Fix data parallel + tensor parallel (#4499)

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -16,6 +16,7 @@
 from __future__ import annotations
 
 import bisect
+import os
 from contextlib import contextmanager
 from typing import TYPE_CHECKING, Callable
 
@@ -81,7 +82,9 @@ def patch_model(
             # tp_group.ca_comm = None
             yield torch.compile(
                 torch.no_grad()(model.forward),
-                mode="max-autotune-no-cudagraphs",
+                mode=os.environ.get(
+                    "SGLANG_TORCH_COMPILE_MODE", "max-autotune-no-cudagraphs"
+                ),
                 dynamic=False,
             )
         else: