Files
zeekay 05cd0c24e7 WIP zen6 wire: qwen3_coder tool calls, reasoning effort as the owner set it, chat_template_kwargs, from-config names and DFlash
Untested: the build was stopped before any test ran.

- tools/parsers/qwen3_coder.rs reads the <tool_call><function=..><parameter=..> format of the Qwen3.5 and 3.8
  templates, several calls a message; the matcher takes its format from the chat template and types each
  argument by its tool's schema, as vLLM's qwen3_coder does. No JSON grammar is forced on it.
- ReasoningEffort names all six levels; a template that resolves xhigh itself gets the name as sent, one
  that does not gets low, medium or high. The router maps high to medium, max to xhigh, minimal to low.
- /v1/chat/completions reads enable_thinking and reasoning_effort from chat_template_kwargs.
- from-config: a model entry takes name (the id requests use), and runtime dflash and prompt-lookup attach.
- scripts/zen_compare.py and scripts/embed_compare.py: the same probes against two servers.
2026-10-12 10:00:03 -07:00
..
2026-09-24 15:01:47 -07:00
2026-09-24 15:01:47 -07:00