Move LoRA out of experimental section

[willccbb]

Move LoRA config out of "experimental" section

---

Note

Moves LoRA config from model.experimental.lora to model.lora and updates trainers, validators, weight broadcast/checkpointing, and example configs accordingly.

• Config:
  • Remove ExperimentalConfig; add top-level model.lora: LoRAConfig | None in ModelConfig.
  • Update validators to reference model.lora (adapter saving, NCCL/adapter-only checks).
• Model/Trainer:
  • Apply LoRA via config.lora in setup_model before FSDP.
  • RL flow: auto_setup_lora now reads trainer.model.lora; sets adapter-only broadcast and inference LoRA settings.
  • RL/SFT train: pass config.model.lora to checkpoint/broadcast setup.
• Configs:
  • Replace [trainer.model.experimental.lora] with [trainer.model.lora] across configs/... and examples/... TOMLs.
• Docs:
  • Add CHANGELOG entry documenting the move from model.experimental.lora to model.lora.

^{Written by Cursor Bugbot for commit 9758514. This will update automatically on new commits. Configure here.}

[cursor]

Bug: Checkpoint setup uses removed LoRA config

setup_ckpt_managers is still called with config.model.experimental.lora, but ExperimentalConfig no longer defines lora after this change. This makes config.model.experimental.lora raise an AttributeError at runtime and can prevent RL/SFT training from starting when checkpoint managers are initialized.

src/prime_rl/trainer/rl/train.py#L96-L99

prime-rl/src/prime_rl/trainer/rl/train.py

Lines 96 to 99 in 34fc34a

	logger.info(f"Initializing checkpoint managers ({config.ckpt})")
	ckpt_manager, weight_ckpt_manager = setup_ckpt_managers(
	config.output_dir, config.ckpt, config.model.experimental.lora
	)

 

---