Fix #458

[Binyang2014]

Fix mscclpp_benchmark allreduce test

[Copilot]

Pull Request Overview

This PR fixes the allreduce test in the mscclpp_benchmark by updating parameter types, changing the reduction store to a broadcast, and adding compile-time branches for vector sizes.

• Changed my_rank, num_ranks, and thread/index variables from size_t to int
• Replaced switchChan.store with switchChan.broadcast in the reduction loop
• Introduced if constexpr branches for different vector_size values and added runtime/assert guards for unsupported sizes

Comments suppressed due to low confidence (1)

python/mscclpp_benchmark/allreduce.cu:838

• New branching logic for different vector_size values lacks corresponding tests. Please add tests for all supported vector_size paths (1, 2, 4, and 8) to ensure correct behavior.

    if (vector_size == 4) {

[Binyang2014]

/azp run

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).