Field note · vLLM
cache_kernels.cu fails to build on aarch64 with CUDA 13.0
A KV-cache batching optimization calls cuMemcpyBatchAsync with a signature the CUDA 13.0 aarch64 driver API does not have. The build stops before anything runs.
- Hardware
- NVIDIA DGX Spark · aarch64
- CUDA
- 13.0.1
- Software
- Docker base nvidia/cuda:13.0.1-devel-ubuntu22.04
/workspace/csrc/cache_kernels.cu(115): error: argument of type "size_t *" is incompatible with parameter of type "CUstream" (aka "CUstream_st *")Symptom#
Building vLLM from source for aarch64 against CUDA 13.0.1 — the nvidia/cuda:13.0.1-devel-ubuntu22.04 base image — fails during kernel compilation. The build never finishes; nothing runs yet. A second error follows immediately after the one above, same file, next line:
/workspace/csrc/cache_kernels.cu(116): error: too many arguments in function callRoot cause#
Commit 4a06e1246, "[Perf] Batch KV cache swap copies via cuMemcpyBatchAsync," calls cuMemcpyBatchAsync with a function signature that doesn't match the CUDA 13.0 driver API on aarch64 — the parameter types are wrong, passing a size_t * where the API expects a CUstream. The likely explanation is a CUDA API version mismatch: cuMemcpyBatchAsync's signature under CUDA 13.0 on aarch64 may simply differ from whatever version this code was written and tested against. That's not confirmed — just the working theory.
Fix#
Revert the commit:
git revert --no-edit 4a06e1246The reverted commit is a performance optimization — batching KV cache swap copies — not a feature the server needs in order to run, so the cost of reverting is low.
This looks aarch64-specific: DGX Spark, and plausibly other aarch64 platforms like Jetson. x86_64 builds may not be affected at all, though that isn't confirmed here either.
