feat: support broadcast with OpenMPI backend implementation

[GordonYang1]

Summary

This PR adds two new collective communication operators: infiniBroadcast and infiniBcast() along with its OpenMPI backend implementation, plus the corresponding example program examples/broadcast.cc.

Public API

infiniResult_t infiniBroadcast(const void *sendbuff, void *recvbuff,
                               size_t count, infiniDataType_t datatype,
                               int root, infiniComm_t comm, void *stream);

infiniResult_t infiniBcast(void *buff, size_t count, infiniDataType_t datatype,
                           int root, infiniComm_t comm, void *stream);

Parameter order matches ncclBroadcast exactly, per the C++ rule in CONTRIBUTING.md ("For functions that correspond to NCCL, parameter order must match the NCCL interface exactly").

Changes

• Broadcast Operator Support

  • add the Broadcast base class in src/base/broadcast.h with
    parameter validation (root range, null pointer, count = 0
    short-circuit);
  • add the OpenMPI backend implementation in src/ompi/impl/broadcast.h,
    using a host-staging path (consistent with the existing AllReduce
    implementation);
  • chunk the buffer by INT_MAX bytes and call MPI_Bcast repeatedly,
    so that arbitrarily large count values are supported;
  • support both in-place (sendbuff == recvbuff) and out-of-place
    calling conventions;
  • in out-of-place mode, allow non-root ranks to pass nullptr as
    sendbuff (this convention is exercised by the test suite).
  • add Bcast which corresponding to the deprecated API in NCCL, but here it is simply an in-place alias for Broadcast.

• Broadcast Example

  • add examples/broadcast.cc covering 3 scenarios:
    • out-of-place infiniBroadcast()
    • in-place infiniBroadcast()
    • legacy in-place infiniBcast()

Known Issues & Future Work

• The OMPI backend uses the host-staging path. This PR does not cover GPU-direct or NCCL paths; an NCCL-based Broadcast backend can be added in a follow-up PR.
• The INT_MAX-byte chunking pattern currently only lives in Broadcast. AllReduce could benefit from the same treatment in the future.
• Broadcast::Execute's argument-validation style is kept consistent with AllReduce::Execute — invalid input returns kInvalidArgument and emits a LOG line. The "invalid root" test case intentionally triggers these log lines, which is expected output, not a failure.

Test Environment & Results

all_gather.log
all_reduce.log
reduce_scatter.log
broadcast.log

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1a129ea4e4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Prevent root-only early exit before collective

In the root branch, CHECK_STATUS can return immediately when the D2H copy or stream sync fails, but non-root ranks still proceed to MPI_Bcast; this mismatched collective participation can hang the whole communicator. This shows up whenever root hits a runtime error (e.g., bad device pointer or prior sticky CUDA error) while other ranks are healthy.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Free staging buffer on error returns

If the final host-to-device copy fails, CHECK_STATUS returns before host_buf is freed, leaking total_bytes on each failed call. With large broadcasts this can consume significant host memory quickly under repeated error conditions.

Useful? React with 👍 / 👎.