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 3, 06:45 UTC
[MRV2][BugFix] Fix default-stream CG capture in P/W LoRA case (#43160)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
May 19, 2026 at 19:19 UTC
39bba710bed5b6018718af3e0fd7984f6082118e
1 file changed
+3
-1
vllm/v1/worker/gpu/cudagraph_utils.py
+3
-1
@@ -358,7 +358,9 @@ class ModelCudaGraphManager(CudaGraphManager):
358
batch_descriptor = None
359
if cg_mode == CUDAGraphMode.PIECEWISE:
360
assert attn_metadata is None
361
- batch_descriptor = BatchDescriptor(num_tokens=num_tokens)
361
+ batch_descriptor = BatchDescriptor(
362
+ num_tokens=num_tokens, has_lora=has_lora
363
+ )
364
with set_forward_context(
365
attn_metadata,
366
self.vllm_config,