Cuda-Compute-OSS (CCO)
An evidence-first research project benchmarking exact-attention latency and memory on consumer NVIDIA GPUs, working toward a hybrid attention operator.
Cuda-Compute-OSS is a research project for efficient long-context attention on consumer NVIDIA GPUs. Its long-term target is a hybrid operator combining exact local attention, a learned causal FFT global path, and a selective retrieval path used only when evidence justifies it; the project states plainly that it does not yet claim FFT replaces or improves attention.
Its current phase keeps output exact and asks whether a candidate implementation can reduce attention latency or memory without losing correctness, using a protected benchmark of nine GPU workload configurations (causal prefill and one-token decode at several context lengths, plus causal and non-causal guards) run on a reference RTX 5070 Ti with fixed warm-up counts, repetitions and independently seeded inputs. The repository includes exact in-core and tiled GEMM kernels, exact, local, spectral, correlation and landmark attention prototypes, and an evaluator with a PR policy that merges qualifying approved feature pull requests automatically.
It is written in Python and MIT licensed.