Expand description
§Async frees and peak memory
Many logup/zerocheck GPU helpers allocate temporary device buffers and launch kernels. Those
buffers are freed via cudaFreeAsync through the VPMM pool, so frees are only stream-ordered. If
new allocations happen before the current stream is synchronized, physical peak memory can
temporarily include both the old and new buffers even though MemTracker only sees the logical
max. Callers that care about physical peak should ensure a current-stream sync after use (e.g.
by calling to_host_on() on outputs or synchronizing the owning stream).
Structs§
Enums§
- Fold
PleError - Fractional
Sumcheck Error - Interaction
GpuError - Round0
Eval Error - Round0
Prep Error - Unstack
Matrix Error
Functions§
- fractional_
sumcheck_ gpu - GKR fractional sumcheck prover. See
docs/cuda-backend/gkr-prover.md(repo root) for the protocol and implementation details. - make_
synthetic_ leaves - Generate random fractional leaves on device for benchmarking.
- prove_
zerocheck_ and_ logup_ gpu