Skip to content

v3.1.5

  • v3.1.5
  • 6788178
  • Verified

    This commit was created on GitHub.com and signed with GitHub’s verified signature.
  • Choose a tag to compare

  • v3.1.5
  • 6788178
  • Choose a tag to compare

  • Verified

    This commit was created on GitHub.com and signed with GitHub’s verified signature.
@eason1981 eason1981 tagged this 17 Nov 13:11
* feat: low-memory defaults below 500 GB of host memory

A host with less than 500 GB of memory, or a container capped below it,
now proves on the GPU with 2 replay workers, an event pool of 4 history
slots, 4 emulator memories and at most 8 waiting deferred records. Above
the line, and on CPU-only builds, the defaults are unchanged.

On one RTX 5090, ere's test_prove_zkvm_interface peaked at 275 GiB with
the old defaults (353 deferred records waited in Bls12381G1Msm, and each
of the 8 replay workers keeps ~8 GiB resident). With these defaults the
heavy cases stay near 50-65 GiB, at the same speed: the GPU, not replay,
bounds those runs.

PICO_HOST_MEM=low|high picks the tier instead of the detected memory; an
embedded prove refuses any other value up front. MLE_DEFERRED_RECORDS,
MEMORY_POOL_SIZE, an explicit event pool size and explicit emulation
workers still win over either tier. Proofs and keys are unchanged.

* perf: prove on the riscv keys the prover already made

Prover::build keygens the program's riscv machine, and the embedded GPU
run keygenned the same program again for card 0 on every new ELF. The
SDK now hands its machine to the run (EmbeddedProgram::riscv), and card 0
adopts it when it keeps the GCS seed and holds no chunk terms; it is
seeded there as RiscvMachine::new would. Other cards and a two-machine
secondary keygen as before.

The batch Phase::Riscv prove on the same machine drops its chunk GCS
terms afterwards, so a later streaming run's drain never sees them.

* perf: keygen only the program table for a new program

Only the riscv program table depends on the ELF. A riscv keygen now keeps
every other AIR's committed preprocessed table, and the pre-hash of an
unchanged preimage, for the next keygen of the same machine in the
process, so a new program generates and commits its program table alone.
The keys are byte-identical to a fresh keygen (tested across three
programs).

* perf: reuse the riscv AIRs' GPU constraint data across programs

Every new program builds a new riscv proving key, and transporting it rebuilt each AIR's
GPU constraint data (rule scheduling, monomial expansion, uploads): about 1.6 s per program.
Only the Program table's preprocessed data depends on the program, so the riscv key now
serves that data from the per-device cache the recursion keys already use, keyed on the
constraint DAG. The cache's default cap rises from 200 to 1024 entries to hold a riscv set
per card next to the recursion set. Proofs and keys are unchanged.

* perf: bump bls12_381 to the fork commit with G1 MSM and subgroup checks on the precompiles

brevis-network/bls12_381#4 runs G1 multi-scalar multiplication and the G1/G2
subgroup checks on the BLS12-381 add/double precompiles inside the zkVM.
Assets 2
Loading