v0.10.5

Provider v0.10.5 (Swift CLI)

Source: d3bbcd0
Build and SDK qualification: https://github.com/Layr-Labs/d-inference/actions/runs/41094173251
Signing: https://github.com/Layr-Labs/d-inference/actions/runs/41094173251
Binary SHA-256: f568a3de578664bbc4246a3e51688aa9595fe51b7aaac8e26bbd829b8733b85d
CodeDirectory SHA-256: 52ab8027450771e6d9312630b1ccbbb87a81be3de8012774d7fa7d62cfb4a841
Bundle SHA-256: c614a0109befb211bbab8c215a1b953576ab52c57d7170937d1e5f5567d49cf1
Metallib SHA-256: 98bab9d77dc5078ad8dd8f1fd17e8f3581415193506e4791a3c4f1ebaab614be
Signed and notarized: yes
macOS deployment target: 14.0 (build, tests and signed runtime smoke: macOS 27)

Install

curl -fsSL https://api.darkbloom.dev/install.sh | bash

What's Changed

• Repair master after #1625: remove console-ui hooks, CI lane and docs citations by @jonathan308 in #1629
• test(provider): stop the unloaded-deletion epoch test hanging the SSD suites by @javiech1 in #1653
• fix(provider): stop Qwen 3.5/3.6 looping to max_tokens on omitted sampling by @javiech1 in #1652
• test(provider): cut ~75 s of serial test time under the 900 s CI watchdog by @jonathan308 in #1644
• feat(provider): darkbloom sandbox — set up, check, pause and remove Darkbloom Sandbox CI jobs by @Gajesh2007 in #1655
• fix(provider): settle model downloads cancelled before waiter attachment by @jonathan308 in #1640
• fix(provider): drop chunk frames processed while disconnected by @jonathan308 in #1646
• fix(provider): preserve prompt-cache hits during batch lookup by @jonathan308 in #1639
• ci: master release lane publishes staging only, no dev release by @davidtai in #1637
• fix(provider): retire SSD cache sweeps in short slices (lookup stalls ~1 s → <28 ms) by @jonathan308 in #1651
• perf(provider): reread only new or changed SSD cache headers in root maintenance (11× faster passes) by @jonathan308 in #1642
• perf(provider): charge the SSD write ledger once per attention donation (~5× faster donations) by @jonathan308 in #1647
• perf(provider): keep a whole SSD read buffer instead of copying it (−6.6% read+decrypt) by @jonathan308 in #1649
• fix(provider): stop counting expired checkpoints as corrupt at startup by @jonathan308 in #1648
• Release qualification names production defaults (master copy of dev PR 1658) by @davidtai in #1659
• perf(provider): overlap uncached weight reads with SHA-256 (1.6× faster full hashes) by @jonathan308 in #1638
• GPT-OSS: match client stop sequences in the Harmony final message only (fixes reasoning-only empty answers) by @jonathan308 in #1645
• perf(provider): probe the SSD LRU victim first and walk the fallback lazily by @jonathan308 in #1641
• perf(provider): let concurrent restores of one SSD checkpoint read it together (8 × 1 GiB: 2.8 s → 0.7 s) by @jonathan308 in #1657
• provider: reconnect faster after a long session or a 1012 close by @davidtai in #1613
• perf(provider): restore dense Qwen conversations from just below the prompt end (−16.5% follow-up TTFT) by @jonathan308 in #1643
• test(provider): SSD retirement fixture implements retireOwnedEntries(_:unusedSince:) (fixes master's test build) by @Gajesh2007 in #1676
• Prepare provider 0.10.5 by @Gajesh2007 in #1678

Full Changelog: v0.10.4...v0.10.5