[Model Runner V2] Minor fix for cudagraph_utils (#29256)

vllm/v1/worker/gpu/cudagraph_utils.py

@@ -1,7 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import gc
-from contextlib import contextmanager
+from unittest.mock import patch
 
 import numpy as np
 import torch
@@ -140,6 +139,7 @@ class CudaGraphManager:
             attn_metadata,
             self.vllm_config,
             num_tokens=batch_size,
+            cudagraph_runtime_mode=CUDAGraphMode.NONE,
             num_tokens_across_dp=num_tokens_across_dp,
         ):
             hidden_states = model(
@@ -148,15 +148,16 @@ class CudaGraphManager:
             )
             if self.hidden_states is None:
                 self.hidden_states = torch.empty_like(hidden_states)
-        torch.cuda.synchronize()
 
         # Capture the graph.
         graph = torch.cuda.CUDAGraph()
         with (
+            patch("torch.cuda.empty_cache", lambda: None),
             set_forward_context(
                 attn_metadata,
                 self.vllm_config,
                 num_tokens=batch_size,
+                cudagraph_runtime_mode=CUDAGraphMode.NONE,
                 num_tokens_across_dp=num_tokens_across_dp,
             ),
             torch.cuda.graph(graph, self.pool),
@@ -183,7 +184,7 @@ class CudaGraphManager:
         if is_global_first_rank():
             sizes_to_capture = tqdm(sizes_to_capture, desc="Capturing CUDA graphs")
 
-        with freeze_gc(), graph_capture(device=self.device):
+        with graph_capture(device=self.device):
             for batch_size in sizes_to_capture:
                 self.capture_graph(
                     batch_size,
@@ -199,13 +200,3 @@ class CudaGraphManager:
         self.graphs[batch_size].replay()
         assert self.hidden_states is not None
         return self.hidden_states[:batch_size]
-
-
-@contextmanager
-def freeze_gc():
-    gc.collect()
-    gc.freeze()
-    try:
-        yield
-    finally:
-        gc.unfreeze()

vllm/v1/worker/gpu/model_runner.py

@@ -298,6 +298,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
             return 0
 
         start_time = time.perf_counter()
+        torch.cuda.empty_cache()
         start_free_gpu_memory = torch.cuda.mem_get_info()[0]
 
         with self.maybe_setup_dummy_loras(self.lora_config):