Summary
gpt_oss.chat broadcasts user input across distributed ranks, but tool/function execution is not rank-gated. When Torch or Triton chat is launched with torchrun, every rank enters the same browser, Python, or apply_patch branch and executes the tool independently.
Impact
A single model tool call can therefore produce N external side effects on N model-parallel ranks:
- browser searches/API calls are duplicated;
- Python tool work is duplicated;
apply_patch can race multiple processes against the same working tree;- non-deterministic tool results can leave ranks with different conversation histories before the next model-parallel generation step.
get_user_input() already establishes the intended rank-0/broadcast pattern for interactive input; tool results need the same treatment.
Proposed resolution
Execute each tool/function call only on rank 0, serialize Harmony Message results with Message.to_dict(), broadcast the result payload, and reconstruct it with Message.from_dict() on every rank. Propagate rank-0 execution failures to every rank as well so nonzero ranks cannot block forever waiting for a result.
Add mocked distributed regression coverage proving nonzero ranks do not execute the tool callback and receive the rank-0 result.
Summary
gpt_oss.chatbroadcasts user input across distributed ranks, but tool/function execution is not rank-gated. When Torch or Triton chat is launched withtorchrun, every rank enters the same browser, Python, orapply_patchbranch and executes the tool independently.Impact
A single model tool call can therefore produce N external side effects on N model-parallel ranks:
apply_patchcan race multiple processes against the same working tree;get_user_input()already establishes the intended rank-0/broadcast pattern for interactive input; tool results need the same treatment.Proposed resolution
Execute each tool/function call only on rank 0, serialize Harmony
Messageresults withMessage.to_dict(), broadcast the result payload, and reconstruct it withMessage.from_dict()on every rank. Propagate rank-0 execution failures to every rank as well so nonzero ranks cannot block forever waiting for a result.Add mocked distributed regression coverage proving nonzero ranks do not execute the tool callback and receive the rank-0 result.