feat(clients): add Together AI fine-tuning provider

[katherineahn]

What

Adds TogetherProvider, a fine-tuning provider for Together AI, mirroring dspy/clients/databricks.py.

Details

finetune() saves training data to JSONL, uploads it, creates a Together fine-tune job, polls until it completes (raises on error/cancelled), and returns the model as together_ai/<name>. together is imported lazily (optional dependency). TrainingJobTogether.status() reports job status, and is_provider_model() recognizes together_ai/ model names.

Tests

tests/clients/test_togetherai.py — a pure-logic test for is_provider_model, plus a mocked test that runs the full finetune() flow against a fake together client (no live API calls; runs in CI).

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR adds Together AI as a fine-tuning provider. The main changes are:

• New TogetherProvider and TrainingJobTogether classes for Together fine-tuning.
• JSONL training data upload, job creation, polling, and prefixed model-name return.
• Tests for provider model detection, pre-start status, and the mocked fine-tuning flow.

Confidence Score: 4/5

Mostly safe to merge after the model-name handling bug is fixed.

The main fine-tuning flow is covered by tests, but prefix removal can submit the wrong base model for names containing together_ai/ after the leading provider prefix.

dspy/clients/togetherai.py

T-Rex Logs

What T-Rex did

• We ran a focused pytest harness against TogetherProvider.finetune with a mocked client and found that the model argument was org/model instead of the expected org/together_ai/model, and the assertion failed.
• We validated the broader contract and confirmed that the provider tests and lint checks completed successfully with exit code 0.
• Artifacts documenting the repro harness, verbose test output, and the pytest and lint results were uploaded.

_{Ran code and verified through T-Rex}

Important Files Changed

Filename	Overview
dspy/clients/togetherai.py	Adds the Together fine-tuning provider flow; prefix stripping still over-applies to internal together_ai/ substrings.
tests/clients/test_togetherai.py	Adds mocked coverage for provider model detection, pre-start status, and the successful fine-tuning path.

_{Reviews (5): Last reviewed commit: "fix(clients): strip together_ai/ prefix ..." | Re-trigger Greptile}

[katherineahn]

Thanks for the review! Addressed:

• Training data is now validated against its declared format via validate_data_format.
• The local JSONL is deleted after upload (try/finally).
• finetune() raises a clear error if a completed job returns no output model name.

On the fresh Together client per poll tick — left as-is intentionally: it's created once per 60s poll and is cheap (mirrors the Databricks provider), so reusing one would add complexity for negligible benefit.

[katherineahn]

Hi @isaacbmiller — thanks for all the review-and-merge work on the repo lately! Checking in on this one when you have a moment.

The Greptile feedback is addressed, and I've just rebased onto the latest main and verified the tests pass locally. The "Lint, Test, and Build" workflow is queued awaiting approval — would you be able to kick it off when you get a chance?

For context: this provider mirrors the existing databricks.py conventions closely, the tests are fully mocked (no live API calls), and I'm happy to maintain it going forward. Happy to make any changes if something's needed — thanks!

[katherineahn]

Quick update: pushed a fix for the model-prefix issue Greptile flagged. Added a _remove_provider_prefix helper (mirroring OpenAIProvider) so the bare model name is passed to Together's fine-tuning API, plus a test asserting that. All Greptile review threads are resolved now. The Lint/Test/Build workflow is still queued awaiting approval to run whenever a maintainer has a moment. Thanks!

[greptile-apps]

Strip only the prefix
_remove_provider_prefix() removes every together_ai/ segment in the model string, not just the routing prefix. A valid bare Together model path that itself contains this substring, e.g. org/together_ai/model, is rewritten to org/model before fine_tuning.create, so the API receives the wrong base model.

Suggested change

	return model.replace(provider_prefix, "")
	return model[len(provider_prefix) :] if model.startswith(provider_prefix) else model

Artifacts

Repro: focused pytest harness for TogetherProvider.finetune model argument

• Evidence file captured while the check ran.

Repro: verbose pytest output showing wrong create model argument and assertion failure

• The full command output behind this check.

_{Ran code and verified through T-Rex}