Skip to content

Verathos v0.1.35 — Concurrent Proof Serving

Choose a tag to compare

View all tags
@Keplerteron Keplerteron released this 11 Aug 08:15
· 36 commits to main since this release
v0.1.35
4fade0a

This release improves inference reliability while Gleipnir hard audits run
under mixed production traffic. Proof semantics, signed tolerances and the
wire format remain unchanged.

Highlights

  • Keeps already-admitted inference running when a hard-audit challenge arrives
    and rejects only new post-challenge admissions with a retryable response.
  • Makes authenticated replay deterministic under concurrent requests and
    retained prefix-cache hits, including hybrid GDN/attention execution.
  • Cancels disconnected streaming requests at the inference engine and promptly
    releases their admission and KV-cache state.
  • Routes with validator-observed performance and live engine/KV availability,
    with bounded failover only before the first response token.
  • Applies the validator-derived hard-proof deadline consistently across
    admission, replay, proving and transport.
  • Assigns distinct GPUs when adding multiple endpoints through the setup
    wizard and preserves managed hard-proof proxy timeouts for every endpoint.
  • Serializes shared-checkout installations and gives each endpoint its own
    deterministic restart slot during automatic miner updates.
  • Warns miners when a rate-limited public RPC is selected and recommends an
    operator-owned Subtensor node for production use.

Updating

  • Miners with auto-update enabled update automatically to v0.1.9.
  • Validators and proxies with auto-update enabled update automatically to
    v0.1.33.
  • No new runtime flags, proof artifacts or model requalification are required.

Validation

  • Focused release and auto-update suite: 702/702 passed.
  • Exact-origin packaging: 299/299 Python parses, 12/12 wheel checks,
    117 protected extensions and zero missing shared libraries.
  • FP8 concurrent hard-audit gate: 8/8 passed across independent nonce
    selections.
  • Mixed-load gate: four admitted light requests completed beside a passing hard
    proof; post-challenge admission returned a retryable response and serving
    recovered immediately afterward.
  • Current-build FP8, Qwen AWQ and Huihui AWQ-MTP hard and mixed-load regression
    gates passed without a proof-relation, tolerance or wire-format change.

Full Changelog: v0.1.34...v0.1.35

Assets 2
Loading