Menu
SkillsMP has collected 11 skills from NVIDIA-BioNeMo/boltz-cp. Open a skill to review its source and details.
Showing 11 of 11 collected skills.
Benchmark a CP inference or training workflow to find the maximum token count that fits at a given CP size and to record end-to-end walltime. Sweeps input size (tokens N, atoms, MSA depth S) over real or synthesized data, ramps N until OOM to report the max,…
Probe and establish the distributed test infrastructure for CP development. Inventories local GPUs (count, model, memory, NVLink topology), checks the software stack (Python, PyTorch+CUDA, NCCL, torch.distributed), and runs shipped smoke tests for…
Orchestrate the END-TO-END integration of context parallelism into a custom co-folding / structure-prediction model: turn the whole effort into a prioritized, dependency-sorted worklist, then drive it phase by phase — map the model (learn_context), verify…
Orchestrate a team of agents to integrate many CP modules and tests in parallel. Builds the work-list from the module map, derives a data-flow dependency graph, topologically sorts it into GPU-slot-bounded waves, and spawns a coder+reviewer pair per work…
Stand up the distributed-model lifecycle for a CP training/inference run: device placement before DTensor wrapping, the all-trainable-params-are-DTensors invariant (with placeholder/freeze for unimplemented modules), checkpoint save (DTensor → plain) and load…
Implement DTensor-based context-parallel model modules that mirror a serial reference. Maps each serial layer/module to its Boltz-CP counterpart with exact input/output shapes, placements, collectives, and backward memory budget; writes the CP implementation…
Explore a custom co-folding / structure-prediction model to understand its inference and training workflows, entry points, training framework (PyTorch Lightning vs DeepSpeed), data-feature format, and featurization pipeline — then map those components onto…
Memory-profile a context-parallel (CP) inference (or training) workflow with the PyTorch CUDA caching-allocator history, then attribute the top-N memory peaks to specific modules and lines of code. Wraps the end-to-end forward in…
Profile a context-parallel (CP) inference (or training) workflow with NVIDIA Nsight Systems (nsys). Locates or installs the nsys binary, mocks random features + small Glorot-init weights (or loads a real checkpoint), wraps the end-to-end forward in a torchrun…
Implement distributed data-feature sharding for CP: assign DTensor placements to every model feature, build the placement-definition dictionary, implement atom-feature pack/pad/scatter, per-shard and cross-axis divisibility padding, the DTensor Dataset /…
Write and run multi-rank pytest parity tests that prove a CP implementation is numerically equivalent to its serial reference, using mp.spawn / spawn_multiprocessing as in the Boltz-CP test framework. Covers unit, layer-integration, module-integration, and…