Issues

• Unlimited-OCR-2: page-window decoding and KV-cache reclamation for long documents

  Blaizzy/mlx-vlm#2604 · Filter by author Hor1zonZzz opened on Oct 27, 2026
  1

• Speculative decoding with response_format (JSON schema) for MTP drafters

  Blaizzy/mlx-vlm#2594 · Filter by author JarJarFran opened on Oct 25, 2026
  1

• Mandatory datasets and mlx-audio dependencies restrict unrelated upgrades in inference-only environments

  Blaizzy/mlx-vlm#2591 · Filter by author jrp2014 opened on Oct 24, 2026
  3

• Quantization Aware Training?

  Blaizzy/mlx-vlm#2589 · Filter by author Goekdeniz-Guelmez opened on Oct 24, 2026
  1

• Follow-up proposals from the cross-model compatibility sweeps

  Blaizzy/mlx-vlm#2586 · Filter by author jrp2014 opened on Oct 23, 2026
  1

• APC: make_warm_batch_kv_cache_multi pads values with the layer-0 key shape — GLM-4.7-Flash (MLA) batches fail with [concatenate] HTTP 500s

  Blaizzy/mlx-vlm#2585 · Filter by author bencodestx opened on Oct 23, 2026
  2

• 20261023 -- Diagnostic update: 51-model sweep on 0.7.10 @ 4319ef2c: no crashes; #2228 fixed by #2545; #2174 down to a CLI option

  Blaizzy/mlx-vlm#2584 · Filter by author jrp2014 opened on Oct 23, 2026
  1

• qwen4_exp + DFlash: four defects block the native drafter path (and one silently corrupts output)

  Blaizzy/mlx-vlm#2581 · Filter by author yeeeeff opened on Oct 23, 2026
  1

• Add Hi3D v1.0 (Sparc3D mesh refinement), I'm working on it

  Blaizzy/mlx-vlm#2578 · Filter by author cguldogan opened on Oct 23, 2026
  2

• Kev-27B decision model (jaredpalmer/kev-27b) on the shared decision API

  Blaizzy/mlx-vlm#2559 · Filter by author JarJarFran opened on Oct 21, 2026
  1

• MTP to the trainer

  Blaizzy/mlx-vlm#2554 · Filter by author Goekdeniz-Guelmez opened on Oct 19, 2026
  1

• The best and the worst models for captioning images

  Blaizzy/mlx-vlm#2546 · Filter by author jrp2014 opened on Oct 17, 2026
  1

• Private contact for a security report

  Blaizzy/mlx-vlm#2534 · Filter by author RobbieJ opened on Oct 16, 2026
  1

  

• Batched MTP speculative decoding: surviving rows use stale M-RoPE rope_deltas after an earlier row finishes

  Blaizzy/mlx-vlm#2532 · Filter by author kod201 opened on Oct 16, 2026
  1

• 20261016 -- Diagnostic update: image recognition and description sweep on 47 models, with a lot of library changes this week

  Blaizzy/mlx-vlm#2530 · Filter by author jrp2014 opened on Oct 16, 2026
  1

• convert --mtp drops q_mode, silently pairing a non-affine main model with an affine drafter

  Blaizzy/mlx-vlm#2526 · Filter by author magnuslundstedt opened on Oct 15, 2026
  1

• qwen3_vl_embedding: batched text embeddings after an image call are wrong (cached position IDs reused)

  Blaizzy/mlx-vlm#2523 · Filter by author sayyidfareed opened on Oct 14, 2026
  2

• Reranker: Qwen3-Reranker-0.6B scores every document the same; its chat_template.jinja expects query/document roles

  Blaizzy/mlx-vlm#2520 · Filter by author sayyidfareed opened on Oct 14, 2026
  1

• Embeddings: include_prompt=false is refused even for CLS pooling (granite-embedding-311m-multilingual-r2)

  Blaizzy/mlx-vlm#2519 · Filter by author sayyidfareed opened on Oct 14, 2026
  2

• BatchGenerator greedy decoding: identical prompts at B=2 and B=4 produce different tokens than B=1 (B=8 matches)

  Blaizzy/mlx-vlm#2518 · Filter by author griffinwork40 opened on Oct 14, 2026
  1

• batch_generate: left-padded rows decode at wrong RoPE positions for Qwen3-VL text-only prompts (rope_deltas subtracts the padding twice)

  Blaizzy/mlx-vlm#2517 · Filter by author MarcusXLee opened on Oct 14, 2026
  1

• Qwen3-VL LoRA: mx.eval(deepstack_visual_embeds) in get_input_embeddings costs ~10 GB per training step

  Blaizzy/mlx-vlm#2512 · Filter by author wrwei opened on Oct 14, 2026
  1

• GLM-OCR SFT training runs without a causal mask: the trainer's attention mask replaces it

  Blaizzy/mlx-vlm#2509 · Filter by author fabriziosalmi opened on Oct 13, 2026
  1

• Qwen3.x-27B MTP: fused GDN decode keeps a second copy of the in-projection weights (~2.28 GB)

  Blaizzy/mlx-vlm#2508 · Filter by author shax71 opened on Oct 13, 2026
  2

• Server: Devstral / tekken checkpoints fail every request (HTTP 500, 'PixtralImageProcessor' has no attribute 'pad_token') and three defects behind it

  Blaizzy/mlx-vlm#2507 · Filter by author eejd opened on Oct 13, 2026
  1