@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 7b98a65ae6
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 16, 02:00 UTC

[torch.compile] disable logging when cache is disabled (#12043)

Signed-off-by: youkaichao <youkaichao@gmail.com>

youkaichao committed Jan 18, 2025 at 04:29 UTC 7b98a65ae6f011fb31fefa1f563b7d6e554df434
1 file changed +21 -9
vllm/compilation/backends.py
+21 -9
@@ -251,15 +251,27 @@ def wrap_inductor(graph: fx.GraphModule,
251 def _get_shape_env() -> AlwaysHitShapeEnv:
252 return AlwaysHitShapeEnv()
253
254 - with patch(# for hijacking the hash of the compiled graph
255 - "torch._inductor.codecache.compiled_fx_graph_hash",
256 - hijack_compiled_fx_graph_hash), \
257 - patch(# for providing a dummy shape environment
258 - "torch._inductor.codecache.FxGraphCache._get_shape_env",
259 - _get_shape_env), \
260 - patch(# for forcing the graph to be cached
261 - "torch._inductor.codecache.FxGraphCache._check_can_cache",
262 - _check_can_cache):
254 + with ExitStack() as stack:
255 + if not cache_data.disabled:
256 + # compilation cache is enabled, patch several functions
257 +
258 + # for hijacking the hash of the compiled graph
259 + stack.enter_context(
260 + patch("torch._inductor.codecache.compiled_fx_graph_hash",
261 + hijack_compiled_fx_graph_hash))
262 +
263 + # for providing a dummy shape environment
264 + stack.enter_context(
265 + patch(
266 + "torch._inductor.codecache.FxGraphCache._get_shape_env",
267 + _get_shape_env))
268 +
269 + # for forcing the graph to be cached
270 + stack.enter_context(
271 + patch(
272 + "torch._inductor.codecache.FxGraphCache._check_can_cache",
273 + _check_can_cache))
274 +
275 compiled_graph = compile_fx(graph,
276 example_inputs,
277 config_patches=current_config)