[core][torch.compile] discard the compile for profiling (#7796)

.buildkite/run-tpu-test.sh

@@ -12,5 +12,4 @@ remove_docker_container
 # For HF_TOKEN.
 source /etc/environment
 # Run a simple end-to-end example.
-docker run --privileged --net host --shm-size=16G -it -e HF_TOKEN=$HF_TOKEN --name tpu-test vllm-tpu \
+docker run --privileged --net host --shm-size=16G -it -e HF_TOKEN=$HF_TOKEN --name tpu-test vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git && python3 /workspace/vllm/tests/tpu/test_compilation.py && python3 /workspace/vllm/examples/offline_inference_tpu.py"
-    python3 /workspace/vllm/examples/offline_inference_tpu.py

tests/tpu/test_compilation.py

@@ -0,0 +1,34 @@
+import glob
+import os
+import runpy
+import tempfile
+
+import depyf
+
+temp_dir = tempfile.mkdtemp()
+with depyf.prepare_debug(temp_dir):
+    cur_dir = os.path.dirname(__file__)
+    parent_dir = os.path.dirname(cur_dir)
+    root_dir = os.path.dirname(parent_dir)
+    example_file = os.path.join(root_dir, "examples",
+                                "offline_inference_tpu.py")
+    runpy.run_path(example_file)
+
+compiled_code = sorted(
+    glob.glob(os.path.join(temp_dir, "__transformed_code*.py")))
+full_code = glob.glob(os.path.join(temp_dir, "full_code*.py"))[0]
+# we should only trigger Dynamo compilation three times:
+# one for the profiling phase (and the compiled artifact will be discarded)
+# one for the prefill phase with symbolic shapes
+# one for the decode phase with symbolic shapes
+# and later calls should not trigger Dynamo compilation again.
+# NOTE: it might still trigger XLA compilation.
+
+# check we have three compiled code
+assert len(compiled_code) == 3
+
+# check the first compilation is discarded
+with open(full_code) as f:
+    full_code_content = f.read()
+    profile_function = compiled_code[0].split(".")[0]
+    assert profile_function not in full_code_content

vllm/worker/model_runner.py

@@ -1097,6 +1097,10 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                 device=self.device)
         self.execute_model(model_input, kv_caches, intermediate_tensors)
         torch.cuda.synchronize()
+
+        # reset and discard the guard and compiled bytecode for profiling runs
+        torch._dynamo.reset()
+
         return
 
     def remove_all_loras(self):

vllm/worker/tpu_worker.py

@@ -143,6 +143,10 @@ class TPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
         num_cpu_blocks = int(self.cache_config.swap_space_bytes //
                              block_size_bytes)
         num_cpu_blocks = (num_cpu_blocks // 8) * 8  # Round down to 8.
+
+        # reset and discard the guard and compiled bytecode for profiling runs
+        torch._dynamo.reset()
+
         return num_tpu_blocks, num_cpu_blocks
 
     def initialize_cache(