feat(tinker): Add support for built-in loss functions and checkpoint control

[corbt]

Summary

This PR adds two new features to TinkerBackend that are fully backwards-compatible with existing code.

1. Built-in Loss Functions

Adds support for Tinker's optimized built-in loss functions via new parameters on TinkerBackend.train():

• tinker_loss_fn: Select from "importance_sampling", "ppo", "cispo", "dro"
• tinker_loss_fn_config: Pass loss-specific config (e.g., {"clip_low_threshold": 0.0, "clip_high_threshold": 6.0})

Benefits:

• Uses forward_backward_async instead of forward_backward_custom_async
• ~1.5x fewer FLOPs per training step
• Up to 3x faster wall time (per Tinker docs)

Default behavior unchanged - when tinker_loss_fn=None (default), continues to use ART's custom loss implementation.

2. Checkpoint Control

The existing save_checkpoint parameter now controls checkpoint behavior in TinkerBackend:

• save_checkpoint=True (default): Saves full state + optimizer (enables training resumption)
• save_checkpoint=False: Only saves sampler weights (fast, for inference only)

This enables faster training when full checkpoints are only needed at specific intervals (e.g., at eval steps).

Usage

backend = art.TinkerBackend()

# Use Tinker's built-in CISPO (faster)
result = await backend.train(
    model,
    trajectory_groups,
    learning_rate=5e-6,
    tinker_loss_fn="cispo",
    tinker_loss_fn_config={"clip_low_threshold": 0.0, "clip_high_threshold": 6.0},
    save_checkpoint=False,  # Fast: only saves sampler weights
)

# Use ART's custom loss (default, backwards compatible)
result = await backend.train(
    model,
    trajectory_groups,
    learning_rate=5e-6,
    ppo=False,
    epsilon=1.0,
)

Files Changed

• src/art/dev/train.py: Added tinker_loss_fn, tinker_loss_fn_config, tinker_save_checkpoint to TrainConfig
• src/art/tinker/backend.py: Overrode train() with new parameters
• src/art/tinker/service.py: Added dispatch logic for built-in vs custom loss, added _save_sampler_weights_only() method

Backwards Compatibility

All existing code continues to work unchanged. The new parameters are optional with sensible defaults that preserve current behavior.

[bradhilton]

looks like some checks aren't passing. and have we tested this?

[corbt]

have we tested this?

No, I tested some of the more foundational pieces in earlier PRs (eg. the move from model.train to backend.train) but since this one is harder to test independently I was planning on testing it as part of the Simple code migration, which will depend on this functionality.

I can wait to merge until that run is working probably. (This will likely turn into a mega-PR with lots of unrelated changes along the way though as I find more pieces I need to update.)

[corbt]

replaced with #532