Issues

• Gemma 4 checkpoint with untied lm_head (tie_word_embeddings: false) fails to convert

  ml-explore/mlx-lm#2052 · Filter by author JJ2One opened on Nov 11, 2026

• [Feature] Energy-per-token metrics for mlx_lm.benchmark

  ml-explore/mlx-lm#2051 · Filter by author MRH113 opened on Nov 11, 2026

• [Feature] Backward pass for gated_delta custom kernel (enables training)

  ml-explore/mlx-lm#2050 · Filter by author zxy1123 opened on Nov 11, 2026

• gpt_oss: tie_word_embeddings is ignored (lm_head always created)

  ml-explore/mlx-lm#2045 · Filter by author electron-rare opened on Nov 8, 2026
  3

• RFC: Add Kimi K3.1 Pro text support and evaluate DSA reuse

  ml-explore/mlx-lm#2042 · Filter by author Kalwaleed opened on Nov 7, 2026

• Phi3Small handles left-padding incorrectly in block-sparse attention

  ml-explore/mlx-lm#2037 · Filter by author Knightkolla opened on Nov 4, 2026
  1

• [Feature Request] Expose a scriptable/programmatic Python API for mlx_lm.benchmark

  ml-explore/mlx-lm#2035 · Filter by author avk-dwvd opened on Nov 4, 2026

• fix(server): return an error for a prompt longer than the model context

  ml-explore/mlx-lm#2032 · Filter by author tobocop2 opened on Nov 3, 2026

• server: --adapter-path silently ignored unless model name is default_model

  ml-explore/mlx-lm#2031 · Filter by author haoxu0 opened on Nov 3, 2026

• LoRA loss counts one padding token per sequence

  ml-explore/mlx-lm#2029 · Filter by author Fedeparg opened on Nov 2, 2026
  1

• RotatingKVCache: multi-token mask hides the keep (attention-sink) slots during chunked prefill

  ml-explore/mlx-lm#2028 · Filter by author cyberanonymous05 opened on Nov 2, 2026
  1

• Tool-call span with invalid JSON is dropped entirely — not in tool_calls, not in content, finish_reason still reports tool_calls

  ml-explore/mlx-lm#2019 · Filter by author rsdancey opened on Oct 29, 2026

• Support for MemTensor/MemT-1.0-Nano (memt_qwen3_5_moe)

  ml-explore/mlx-lm#2005 · Filter by author Bayway opened on Oct 24, 2026

• MiniCPM4-0.5B RoPE default differs from the reference

  ml-explore/mlx-lm#2002 · Filter by author WindChimeRan opened on Oct 22, 2026

• Jamba MoE routing weights differ from the reference

  ml-explore/mlx-lm#2001 · Filter by author Knightkolla opened on Oct 22, 2026

• gpt_oss: sanitize() breaks on BF16 checkpoints in current transformers layout ([E, H, 2F]) — '72 parameters not in model'

  ml-explore/mlx-lm#1994 · Filter by author bebopbabailey opened on Oct 19, 2026
  1

• qwen3_coder parser silently truncates parameter value at its first </parameter> (non-greedy _parameter_regex)

  ml-explore/mlx-lm#1992 · Filter by author richgoodson opened on Oct 18, 2026
  5

• Proposal: ternary weights ({-1,0,1}, per-group scale) as a quantization mode

  ml-explore/mlx-lm#1966 · Filter by author Jakevin opened on Oct 12, 2026
  4

• LoRA on Qwen3.5 / Qwen3-Next: step time scales with the number of trained GatedDeltaNet layers (backward runs the per-token ops path)

  ml-explore/mlx-lm#1956 · Filter by author timothyzhbw-jpg opened on Oct 8, 2026
  2

• gemma4: convert fails after transformers 5.17 saves the config again

  ml-explore/mlx-lm#1949 · Filter by author Avicennasis opened on Oct 6, 2026

• Per-module quantization overrides are silently dropped when a model's sanitize() renames weights

  ml-explore/mlx-lm#1924 · Filter by author sje397 opened on Sep 26, 2026
  2

• server: uncaught exception in extract_cache kills the generation thread while the HTTP listener keeps accepting — all subsequent requests hang forever with no error

  ml-explore/mlx-lm#1909 · Filter by author lumatic2 opened on Sep 21, 2026
  1

• qwen3_coder tool parser returns raw string for parameters whose schema has no top-level "type" (oneOf/anyOf)

  ml-explore/mlx-lm#1907 · Filter by author petriematt opened on Sep 21, 2026
  1

• Missing general Knowledge Distillation paradigm

  ml-explore/mlx-lm#1900 · Filter by author diogomsmiranda opened on Sep 17, 2026

• mlx_lm.server: option to restrict which models a request may load (allow-list)

  ml-explore/mlx-lm#1892 · Filter by author mrcushen-arch opened on Sep 15, 2026