Fix the program hang caused by duplicate serialization#1949
Conversation
|
@claude approve |
There was a problem hiding this comment.
LGTM. The fix correctly deduplicates Ray object serialization by using explicit ray.put() with a data_batch_refs cache keyed by dp_idx. Previously, each worker in the same data-parallel group would implicitly serialize the same data slice via .remote(), causing redundant serialization that could lead to hangs. The cleanup with del data_batch_refs in the finally block is also appropriate.
|
Claude: Approved PR #1949 as requested. What was requested: Approve the PR. Actions taken: Reviewed the diff and approved with comments. The fix correctly addresses a program hang caused by duplicate Ray object serialization — workers in the same data-parallel group now share a single |
No description provided.