●Change8

3.0.3

Breaking Changes
📦 dspyView on GitHub →
⚠ 1 breaking✨ 5 features🐛 4 fixes🔧 5 symbols

Summary

dspy 3.0.3 adds rollout_id support, improves GEPA optimizer, updates caching behavior, and includes several bug fixes and maintenance updates.

⚠️ Breaking Changes

  • LiteLLM caching was removed from the LM class; code that relied on the deprecated LiteLLM cache must be updated to use the new caching mechanism or disable caching.

Migration Steps

  1. Remove any usage of the deprecated LiteLLM caching flags from `dspy.LM`.
  2. If you relied on the old cache behavior, configure caching explicitly using the new API.
  3. Update code that manually calls `forward` to rely on the patched behavior or adjust calls accordingly.
  4. When performing multiple rollouts, you may need to adjust temperature settings if you previously set them manually.

✨ New Features

  • Added `rollout_id` parameter to allow namespaced bypass of the LM cache.
  • Temperature is automatically increased for multiple rollouts and a warning is emitted when a single rollout is performed.
  • GEPA optimizer upgraded: new version includes a progress bar and ensures failing metrics stop execution.
  • Added type hints and corrected the docstring for the SIMBA optimizer.
  • Reduced the `max_tokens` requirement for reasoning language models.

🐛 Bug Fixes

  • Fixed `__call__` forwarding in the GEPA tutorial example.
  • Patched the `forward` method handling to avoid warnings on explicit `forward` calls.
  • Corrected Predict configuration merging logic and documented `rollout_id` usage.
  • Replaced `ujson` with `orjson` for faster JSON handling.

🔧 Affected Symbols

dspy.LMdspy.GEPAdspy.SIMBAdspy.predictdspy.forward