Expose optimizer params

[RobotSail]

Users should be able to pass settings to the AdamW optimizer.

[RobotSail]

@claude Is this still needed?

[RobotSail]

@claude-code do you think we need to do this?

[claude]

Claude finished @RobotSail's task in 59s —— View job

---

Analyzing whether optimizer params need to be exposed

• Investigate current optimizer configuration in the codebase
• Check if optimizer params are already configurable
• Provide recommendation

Analysis

This issue is largely already resolved. The AdamW optimizer parameters (beta1, beta2, eps, weight_decay) have been added to the codebase since this issue was opened. Here's the current state:

Layer	beta1	beta2	eps	weight_decay	Status
TrainingArgs dataclass	training_types.py:85-97				Exposed
CLI (main())	train.py:1021-1024				Exposed
setup_training_components()	setup_model_for_training.py:1236-1239				Plumbed through
api_train.py (run_training)	--	--	--	--	Not passed to CLI

The one remaining gap: api_train.py:run_training() builds the torchrun command from TrainingArgs but does not forward beta1, beta2, eps, or weight_decay to the CLI. So users of the programmatic API (run_training()) silently get the defaults regardless of what they set in TrainingArgs.

Recommendation

This issue can be closed once that api_train.py gap is fixed. It's a small change — just appending four more flags to the command list (similar to how --learning-rate is already passed). If you'd like, I can fix it right now.