Search results

LoRaLayer uses raw alpha instead of alpha/rank — 8x scaling error with default settings

Status: Open.

#845 In Blaizzy/mlx-vlm;

· kikoncuo opened on Mar 21, 2026

LoRA scaling uses alpha instead of alpha divided by rank, causing default settings to over-scale updates and produce unusuable adapters

Status: Open.

#841 In Blaizzy/mlx-vlm;

· quist00 opened on Mar 20, 2026

In multi-modal training of qwen models tokenizer-level chat templating can fail for list-based multimodal content

Status: Open.

#840 In Blaizzy/mlx-vlm;

· quist00 opened on Mar 20, 2026

mlx_vlm.server returns 500 on GLM-OCR region crops due to UTF-8 decode in BPEStreamingDetokenizer.add_token

Status: Open.

#837 In Blaizzy/mlx-vlm;

· anthonyarguedas opened on Mar 19, 2026

Multi-turn conversations with images fail: 'Image features and image tokens do not match'

bugSomething isn't workingSomething isn't working

Status: Open.

#833 In Blaizzy/mlx-vlm;

· nakul-malhotra opened on Mar 19, 2026

Cross-turn image KV cache with Qwen3.5: vision tower re-runs on every turn

Status: Open.

#832 In Blaizzy/mlx-vlm;

· AirRunner opened on Mar 19, 2026

GLM-OCR LoRA fine-tuning: image tokens not properly handled during training

Status: Open.

#830 In Blaizzy/mlx-vlm;

· zkrige opened on Mar 18, 2026

Improve mlx-vlm type hints for generate() / stream_generate() so Pyright/Pylance can catch API mistakes

Status: Open.

#825 In Blaizzy/mlx-vlm;

· jrp2014 opened on Mar 15, 2026

LoRA training broken for Qwen3.5 VLM: 3 bugs causing corrupted generation

Status: Open.

#824 In Blaizzy/mlx-vlm;

· kikoncuo opened on Mar 15, 2026

20260313 -- Diagnostic Report

Status: Open.

#819 In Blaizzy/mlx-vlm;

· jrp2014 opened on Mar 13, 2026

📋 Documentation Enhancement Suggestion

Status: Open.

#812 In Blaizzy/mlx-vlm;

· croviatrust opened on Mar 11, 2026

Some suggestions for mlx-vlm diagnostic improvements

Status: Open.

#802 In Blaizzy/mlx-vlm;

· jrp2014 opened on Mar 7, 2026