Skip to content

fix(server): do not auto-select qwen3 parser for instruct-2507 - #564

Open
SanjayS45 wants to merge 1 commit into
FlashML-org:mainfrom
SanjayS45:fix/qwen3-instruct-2507-reasoning-auto
Open

SanjayS45 wants to merge 1 commit into
FlashML-org:mainfrom
SanjayS45:fix/qwen3-instruct-2507-reasoning-auto

Conversation

@SanjayS45

@SanjayS45 SanjayS45 commented Sep 28, 2026 •

Copy link
Copy Markdown

emits plain answers, so auto qwen3 left message.content empty.
KarrAcaRn pushed a commit to KarrAcaRn/FreeToken-ByAI that referenced this pull request Oct 3, 2026
KarrAcaRn pushed a commit to KarrAcaRn/FreeToken-ByAI that referenced this pull request Oct 3, 2026
Logs FlashML-org#565 (reimplemented), FlashML-org#564, FlashML-org#562, FlashML-org#559 (with fixups), FlashML-org#558
(superseded by FlashML-org#305) and FlashML-org#305 (pulled forward, adopted).

Assisted-by: Claude Opus 5.5
KarrAcaRn pushed a commit to KarrAcaRn/FreeToken-ByAI that referenced this pull request Oct 5, 2026
…er for every qwen3 instruct checkpoint

FlashML-org#564 matched only Instruct-2507. Qwen3-VL-*-Instruct and Qwen3-Coder-*-Instruct
templates never open a think block either, so auto picked qwen3, which assumes
an open think block, and every plain answer came back as reasoning_content with
an empty message.content (checked against Qwen/Qwen3-VL-8B-Instruct's template).

Assisted-by: Claude Opus 5.5

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[BUG] --reasoning-parser auto gives non-thinking Instruct checkpoints a thinking parser, so message.content comes back empty

1 participant