Yuge Zhang comments

Results 279 comments of


                                            Yuge Zhang

The server training ran out of memory (OOM) and crashed.

I didn't see the OOM issue from this log. Do you have more contexts to help us investigate?

ValueError: `position_ids should be a 3D tensor of shape (4, batch_size, seq_length)` when using agent-lightning 0.2.1 + Qwen2.5-VL

Qwen2.5VL is known to be problematic with current VERL integration. The major problem is we didn't upgrade our VERL integration after VERL has upgraded themselves. Related to #105

Diff components

Looks reasonable. Would you feel that `` is enough for you? Getting children components in react has been always tricky and it's unclear whether the diff should be based on...

AttributeError: module 'agentlightning' has no attribute 'setup_logging'

If you are running with v0.2.2, you should check out examples from v0.2.2: https://github.com/microsoft/agent-lightning/tree/v0.2.2/examples

Bump pdfjs-dist from 3.11.174 to 4.2.67

This is a breaking change that requires much effort.

How to turn off the thinking mode

So you want to remove the thinking tokens from the response for training? I think that will cause discrepency between training and inference.

How to turn off the thinking mode

Has verl figured that out? If they haven't, we are unable to help despite we want to.

Completed rollouts > total_tasks_queued

I think this is a known issue. Remember what rollout IDs have been sent and reject those unregistered rollouts might be a workaround. In the long run, we need to...

Skip vllm 0.12.0

/ci

Increase graceful timeout on CI

/ci