CGinS — CUDA Ghost in the Shell explores how language models can translate PyTorch operations into CUDA kernels using runtime context and correctness feedback. This repository brings together the research implementation, generated-kernel workspace, and a Jac interface.

Implementation · Research paper · Original documentation · Jac interface guide

The engineering loop

flowchart LR
    A["Profile<br/>PyTorch operation"] --> B["Capture<br/>tensor context"]
    B --> C["Generate<br/>CUDA"]
    C --> D["Compile<br/>and validate"]
    D -->|Feedback| C
    D --> E["Inspect<br/>the kernel"]

Loading

The interesting boundary is between generated code and executable evidence: tensor inputs, compilation results, and comparison with the reference operation all participate in the feedback loop.

Navigate the implementation

Location	Purpose
CGinS-gui/src	Generation and optimization implementation
CGinS-gui/benchmarks	Profiling and benchmark workspace
CGinS-gui/kernels	CUDA kernel workspace
CGinS-gui/cgins_runtime	Runtime integration
CGinS-gui/main.jac	Jac application entry point
CGinS-gui/cgins-frontend	Frontend implementation

Getting oriented

git clone https://github.com/Dhravidk/openTorch.git
cd openTorch/CGinS-gui

Start with the implementation README and Jac guide for environment and provider setup. GPU execution requires a compatible NVIDIA/CUDA environment; generation uses a configured model provider.

Research context

This is an exploratory research implementation. Consult the included paper for its experimental setting and reported results. Performance depends on the operator, workload, hardware, and baseline; the presence of a generated kernel does not establish an application-level speedup.

The original implementation and its documentation remain under CGinS-gui/.