Update mlx dependencies (#350)
* Handle request-owned Qwen MRoPE state * Use mlx-vlm cache classes for batched vision * Exclude quantized Qwen caches from singleton shortcut * Remove obsolete LFM2-VL projector patch * Refresh requirements * Handle text position IDs in prompt slicing
N
Neil Mehta committed
86bb0969dfb7d16faab0f34c4f36fd2b3c9a3fc8
Parent: 60d28b3
Committed by GitHub <noreply@github.com>
on 7/22/2026, 8:13:53 PM