Skip to content

Non-leader worker initialization failure can wedge MPI ranks #16342

Description

@Stelath

Description

When worker_main fails while constructing GenerationExecutorWorker on a non-leader rank, that rank returns while rank 0 can remain blocked in an MPI collective that the failed rank can no longer enter. The executor process then remains wedged instead of failing promptly, so the serving worker may require manual termination.

Proposed fix

Abort the MPI communicator from a non-leader rank when worker initialization raises. Rank 0 retains the existing behavior of reporting the initialization exception through worker_init_status_queue.

Validation

  • Focused unit regression covering the non-leader failure path.
  • Real 4-rank MPI integration test with rank 3 constructor failure: rank 3 invoked MPI_Abort, and the MPI job exited with code 1 before a 20-second timeout.

This was originally observed with disaggregated TensorRT-LLM serving, where a rank initialization failure otherwise left the node-level worker wedged.

Metadata

Metadata

Assignees

No one assigned

    Labels

    Disaggregated serving<NV>Deploying with separated, distributed components (params, kv-cache, compute). Arch & perf.

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions