[Feature Request] do not switch away from Fable when working on trying to add Support for open-weight models with custom attention mechanisms

Question

**Bug Description**
"I was just working on modifying mlxcel to try to support an open weight model with a new sparse attention approach, no security stuff, nothing to do with anything unsafe or even unsavoury.  Kind of the whole point of having a coding agent solving complex issues, no?"

**Environment Info**
- Platform: darwin
- Terminal: tmux
- Version: 2.1.179
- Feedback ID: b614cf62-6928-42ea-885f-c19f2911335f

**Errors**
```json
[]
```