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 14, 05:30 UTC
Fix nits from #20059 (#23548)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Aug 26, 2025 at 11:27 UTC
b00e69f8ca55f4a82847d39466f57ceb748324c1
1 file changed
+6
-8
vllm/config/compilation.py
+6
-8
@@ -225,7 +225,8 @@ class CompilationConfig:
225
# CudaGraph compilation
226
cudagraph_mode: Optional[CUDAGraphMode] = None
227
"""
228
- The mode of the cudagraph.
228
+ The mode of the cudagraph:
229
+
230
- NONE, no cudagraph capture.
231
- PIECEWISE. (v1 default)
232
- FULL.
@@ -384,13 +385,10 @@ class CompilationConfig:
385
if pass_config_exclude:
386
exclude["pass_config"] = pass_config_exclude
387
387
- # The cast to string is necessary because Pydantic is mocked in docs
388
- # builds and sphinx-argparse doesn't know the return type of decode()
389
- return str(
390
- TypeAdapter(CompilationConfig).dump_json(
391
- self,
392
- exclude=exclude, # type: ignore[arg-type]
393
- exclude_unset=True).decode())
388
+ return TypeAdapter(CompilationConfig).dump_json(
389
+ self,
390
+ exclude=exclude, # type: ignore[arg-type]
391
+ exclude_unset=True).decode()
392
393
__str__ = __repr__
394