v0.2.0

Feature and correctness release on top of v0.1.12, with additional image-capable renderers, richer inference metadata, and renderer parity fixes.

Compatibility

• Supports Python 3.11–3.14.
• generate() now returns completion_logprobs as a NumPy float32 array rather than a Python list. Consumers that need a list can use .tolist() (#201).

New renderers

• GLM-5.3 Flash with image support (#191).
• Muse Glimmer with image support (#194).

Added

• Transport sampler probability heads for score centering, including compact logprobs and complete sampling replay support (#180).
• Pass through vLLM weight_versions spans over completion tokens (#195).

Fixed

• Preserve the newline before the second and later GLM-4.5 tool calls when an assistant message contains both text and multiple calls (#203).
• Render non-leading system messages in Nemotron 3 Nano, Super, Ultra, and 3.5 Lightning consistently with their official templates (#200).
• Align Muse Glimmer tool parsing and answer whitespace with vLLM (#197).

Performance

• Read packed completion logprobs directly, avoiding an unnecessary float64/list conversion (#174, #201).
• Load client errors lazily so importing renderer configs does not eagerly import OpenAI (#199).

Full changelog: renderers-v0.1.12...renderers-v0.2.0