@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 53cd7f868b

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 11, 07:30 UTC

[compile] Recompile graph module during Dynamo cache loading. (#30743)

Signed-off-by: Zhengxu Chen <zhxchen17@fb.com>

Zhengxu Chen committed Dec 17, 2025 at 05:00 UTC 53cd7f868b3632cbbe982cffaee8e16fb49dd694

1 file changed +1

vllm/compilation/caching.py

@@ -104,6 +104,7 @@ class VllmSerializableFunction(SerializableCallable):

104 state = pickle.loads(data)

105 fake_mode = FakeTensorMode(shape_env=ShapeEnv())

106 state["graph_module"] = GraphPickler.loads(state["graph_module"], fake_mode)

107 + state["graph_module"].recompile()

108 state["example_inputs"] = GraphPickler.loads(state["example_inputs"], fake_mode)

109 vllm_backend = VllmBackend(get_current_vllm_config(), state["prefix"])

110