diff --git a/vllm/compilation/decorators.py b/vllm/compilation/decorators.py
index cb629ea284fa1..1370862d580a5 100644
--- a/vllm/compilation/decorators.py
+++ b/vllm/compilation/decorators.py
@@ -189,7 +189,6 @@ def _support_torch_compile(
             CompilationLevel.NO_COMPILATION, CompilationLevel.DYNAMO_AS_IS
         ] or not supports_dynamo() or _should_ignore_torch_compile(
             self.__class__)
-        self.do_not_compile = True
         if self.do_not_compile:
             return
 
diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index 93b40c6241b73..3e2947b6d2551 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -1918,11 +1918,12 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                                        scheduler_output, is_dummy_run)
             # if is_global_first_rank():
             #     logger.info(f"RUNNING FULL BATCH {num_scheduled_tokens}")
+            skip_cuda_graphs = self.parallel_config.enable_microbatching
             with set_forward_context(attn_metadata,
                                      vllm_config=self.vllm_config,
                                      num_tokens=num_scheduled_tokens or 1,
                                      num_tokens_across_dp=num_tokens_across_dp,
-                                     skip_cuda_graphs=True):
+                                     skip_cuda_graphs=skip_cuda_graphs):
                 return self.model(
                     input_ids=input_ids,
                     positions=positions,