Issues

• warmup skipped in drafter mode: full-block S=32 requests 1024 threads, refused at 832 on M1/M2

  ARahim3/mlx-dspark#41 · Filter by author Tobias-Conrad opened on Oct 4, 2026
  1

• M1 Max 64GB: speculative decoding is a net loss on Qwen3.8-27B-8bit (all modes/caps < 1.0×)

  ARahim3/mlx-dspark#38 · Filter by author ztaro21 opened on Sep 16, 2026

• prefix cache: rung ladders lost on partial-hit inheritance and WARN sheds; checkpoint persistence across restarts (patch, still applies to 0.18.1)

  ARahim3/mlx-dspark#36 · Filter by author felix-ab opened on Sep 10, 2026
  1

• dflash: drafter RotatingKVCache offset runs ahead of an empty cache after a prefix-cache restore miss -> 'Negative dimensions not allowed' (500)

  ARahim3/mlx-dspark#33 · Filter by author leoli-dev opened on Sep 3, 2026
  1

• SIGSEGV in matmul_bnns (BNNSFilterApplyTwoInputBatch) during CPU co-prefill - macOS 26.5.2, M4 Pro

  ARahim3/mlx-dspark#31 · Filter by author amusman opened on Aug 30, 2026
  4

• Show prefill progress for long external requests

  ARahim3/mlx-dspark#29 · Filter by author cryptedx opened on Aug 27, 2026
  1

• LM Studio models get downloaded twice on first load

  ARahim3/mlx-dspark#28 · Filter by author koc-Z3 opened on Aug 24, 2026
  1

• Claude Code's /effort is ignored — support output_config.effort

  ARahim3/mlx-dspark#25 · Filter by author carnastone opened on Aug 22, 2026
  1

• Engine v0.13 dflash (DFlash 2) mode: long generations stall ~100 s mid-stream — clients with idle timeouts fail

  ARahim3/mlx-dspark#19 · Filter by author TianjinAI opened on Aug 19, 2026
  5

• Cannot download or load any models, greyed out.

  ARahim3/mlx-dspark#15 · Filter by author zacharisharris opened on Aug 18, 2026
  5

• Load models from LM Studio

  ARahim3/mlx-dspark#12 · Filter by author neokree opened on Aug 17, 2026
  1

• M5 Max (64 GB) benchmark run — Qwen3.8-27B-8bit, plus a same-prompt llama.cpp comparison

  ARahim3/mlx-dspark#8 · Filter by author kfirsch opened on Aug 15, 2026
  1

• Server prefix cache never hits (hits: 0) — TTFT does not improve on repeated identical requests

  ARahim3/mlx-dspark#7 · Filter by author joeOGsan opened on Aug 15, 2026
  7

• Proposal: serve Muse-Glimmer vision inputs in-process with the existing target

  ARahim3/mlx-dspark#6 · Filter by author gordoabc opened on Aug 15, 2026
  1

• DSpark mode: intermittent 'generation failed: list index out of range' (500) under sustained batch load

  ARahim3/mlx-dspark#5 · Filter by author franciscosimoesmoura opened on Jul 18, 2026
  1

• gemma4 target fails to load: mlx-vlm Gemma4UnifiedProcessor breaks under transformers 5.12.1

  ARahim3/mlx-dspark#4 · Filter by author jnyer27 opened on Jul 12, 2026

• Can't Get Gemma Model working

  ARahim3/mlx-dspark#3 · Filter by author jnyer27 opened on Jul 11, 2026

• Simple Benchmak on Macbook Air M5

  ARahim3/mlx-dspark#2 · Filter by author yeungtuzi opened on Jul 4, 2026