README.md

AITER dependency patches

SlimServe builds AITER from the exact upstream commit recorded by AITER_COMMIT in docker/Dockerfile.rocm_base, then applies every patch in this directory before compiling the wheel.

custom-all-reduce-graph-buffer-order.patch keeps CUDA/HIP graph input and output buffers in capture order. AITER previously stored inputs and outputs in separate growing lists. The second all-reduce input therefore reused the graph slot already assigned to the first output, which could produce an illegal peer memory access during graph replay. The patch also extends AITER's multi-GPU test to capture two consecutive all-reduces.

To reproduce the dependency outside the container build:

git -C /path/to/aiter checkout --detach \
  17f24ec6e93f48722a6b4ec8e54738434194f3d6
git -C /path/to/aiter apply --unidiff-zero \
  /path/to/SlimServe/docker/patches/aiter/custom-all-reduce-graph-buffer-order.patch

The zero-context form keeps the vendored patch free of trailing whitespace. It must continue to pass git apply --unidiff-zero --check when AITER_COMMIT is updated. The container build enforces that invariant before compilation.