unsloth/studio/backend/core/inference
Daniel Han 9dff999ed6 Address review: gate synthesis nudge to small models and skip on errors
- Synthesis nudge (both route-level and loop-level) now only applies to
  small models (<9B). Large models handle multi-step tool use correctly
  (e.g. web_search(query) -> web_search(url=...) -> python(...)) and
  should not be told "Tool calls are no longer needed."

- Added synthesise_after_tool_result parameter to
  generate_chat_completion_with_tools(). Routes pass True only when
  _is_small_model, so the loop-level nudge is also model-size-aware.

- Renamed _has_tool_result_in_current_turn to
  _has_successful_tool_result_in_current_turn. Now checks tool result
  content against error prefixes (Error, Failed to fetch, Blocked, etc.)
  so error-only tool results don't prematurely force synthesis -- the
  model can retry with a different query instead.
2026-04-16 11:31:57 +00:00
..
__init__.py Final cleanup 2026-03-12 18:28:04 +00:00
_html_to_md.py fix: studio web search SSL failures and empty page content (#4754) 2026-04-01 06:12:02 -07:00
anthropic_compat.py Studio: Expose openai and anthropic compatible external API end points (#4956) 2026-04-13 21:08:11 +04:00
audio_codecs.py studio: per-model inference defaults, GGUF slider fix, reasoning toggle (#4325) 2026-03-16 06:37:55 -07:00
defaults.py UI Changes (#4782) 2026-04-02 08:05:55 -07:00
inference.py Pin bitsandbytes to continuous-release_main on ROCm (4-bit decode fix) (#4954) 2026-04-10 06:25:39 -07:00
llama_cpp.py Address review: gate synthesis nudge to small models and skip on errors 2026-04-16 11:31:57 +00:00
orchestrator.py fix(studio): prioritize curated defaults over HF download ranking in Recommended (#4792) 2026-04-02 10:46:53 -07:00
tools.py fix(studio): harden sandbox security for terminal and python tools (#4827) 2026-04-03 13:33:42 -07:00
worker.py split venv_t5 into tiered 5.3.0/5.5.0 and fix trust_remote_code (#4878) 2026-04-07 20:05:01 +04:00