chore: strip [default0] prefix from trainer logs

[mikasenghaas]

Summary

• Strip the [default0]: prefix that torchrun's --tee flag prepends to every log line (note, that these can still be seen for in the per-rank logs)
• Uses sed -u in the tail/tee pipelines to remove the prefix in real-time
• Also sets --role=trainer for clearer per-rank log files in the torchrun log directory
• Applied to both local launchers (rl.py, sft.py) and SLURM templates

Before:

[default0]:00:19:20    INFO Starting training loop (max_steps=100)
[default0]:00:20:21 SUCCESS Step 0 | Time: 60.98s | Loss: 0.0000 | ...

After:

00:19:20    INFO Starting training loop (max_steps=100)
00:20:21 SUCCESS Step 0 | Time: 60.98s | Loss: 0.0000 | ...

🤖 Generated with Claude Code

---

Note

Low Risk
Logging-only changes that adjust torchrun args and shell pipelines; main risk is portability/quoting issues when spawning shell=True or running on systems without compatible sed.

Overview
Trainer launch commands now pass torchrun --role=trainer so per-rank log files are labeled more clearly.

For both local (rl.py, sft.py) and multi-node SLURM templates, the displayed/tee’d trainer log stream is piped through sed -u to strip the leading [...] rank prefix (e.g., [default0]:) while keeping the underlying per-rank torchrun logs intact.

^{Written by Cursor Bugbot for commit 0332018. This will update automatically on new commits. Configure here.}

[samsja]

nah we need this for when we have multi rank trainer writting to the same file. Also I am against doing too many magic beyond tail. The file is the file