torchspec-project/TorchSpec

Decouples inference and training via a disaggregated pipeline that streams hidden states from vLLM or SGLang inference engines to distributed training workers through Mooncake's in-memory store, enabling independent scaling of each component. Integrates directly with PyTorch FSDP for distributed training, uses vLLM's Worker Extension API to avoid RPC serialization overhead, and supports vocabulary pruning with HuggingFace checkpoint conversion. Includes production examples for Qwen3, Kimi-K2.5, and MiniMax-M2.5 models with configurable training modes for resuming interrupted runs or continual training from existing weights.