refactor: remove "auto" option from fused_lm_head_token_chunk_size

[samsja]

Summary

Removes the "auto" option from model.fused_lm_head_token_chunk_size. This value previously resolved to 8192 for RL training via the auto_setup_fused_lm_head_token_chunk_size config validator. It was redundant — users can simply set the integer explicitly.

Changes

• packages/prime-rl-configs/src/prime_rl/configs/trainer.py:
  • Type changed from int | Literal["auto", "disabled"] → int | Literal["disabled"]
  • Removed auto_setup_fused_lm_head_token_chunk_size validator (was a no-op for non-"auto" values)
  • Updated docstring to remove "auto" mention
• docs/scaling.md: Removed the commented # fused_lm_head_token_chunk_size = "auto" example line
• CHANGELOG.md: Added breaking-change entry

Migration

Users who relied on "auto" should set fused_lm_head_token_chunk_size = 8192 explicitly:

[trainer.model]
fused_lm_head_token_chunk_size = 8192  # was: "auto"

Notes

• The default remains 1024 (unchanged)
• "disabled" (vanilla LM head) and explicit integers continue to work as before
• SFT training still silently disables this (sets to "disabled")
• No existing config files in the repo used "auto" for this field
• Follow-up to feat: default ep to auto for MoE models #2868

---

Note

Low Risk
Small, localized trainer config validation and documentation change with a clear migration path; no runtime training logic beyond dropping the auto-resolve step.

Overview
Breaking config change: model.fused_lm_head_token_chunk_size no longer accepts "auto". The field is now int | Literal["disabled"] only; the default stays 1024.

The auto_setup_fused_lm_head_token_chunk_size validator that rewrote "auto" → 8192 for RL is removed. Anyone who relied on "auto" should set fused_lm_head_token_chunk_size = 8192 explicitly. "disabled" and explicit integers behave as before; SFT still forces this off.

Docs (scaling.md) drop the commented "auto" example; CHANGELOG documents the migration.

^{Reviewed by Cursor Bugbot for commit b83e6ab. Bugbot is set up for automated code reviews on this repo. Configure here.}