mirror of
https://github.com/zenlm/gym.git
synced 2026-11-04 17:11:44 +00:00
Major updates: - Implemented BitDelta 1-bit quantization (10-25× memory reduction) - Added DeltaQuant flexible quantization (INT2/4/8, Binary, Ternary) - Created DeltaSoup for Byzantine-robust community aggregation - Configured accurate Qwen3 models for Zen family (Nano, Eco, Coder, Omni, Next) - Added GRPO/GSPO training algorithms with memory optimizations - Simplified CLI from gym-cli to gym - Updated WebUI with Zoo Labs black monochromatic theme - Integrated Unsloth optimizations for 2-3× speedup - Added comprehensive test suite for quantization methods Zoo Labs Foundation Inc. - Copyright 2025
3.7 KiB
3.7 KiB
Zen Model Repository Structure
Zoo Labs Foundation - AI Model Collection
📁 Model Organization
Each model has its own repository with standardized structure:
models/
├── nano/ # Zen Nano (0.6B) - Edge deployment
├── eco/ # Zen Eco (4B) - Production balance
├── coder/ # Zen Coder (7B MoE) - Code generation
├── omni/ # Zen Omni (14B MoE) - Multimodal
└── next/ # Zen Next (32B MoE) - Advanced reasoning
🚀 Quick Training Commands
# Train any model by navigating to its directory
gym train models/nano/configs/gspo_training.yaml # Zen Nano (0.6B)
gym train models/eco/configs/gspo_training.yaml # Zen Eco (4B)
gym train models/coder/configs/gspo_training.yaml # Zen Coder (7B MoE)
gym train models/omni/configs/gspo_training.yaml # Zen Omni (14B MoE)
gym train models/next/configs/gspo_training.yaml # Zen Next (32B MoE)
📊 Model Comparison
| Model | Size | Architecture | VRAM | Speed | Best For |
|---|---|---|---|---|---|
| Nano | 0.6B | Dense | 2GB | 120 tok/s | Edge, Mobile |
| Eco | 4B | Dense | 8GB | 85 tok/s | Production APIs |
| Coder | 7B | MoE | 16GB | 65 tok/s | Code Generation |
| Omni | 14B | MoE | 24GB | 45 tok/s | Multimodal |
| Next | 32B | MoE | 48GB | 30 tok/s | Complex Tasks |
🔧 Directory Structure
Each model directory contains:
model_name/
├── README.md # Model documentation
├── configs/
│ ├── gspo_training.yaml # GSPO training config
│ ├── grpo_training.yaml # GRPO training config (if applicable)
│ └── inference.yaml # Inference configuration
├── checkpoints/ # Saved model checkpoints
├── scripts/ # Model-specific scripts
└── examples/ # Usage examples
🎯 Training Algorithms
-
GSPO (Group Sequence Policy Optimization)
- Best for: Nano, Eco, MoE architectures (Omni, Next)
- Benefits: 40-60% memory reduction, sequence-level optimization, MoE stabilization
-
GRPO (Group Relative Policy Optimization)
- Best for: Coder (when precision needed)
- Benefits: Token-level precision, faster convergence
📈 Performance Benchmarks
Language Understanding (MMLU)
- Next (32B MoE): 81.2%
- Omni (14B MoE): 74.6%
- Coder (7B MoE): 68.3%
- Eco (4B): 62.5%
- Nano (0.6B): 45.2%
Code Generation (HumanEval)
- Next (32B MoE): 78.9%
- Coder (7B MoE): 72.4% ⭐
- Omni (14B MoE): 65.2%
- Eco (4B): 48.7%
- Nano (0.6B): 28.3%
🛠️ Advanced Training Options
Multi-GPU Training
CUDA_VISIBLE_DEVICES=0,1,2,3 gym train models/[model]/configs/gspo_training.yaml
DeepSpeed Integration
gym train models/[model]/configs/gspo_training.yaml \
--deepspeed configs/deepspeed/ds_z3_config.json
Custom Datasets
gym train models/[model]/configs/gspo_training.yaml \
--dataset your_dataset \
--dataset_dir /path/to/data
Quantization (4-bit)
gym train models/[model]/configs/gspo_training.yaml \
--quantization_bit 4 \
--load_in_4bit true
📝 Creating Custom Configurations
To create a custom training configuration:
- Copy the base config from your model's directory
- Modify parameters as needed
- Run with your custom config:
gym train path/to/your/custom_config.yaml
🔗 Resources
Copyright © 2025 Zoo Labs Foundation Inc.
501(c)(3) Non-Profit Organization