All-Reduce
all_reduce validates a two-rank GPU sum-and-broadcast collective. Pantheon
initializes GPU 0 with 1 and its peer GPU with 2, reduces the values to 3,
then confirms that both ranks receive the same result.
Transport behavior
When the selected GPU pair exposes bidirectional peer access, Pantheon uses direct peer DMA. If peer access is unavailable, it uses a host-staged fallback and identifies that mode in the result output. The fallback is useful for correctness and topology diagnosis, but its throughput is not a direct P2P bandwidth result.
Verification
--verify checks the complete output buffer on both GPUs. For a verification
self-test, add --inject_error; Pantheon corrupts one value on the peer rank,
reports Verification: FAIL, and exits non-zero.
Requirements
- At least two visible GPUs.
- The chosen GPU and its next visible peer GPU must be accessible to the runtime.
- CUDA or ROCm/HIP runtime support. The workload uses portable HIP APIs.