Efficient attention, measured before it is rewarded.
CCO is replacing a generic square-matrix competition with a protected exact-attention foundation on an NVIDIA GeForce RTX 5070 Ti.
Current benchmark
attention-foundation-v1.1-rtx5070ti
- 4 causal prefill workloads: 1K, 2K, 4K, 8K
- 3 one-token decode workloads: 1K, 4K, 8K KV context
- 2 correctness/regression guards
- Exact output gate before latency or VRAM is considered
- Balanced same-machine
main → PR → PR → mainmeasurement - Protected run seed, fresh inputs, and validation of every timed output
Current truth
CCO contains exact/tiled GEMM, legacy square-matrix strategies, and attention prototypes. It does not yet prove an FFT attention improvement or an LLM improvement.
The hardened RTX 5070 Ti protocol is calibrated, adversarially certified, and active. Only protected results can authorize a merge.
Contribution status
Focused feature proposals for approved attention-foundation issues may enter protected evaluation. Bug-fix, docs, cleanup, legacy transform, and protected-evaluator PRs are not mining lanes.