QVAC-18580 chore[skiplog]: release sdk 0.10.2 by simon-iribarren · Pull Request #1939 · tetherto/qvac

simon-iribarren · 2026-05-07T10:56:37Z

Note: be concise and prefer bullet points.

🎯 What problem does this PR solve?

Cuts the v0.10.2 hotfix of @qvac/sdk on the release-sdk-0.10.2 branch. Brings a single high-priority fix from main to the 0.10.x line — the delegated-inference connection regression that was blocking the Workbench 0.7.0 release.

#1934 — fix: drop blocking dht.fullyBootstrapped() wait from delegated connect (loadModel.delegation.connection ~3.2× faster: 3.82s → 1.18s mean, restoring v0.9.0-class latency).

📝 How does it solve it?

Cherry-picks #1934 (fix, QVAC-18580) from upstream/main onto release-sdk-0.10.1 (5eb6b52e5).
Bumps packages/sdk/package.json: 0.10.1 → 0.10.2.
Adds packages/sdk/changelog/0.10.2/{CHANGELOG.md, CHANGELOG_LLM.md}.
Rebuilds the root packages/sdk/CHANGELOG.md aggregate with v0.10.2 at the top.
packages/sdk/NOTICE is unchanged — no dependency changes since v0.10.1, so the existing attribution list is byte-identical and stays accurate.

🧪 How was it tested?

The underlying fix (#1934) was benchmarked locally before merge (10 consumer↔provider runs each, against the published v0.10.1 build):

Build	Mean conn. time	p50	p95
v0.10.1 (baseline)	3.82s	3.71s	4.94s
v0.10.2 (this fix)	1.18s	1.12s	1.49s

Failure-path tests (delegated-connection-failure, delegated-provider-not-found, delegated-load-model-fallback-local) were re-run against the patched SDK and pass — PEER_NOT_FOUND for non-existent peers and connection timeouts for unreachable ones still fire on the same code path. Full verification notes are on PR #1934.

This release branch carries only the cherry-pick + release metadata (4 files: package.json, root CHANGELOG.md, two changelog/0.10.2/ files). No other functional changes from main are included.

Once merged → triggers GPR publish for @tetherto/sdk@0.10.2. A backmerge PR (release-sdk-0.10.2 → main) will follow to publish to npm and keep main aligned.

…ated connect (tetherto#1934) The `await swarm.dht.fullyBootstrapped()` block in `ensureRPCConnection` (introduced in tetherto#1729) serialized 1-3s of cold-start time onto every first delegated call, regressing `loadModel.delegation.connection` vs 0.9.0. Local bench (5 cold consumer runs against the same long-running provider, model already in cache): - Baseline (current main): mean = 3.82s, p50 = 3.91s - This patch: mean = 1.18s, p50 = 1.15s (-2.64s, ≈3.2× faster, all runs succeeded) WB team reported the same regression as 2.46s → 8.34s on their environment when bumping @qvac/sdk 0.9.0 → 0.10.0; this fix is expected to land them in the 2-3s range, back inside the 0.9.0 baseline. `getSwarm()` is still invoked for its side-effect (lazy swarm creation on the first call). `dht.connect()` itself bootstraps on demand if the routing table isn't warm enough, so removing the explicit await loses no correctness — the explicit guard was conservative scaffolding from tetherto#1729 that never actually mattered on the SDK's normal init path, where `getSwarm()` is called early during registry/runtime initialisation. Refs QVAC-18144. (cherry picked from commit 318f0c6)

- Bump packages/sdk/package.json: 0.10.1 -> 0.10.2. - Add packages/sdk/changelog/0.10.2/{CHANGELOG.md, CHANGELOG_LLM.md}. - Rebuild root packages/sdk/CHANGELOG.md aggregate with v0.10.2 at top. Hotfix release for the delegated-inference connection regression introduced in v0.10.0 (tetherto#1934). NOTICE file unchanged — no dependency changes since v0.10.1.

github-actions · 2026-05-07T10:57:10Z