    # is just a copy of the reference.
+                    self._current_workspaces[ubatch_id] = None
+                    del current_workspace
                     self._current_workspaces[ubatch_id] = torch.empty(
                         (required_bytes,), dtype=torch.uint8, device=self._device
                     )
-                elif self._workspace_size_bytes(current_workspace) < required_bytes:
-                    current_workspace.resize_(required_bytes)
 
             if envs.VLLM_DEBUG_WORKSPACE:
                 logger.info(