boltz-cp

Integrate DTensor-based context parallelism (CP) into a custom co-folding / structure-prediction model — sharding the O(N²) pair and attention tensors across a multi-GPU device mesh via explicit ring/transpose collectives. A top-level cpize_model_workflow conductor prioritizes, sequences, and gates the whole effort end-to-end, delegating each phase to focused skills. Explores the user's model and data pipeline, builds and probes distributed test infrastructure (local GPUs or SLURM), implements DTensor data-feature sharding and CP model modules mapped to a located CP reference implementation (e.g. Boltz-CP, via $BOLTZ_CP_REPO), writes serial-as-ground-truth multi-rank parity tests, dispatches the work across an orchestrated coder+reviewer team, and benchmarks max-token / walltime. Ships a SessionStart hook that injects the CP hard-rules and an opt-in PreToolUse guard protecting serial ground-truth files and distributed-test timeouts.

10 skills · hooks