- Adds optional upstream batch runtime (use_batch_runtime) with configurable
attention backend, vision attention, scheduler, group_size, and strict_attn
- Applies runtime compatibility patches to model snapshot files so the node
works with current transformers releases without manual cache edits
- Adds top_k sampling parameter to both predict and predict_batch paths
- Forces temperature=0.0 on CUDA to avoid device-side multinomial asserts
- Retries automatically with generation_mode=slow on empty coordinate output
or degenerate repeated-token responses
- Changes attention default from sdpa to eager for broader compatibility
- Bumps version to 0.1.3
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
