All issues

Search results

/v1/audio/transcriptions crashes with cryptic TypeError when pointed at a chat/omni model (e.g. Qwen3-Omni)

Status: Open.

#2319 In Blaizzy/mlx-vlm;

· hldz-top opened on Sep 20, 2026

Feature request: native Laya RLCD decision-checkpoint inference

Status: Open.

#2317 In Blaizzy/mlx-vlm;

· eloe opened on Sep 20, 2026

deepseek-vl2 (deepseek_vl_v2): any image input gives degenerate or empty output; text-only prompts work; bf16 and 4-bit both affected

Status: Open.

#2315 In Blaizzy/mlx-vlm;

· jrp2014 opened on Sep 19, 2026

GenerationBatch._eval_pending_state leaks whole KV caches via a self-referencing closure (64 GB peak at batch 32 instead of 11 GB)

Status: Open.

#2310 In Blaizzy/mlx-vlm;

· cjnuk opened on Sep 19, 2026

Qwen3.5/3.6 MTP: ~10% per-cycle speculative-decode slowdown from 0.6.3 → 0.7.1 (measured via a downstream integration; cause not localized)

Status: Open.

#2307 In Blaizzy/mlx-vlm;

· shironexus-gif opened on Sep 18, 2026

20260918 -- Diagnostic update: 49-model sweep on 0.7.1 @ 4774f3d7 — three fixes verified, no timing regressions, four faults remaining

Status: Open.

#2305 In Blaizzy/mlx-vlm;

· jrp2014 opened on Sep 18, 2026

Document video= in apply_chat_template: num_videos=1 produces no video placeholder with Qwen2.5-VL

Status: Open.

#2302 In Blaizzy/mlx-vlm;

· viavicdev opened on Sep 18, 2026

deepseekocr: hard-coded BOS and tokenizer whitespace loss cause hallucinated page preamble with jina-ocr-v1 (DeepSeek-OCR fine-tune)

Status: Open.

#2300 In Blaizzy/mlx-vlm;

· cjnuk opened on Sep 18, 2026

HyperConnection: the exact fused HC+norm path is skipped for single-token decode (L == 1); measured 1–3% decode on GLM-5.3-Flash from fusing it

Status: Open.

#2281 In Blaizzy/mlx-vlm;

· freddyhaddad opened on Sep 16, 2026

glm5_next: remaining KDA decode kernel-launch overhead after #2127 (spec + measurements for a fused/compiled step)

Status: Open.

#2254 In Blaizzy/mlx-vlm;

· freddyhaddad opened on Sep 14, 2026

0.7.0 @ 45d6e125: Nemotron-3-Nano-Omni 4-bit rejects 729 flat backbone/lm_head weights; Mistral-Small-3.2 conversion's text-only chat template raises TypeError on mistral3 list content

Status: Open.

#2246 In Blaizzy/mlx-vlm;

· jrp2014 opened on Sep 13, 2026

server /v1/chat/completions intermittently hangs (0% CPU, no tokens, no error) when request includes a tools array

Status: Open.

#2239 In Blaizzy/mlx-vlm;

· CharafChnioune opened on Sep 13, 2026