Skip to content

v0.1.33 - Maximum-Concurrency Proof Serving

Choose a tag to compare

@Keplerteron Keplerteron released this 10 Aug 05:42

Highlights

  • Keeps Gleipnir proof capture and nonce-free light precommits valid at the miner's advertised vLLM concurrency.
  • Separates concurrency-scaled response capture from serialized hard-proof replay storage.
  • Resolves each retained request through vLLM's canonical scheduler slot across CUDA-graph batch sizes, preventing stale or incorrect K/V replay.
  • Bounds HTTP admission to the engine's actual scheduler capacity and preserves a dedicated completed-proof retry slot.
  • Prevents validator event-loop congestion from making already-arrived precommits appear late.
  • Adds fail-closed startup validation between registered projection operations and authenticated static-weight commitments.
  • Does not change proof arithmetic, signed artifacts, tolerances, wire format, canary policy, or hard-audit coverage.

Versions

  • Miner: 0.1.7
  • Validator/proxy: 0.1.32
  • Spec: 0.1.32

Updating

Miners and validators running with auto-update enabled will fetch this release and restart automatically at a safe idle window. Fresh installations receive the corrected runtime directly through the official setup process.

No new configuration flags or manual artifact changes are required.

Validation

  • Role-version and auto-update suite: 42/42 passed.
  • Focused concurrency, capture, routing, transport, and verifier checks: 361/361 passed.
  • Final release recheck: 122/122 passed.
  • AWQ maximum-concurrency test: 63 light requests + 1 hard audit passed, 22.99 s proving, 2.17 s verification, 7.07 MiB wire, 1.12 GiB GPU-memory delta.
  • FP8 maximum-concurrency test: 63 light requests + 1 hard audit passed, 16.33 s proving, 2.32 s verification, 6.92 MiB wire, 780 MiB GPU-memory delta.
  • Exact-origin packaging audit: 511 tracked files, 299/299 Python parses, 12/12 wheel checks, and 117/117 protected shared objects.

Full Changelog: v0.1.32...v0.1.33