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 15, 07:15 UTC
Improve "failed to get the hash of the compiled graph" error (#18956)
Signed-off-by: rzou <zou3519@gmail.com>
Richard Zou committed
May 30, 2025 at 11:00 UTC
84ec470fca4f3c1873863086a420a40206f1bf01
1 file changed
+8
-2
vllm/compilation/compiler_interface.py
+8
-2
@@ -415,8 +415,14 @@ class InductorAdaptor(CompilerInterface):
415
# compilation cache. So turn off the checks if we disable the
416
# compilation cache.
417
if not envs.VLLM_DISABLE_COMPILE_CACHE:
418
- assert hash_str is not None, (
419
- "failed to get the hash of the compiled graph")
418
+ if hash_str is None:
419
+ raise RuntimeError(
420
+ "vLLM failed to compile the model. The most "
421
+ "likely reason for this is that a previous compilation "
422
+ "failed, leading to a corrupted compilation artifact. "
423
+ "We recommend trying to "
424
+ "remove ~/.cache/vllm/torch_compile_cache and try again "
425
+ "to see the real issue. ")
426
assert file_path is not None, (
427
"failed to get the file path of the compiled graph")
428
return compiled_graph, (hash_str, file_path)