[Kimi Bug] Fix k3 torch.AcceleratorError: CUDA error: an illegal memory access was encountered (#53773)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Aug 25, 2026 at 20:38 UTC
a447955acad919a6902d65f0af2d4f76c0335ed3
1 file changed
+2
vllm/v1/worker/gpu/cudagraph_utils.py
+2
@@ -845,6 +845,8 @@ def _teardown_profiling_state(runner: "GPUModelRunner") -> None:
845
"""Release the profiling KV cache and captured graphs while keeping model
846
weights, so the real ``initialize_kv_cache`` starts from a clean slate."""
847
torch.accelerator.synchronize()
848
+ if hasattr(runner.model_state, "_mamba_ctx"):
849
+ runner.model_state._mamba_ctx = None
850
if hasattr(runner, "kv_caches"):
851
runner.kv_caches.clear()
852
if hasattr(runner, "attn_groups"):