Tinker by Thinking Machines
13 hours ago
- #research
- #api
- #machine-learning
- Tinker is a training API designed for researchers, offering control over model training and fine-tuning while managing infrastructure.
- Key functions include forward_backward (performs forward and backward passes), optim_step (updates weights), sample (generates tokens), and save_state (saves training progress).
- Uses LoRA for efficient fine-tuning by training a small add-on instead of altering all original weights, matching full fine-tuning performance with less compute.
- Handles scheduling, tuning, resource management, and infrastructure reliability, orchestrating distributed training on GPU clusters.
- Supports a range of open-source models, from compact to large, with plans to expand the selection.
- Allows downloading model weights during and after training.
- Initially free, with usage-based pricing to be introduced soon.