[Project] PyGo – embedding CPython inside a Go process to build a deep learning framework
I've been working on something a bit unusual: a deep learning framework where Go is the top-level API, Python handles autograd and the model zoo, and C++/CUDA does the raw compute.
The architecture looks like this:
Go API → CGo bridge → CPython (embedded) → pybind11 → CUDA/AVX-512 kernels
The key insight: instead of a Python sidecar in every pod, CPython runs inside the Go binary. Tensors live in shared memory — zero-copy across all three layers.
Why Go on top?
Go is already running most ML infrastructure (K8s, Prometheus, etcd). PyGo makes models first-class citizens there, without rewriting your ops stack. Goroutines make dynamic batching trivial.
Current state:
- LLaMA-3, GPT-2, BERT, ViT, Whisper partially implemented
- Flash Attention v2, GPTQ/AWQ quantisation
- FSDP, DPO, SFT trainers
- Early stage — looking for Go + C++/CUDA contributors
The main unsolved problem is CGo call overhead at the tensor boundary. If anyone has experience embedding CPython in a Go process, I'd love to talk.
Looking for core contributors — especially Go devs with CGo experience, Python autograd engineers, and C++/CUDA kernel writers.
Interested in joining?
Fill this form:https://forms.gle/Tgr2wM2ii64iWxi78
Top comments (0)