All pull requests

Pull requests

• feat: Enable unsloth vLLM standby mode on XPU

  #1747·JoshuaL3000 opened Oct 14, 2026Contributor·

  ·

• MoE LoRA: skip the rank pad when both LoRA GEMMs take the Triton grouped GEMM

  #1746·danielhanchen opened Oct 14, 2026Member·

  ·
  12

• CI: fail loudly when the unsloth install fails instead of skipping its tests

  #1745·danielhanchen opened Oct 14, 2026Member·

  ·
  1

• CI: prune idle superseded pip cache generations, sweep every 2h on the arm label

  #1744·danielhanchen opened Oct 14, 2026Member·

  ·
  1

• feat(mlx): compiled training for Xing4.0 and MiniCPM-V 4.7 (mlx-vlm 0.7.8)

  mlx
  #1743·Lyxot opened Oct 14, 2026Collaborator·

  ·
  7

• Support mlx-vlm 0.7.8

  mlx
  #1742·Lyxot opened Oct 14, 2026Collaborator·

  ·
  4

• fix(mlx): omni models given a Clef head train and serve audio (Gemma 4, Phi-4-multimodal)

  mlx
  #1741·Lyxot opened Oct 14, 2026Collaborator·

  ·
  18

• MLX: never quantize what sits beside the text decoder on a 4-bit multimodal load

  mlx
  #1740·Lyxot opened Oct 14, 2026Collaborator·

  ·
  3

• fix(gemma4): invalidate band-mask verdict after in-place edits

  edge case#
  #1739·taking-lying-flat opened Oct 13, 2026Contributor·

  ·
  1

• fix(mlx): Qwen3-Omni trains on records whose image residuals come as upstream's array

  codex-converged#important#mlx
  #1725·Lyxot opened Oct 13, 2026Collaborator·

  ·
  3

• feat(mlx): add CPO preference training at parity with TRL's CPOTrainer

  codex-converged#mlxnew-feature#
  #1720·Lyxot opened Oct 13, 2026Collaborator·

  ·
  14

• Respect gradient clipping and max_steps settings in unsloth_train

  important#
  #1634·taking-lying-flat opened Oct 8, 2026Contributor·

  ·

• Block Swap: report per-layer placement and copy/stall timing, adapt prefetch depth with prefetch_depth="auto", and refuse a swap that won't fit in host RAM

  new-feature#
  #1619·LeoBorcherding opened Oct 7, 2026Collaborator·

  ·

• AMD MI50 / MI60 (gfx906): opt-in tiled attention kernel for Gemma-4

  AMDedge case#new-feature#
  #1315·namnguyen0503 opened Sep 21, 2026·

  ·
  1

• docs: trim pyproject.toml comments

  conflicts#
  #1283·alkinun opened Sep 18, 2026Contributor·

  ·

• Drop the vision and audio tensors from a text_only merged_16bit export

  conflicts#edge case#
  #1097·Zelys-DFKH opened Aug 25, 2026·

  ·
  6

• Feat(mlx): Support group_by_length for text training

  conflicts#edge case#new-feature#
  #995·BardiaKoopah opened Aug 4, 2026Contributor·

  ·
  7

  

• Fix Qwen3.5 fp16 training dtype mismatches under UNSLOTH_FORCE_FLOAT32

  conflicts#edge case#
  #978·chakshu-dhannawat opened Aug 4, 2026·

  ·
  35

• feat(mlx): quantization-aware training for LoRA on Apple Silicon

  edge case#new-feature#
  #970·BardiaKoopah opened Jul 31, 2026Contributor·

  ·
  19

  

• add confidence weighted CE (DFT) as fused loss function

  conflicts#new-feature#refactor#
  #903·0xEljh opened Jul 13, 2026·

  ·
  24

• dataset_utils: add mask_out_tokens to train_on_responses_only (fixes unslothai/unsloth#6695)

  conflicts#edge case#new-feature#
  #852·pjordanandrsn opened Jul 3, 2026Contributor·

  ·
  10

• feat(mlx): load GPTQ/AWQ pre-quantized checkpoints on Apple Silicon

  conflicts#edge case#new-feature#
  #848·BardiaKoopah opened Jul 2, 2026Contributor·

  ·
  69

  

• dataset_utils: add mask_thinking_tokens to mask the </think> token

  conflicts#edge case#new-feature#refactor#
  #833·Sushankthatipally opened Jun 26, 2026·

  ·
  10

• Feat(mlx): GRPO support for MLX

  conflicts#new-feature#regresses#
  #832·BardiaKoopah opened Jun 26, 2026Contributor·

  ·
  37

  

• fix: decode llama.cpp subprocess output as UTF-8 on Windows

  conflicts#security#
  #764·dylanschroers opened Jun 11, 2026·

  ·
  3