PyGo: A Deep Learning Framework Where Go Calls Python Calls C++

Sarkar-AGI — Sun, 19 Jul 2026 07:35:56 +0000

[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://docs.google.com/forms/d/e/1FAIpQLSe8rxglSofehizPqFu5Rb8XjsWXGaOzlh2ZZXHExrr8Qk8Sng/viewform

TitanCore Core-1 – Trillion-parameter LLM training infra in C++/CUDA with ZeRO-3

Sarkar-AGI — Fri, 22 May 2026 12:07:24 +0000

I built TitanCore Core-1, a lightweight core infrastructure (around 75+ files) written in C++ and custom CUDA kernels to address the VRAM bottleneck in trillion-parameter LLM training.

By implementing Fully Sharded Data Parallelism (FSDP) via ZeRO-3 and bypassing standard framework overhead with fused kernels, I managed to hit 890 GB/s memory bandwidth utilization ($2.6\times$ speedup compared to traditional pipelines).

The code is fully open-source. I would love to get your feedback on the custom memory handling and activation checkpointing logic!
GitHub link https://github.com/Sarkar-AGI/Core-1

DEV Community: Sarkar-AGI

PyGo: A Deep Learning Framework Where Go Calls Python Calls C++

TitanCore Core-1 – Trillion-parameter LLM training infra in C++/CUDA with ZeRO-3