mirror of
https://github.com/zenlm/zen-vl.git
synced 2026-10-30 14:57:35 +00:00
zeekay
f41b6b1d02
- Complete training infrastructure for 4B/8B/30B models - Identity fine-tuning with Zen branding - Function calling with visual context - Visual agent capabilities (GUI interaction, parameter extraction) - Research paper outline and documentation - Makefile automation for complete pipeline Models: - zen-vl-4b-instruct: Base vision-language with identity - zen-vl-4b-agent: Function calling and tool use - zen-vl-8b-instruct/agent: General-purpose VL - zen-vl-30b-instruct/agent: Frontier VL performance Based on Qwen3-VL architecture with: - 256K→1M context window - OCR in 32 languages - Video understanding - Spatial grounding
38 lines
547 B
Plaintext
38 lines
547 B
Plaintext
# Zen VL Requirements
|
|
# Python 3.13 required
|
|
|
|
# Core ML/DL
|
|
torch>=2.0.0
|
|
torchvision>=0.15.0
|
|
torchaudio>=2.0.0
|
|
|
|
# Transformers (dev version for Qwen3-VL)
|
|
transformers>=4.50.0
|
|
accelerate>=0.24.0
|
|
datasets>=2.14.0
|
|
|
|
# Quantization
|
|
bitsandbytes>=0.41.0
|
|
|
|
# Image/Video processing
|
|
pillow>=10.0.0
|
|
opencv-python>=4.8.0
|
|
|
|
# Utilities
|
|
requests>=2.31.0
|
|
tqdm>=4.66.0
|
|
numpy>=1.24.0
|
|
|
|
# HuggingFace
|
|
huggingface-hub>=0.17.0
|
|
|
|
# Optional: MLX for Apple Silicon
|
|
# mlx>=0.0.9
|
|
|
|
# Optional: Flash Attention 2
|
|
# flash-attn>=2.3.0
|
|
|
|
# Development
|
|
pytest>=7.4.0
|
|
black>=23.7.0
|