Search results

Per-layer mixed-quantization checkpoint (Qwen3.5-4B-OptiQ-4bit) loads with correct per-layer bits but generates garbage; uniform 4bit/6bit fine; Python mlx-lm fine

Status: Open.

#450 In ml-explore/mlx-swift-lm;

· ryuichi-sumida opened on Jul 21, 2026

savePromptCache / loadPromptCache drop LMOutput.State — a disk-restored VLM prompt cache decodes at wrong M-RoPE positions

Status: Open.

#443 In ml-explore/mlx-swift-lm;

· NivDvir opened on Jul 18, 2026

[Feature] Is multi-turn tool calling planned?

Status: Open.

#441 In ml-explore/mlx-swift-lm;

· kennycarruthers opened on Jul 18, 2026

UserInput.Processing could use video options

Status: Open.

#429 In ml-explore/mlx-swift-lm;

· davidkoski opened on Jul 15, 2026

Prompt-lookup (n-gram) speculative decoding: draft-model-free speedup for repetition-heavy generation

Status: Open.

#425 In ml-explore/mlx-swift-lm;

· fdagostino opened on Jul 15, 2026

RotatingKVCache becomes untrimmable once the window wraps: speculative-decoding rollback silently no-ops, and prompt-cache prefix reuse is defeated

Status: Open.

#424 In ml-explore/mlx-swift-lm;

· fdagostino opened on Jul 14, 2026

Qwen2.5-VL / Qwen3-VL: M-RoPE state is dropped across ChatSession turns, so warm continuations run at wrong positions

Status: Open.

#420 In ml-explore/mlx-swift-lm;

· spokvulcan opened on Jul 14, 2026

KVCacheSimple.offset (Swift Int) breaks MLX.compile() decode — cache write position frozen across steps

Status: Open.

#406 In ml-explore/mlx-swift-lm;

· Augustas11 opened on Jul 7, 2026

Qwen3.5/Qwen3-VL: no windowed prefill path for continuing a warm KV cache through new images

Status: Open.

#397 In ml-explore/mlx-swift-lm;

· spokvulcan opened on Jul 5, 2026

Gemma4 native audio support

Status: Open.

#393 In ml-explore/mlx-swift-lm;

· tleyden opened on Jul 4, 2026

[BUG] tests fail due to TF32

Status: Open.

#357 In ml-explore/mlx-swift-lm;

· davidkoski opened on Jun 18, 2026

Gemma 4 E2B/E4B QAT checkpoints fail to load: KV-shared layers omit k_proj/k_norm/v_proj/v_norm, but the attention module creates them unconditionally

Status: Open.

#338 In ml-explore/mlx-swift-lm;

· jonyoder opened on Jun 11, 2026