2026-05 Monthly summary for unsloth-zoo: Delivered substantial improvements across memory efficiency for VLM training, fine-tuning configurability, loader safety for pre-quantized models, Apple Silicon stability fixes, and VLM loading/export reliability. These efforts reduced memory pressure, improved model targeting flexibility, and hardened the training pipeline across architectures, enabling safer, faster iteration with larger configurations.

April 2026 performance summary for unsloth-zoo (unslothai/unsloth-zoo). Focused on reliability, performance, and reproducibility across VLM and MLX workflows. Delivered robustness in VLM loading, unified MLX training pipeline, pre-training/patch workflows for Qwen3.5, and expanded IO/export capabilities with hardware-aware dtype handling. These changes reduce training-time errors, streamline deployment, and improve model portability. Key features delivered: - VLM Model Loading Robustness Fix: address missing gradients and __new__ without __init__ during VLM loading to ensure training stability. (Commit 317c7de...) - VLM Development and Runtime Patches: added mlx_compile.py for compile qualification and runtime patches; alias-aware key resolution; refactoring to avoid monkey-patching and stabilize runtime quantization. (Commits 233d0f0...; 856efd7...) - Qwen3.5 Attention Cache None Crash Fix: compute position_ids from scratch to prevent cache=None crash during training. (Commit cab1ad0...) - Qwen3.5 Training Patch Setup: install mlx compile patches before training to fix cache/offset issues. (Commit 5fca3ec...) - MLX Training Workflow Improvements: unify step and data prep for VLM/text; memory-efficient VJP for GatedDeltaNet; remove dead CCE code; add clear_cce_cache for kernel cache management. (Commits 492f98f...; 13e65293...) - Gemma3-family VLM Training Fixes: multiple small fixes to unblock Gemma3-family VLM training (image/attention mask handling, mask alignment, and forward paths). (Commit c44757cd...) - Model IO, Saving and Hub Enhancements: dtype parameter for from_pretrained; full_finetuning mode; revision/random_state plumbing; save_pretrained_merged and hub/push improvements. (Commits 0ac3ddde...; 116d4cd8...; 5e9aeee1...; c0986f75...; 521f0036...) - GGUF Export Workflow and Saving Enhancements: enable GGUF export, ensure gguf package is installed, and expose first_conversion in save_pretrained_gguf. (Commits b0105012...; 175b7721...; cc973ce2...) - Auto-detect Best Data Type by Hardware: auto-pick dtype when None based on chip (M3+ bf16; M1/M2 float16) to optimize performance. (Commit 22123995...) Major bugs fixed: - Fix for missing gradients and __new__ handling during VLM loading, reducing silent training-time gradient issues. - Patch addressing Qwen3.5 attention cache None crash during training to prevent stalls. - Gemma3-family training issues resolved via targeted fixes to image/attention masks and forward paths. - Fix for HF Hub authentication path and save workflow integration to ensure reliable model pushes. Overall impact and accomplishments: - Significantly increased training stability and reproducibility across VLM and MLX workflows, reducing downtime due to loading/gradient issues and cache-related crashes. - Streamlined MLX trainer flow with unified data prep and memory management, enabling more scalable experiments. - Expanded hardware-aware capabilities and export paths (GGUF) for easier deployment and portability. - Improved model IO and hub workflows, supporting finer control over precision, finetuning, and reproducible experiments. Technologies/skills demonstrated: - PyTorch-based VLM pipelines, MLX quantization and runtime patches, and patch governance. - Advanced training workflow engineering, including data path unification and resource management. - Hardware-aware dtype selection and GGUF export integration for production-grade deployment. - Strong emphasis on reproducibility (revision/random_state plumbing, deterministic LoRA/peft handling) and robust model versioning.

March 2026 monthly summary focusing on data ingestion expansion, training workflow stabilization, and improved developer tooling for faster onboarding and reliable model training. Delivered multi-format local dataset upload, enhanced CLI export/training, robust training lifecycle controls, critical state bug fixes, embedding-aware model handling, and project structure cleanup. These changes reduce data preparation time, strengthen observability, and improve maintainability.