Skip to content

v2.3.3 - Clean streaming output and GGUF loop guard

Choose a tag to compare

@goodguy1963 goodguy1963 released this 26 May 22:44
· 91 commits to main since this release

Fixes

  • Replaces prefixed rolling terminal streaming with clean wrapped token output.
  • Removes visible stream heartbeat/status chatter around token streams.
  • Adds GGUF repeated-token degeneration detection and trims repeated trailing loops.
  • Adds safer GGUF sampler defaults (top_k, min_p, repeat_last_n, presence/frequency penalties) for Qwen repetition failures.

Validation

  • Python compile passed for all ThinkingLLM node files.
  • pytest tests/test_new_features.py -q: 42 passed.
  • git diff --check: passed.
  • JSON validation for custom_nodes.json, hf_models.json, and gguf_models.json: passed.
  • Live ComfyUI GPU smoke test on RTX 3060 with Qwen3-4B-Q4_K_M.gguf: streamed clean text and completed in 3.69s.

Full Changelog: v2.3.2...v2.3.3