#2368 [Feature Request] Tracking Issue for Grammar-Constrained Sampling

Beschreibung

To avoid duplicates, having an open tracking issue while the feature is in backlog helps prevent people from creating issues and makes closing dupes faster.

I gathered all issues that can be traced back to lack of grammar sampling during chat completion atm:
- low-bit quants or low-thinking finetunes ocassionally make malformed tool calls (e.g., createdAt instead of created_at, wrong tool call format, extra brackets): #1431
- grammar sampling as robust approach to enforce tool calls separate from reasoning/output: #210, #804

Also PRs: #636, #1328

Non grammar-constrained approaches PRs: #762, #763, #790, #525

Unsure:
- JSON structured output api route: #76, #976
  - resolved by PR #334? then why was `#976` an issue?)

Let me know in comments if anything missing or if i misunderstood something.