This release includes 1 breaking change for platform teams planning a safe upgrade.
✓ No known CVEs patched in this version
Topics
+4 more
ReleasePort's take
Moderate signalCUDA v0.3.0 eliminates command‑lane blocking waits and reduces driver memory allocations, while introducing pinned async copy support for multiple memory layouts.
Why it matters: Performance improvements (no blocked command lane, less allocation) boost throughput; the breaking change requiring PinnedHost type for host‑copy APIs mandates code updates before upgrade.
Summary
AI summaryCUDA waits no longer block the command lane and driver paths allocate less, adding pinned async copy support.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Breaking | High |
Old host‑copy signatures now require accepting PinnedHost type. Old host‑copy signatures now require accepting PinnedHost type. Source: llm_adapter@2026-07-15 Confidence: low |
— |
| Feature | Medium |
Adds pinned async copies for flat, pitched, volume, and array memory with overlap and memset benchmarks. Adds pinned async copies for flat, pitched, volume, and array memory with overlap and memset benchmarks. Source: llm_adapter@2026-07-15 Confidence: high |
— |
| Dependency | Low |
Bumps github.com/ebitengine/purego from 0.10.0 to 0.10.1. Bumps github.com/ebitengine/purego from 0.10.0 to 0.10.1. Source: llm_adapter@2026-07-15 Confidence: high |
— |
| Dependency | Low |
Bumps actions/checkout from 6 to 7. Bumps actions/checkout from 6 to 7. Source: llm_adapter@2026-07-15 Confidence: high |
— |
| Performance | Medium |
CUDA waits no longer block the command lane. CUDA waits no longer block the command lane. Source: llm_adapter@2026-07-15 Confidence: high |
— |
| Performance | Medium |
Common driver paths now allocate less memory. Common driver paths now allocate less memory. Source: llm_adapter@2026-07-15 Confidence: high |
— |
| Performance | Medium |
Cuts hot‑path allocations. Cuts hot‑path allocations. Source: llm_adapter@2026-07-15 Confidence: high |
— |
| Bugfix | Medium |
Avoids executor stalls in the driver. Avoids executor stalls in the driver. Source: llm_adapter@2026-07-15 Confidence: high |
— |
| Bugfix | Low |
HostBuffer calls continue to work with updated host‑copy signatures. HostBuffer calls continue to work with updated host‑copy signatures. Source: llm_adapter@2026-07-15 Confidence: low |
— |
| Bugfix | Low |
Recommends using errors.Is or errors.As for NOT_READY type assertions. Recommends using errors.Is or errors.As for NOT_READY type assertions. Source: llm_adapter@2026-07-15 Confidence: low |
— |
Full changelog
CUDA waits no longer block the command lane, and common driver paths allocate less.
This release also adds pinned async copies for flat,
pitched, volume, and array memory, with overlap and memset benchmarks.
HostBuffer calls still work. Exact interfaces using the old host-copy signatures need to accept PinnedHost. Direct NOT_READY type
assertions should use errors.Is or errors.As.
What's Changed
- Bump github.com/ebitengine/purego from 0.10.0 to 0.10.1 by @dependabot[bot] in https://github.com/eitamring/gocudrv/pull/54
- Bump actions/checkout from 6 to 7 by @dependabot[bot] in https://github.com/eitamring/gocudrv/pull/53
- avoid executor stalls by @eitamring in https://github.com/eitamring/gocudrv/pull/60
- cut hot path allocations by @eitamring in https://github.com/eitamring/gocudrv/pull/61
- add pinned async copies by @eitamring in https://github.com/eitamring/gocudrv/pull/62
New Contributors
- @dependabot[bot] made their first contribution in https://github.com/eitamring/gocudrv/pull/54
Full Changelog: https://github.com/eitamring/gocudrv/compare/v0.2.0...v0.3.0
Breaking Changes
- Exact interfaces using old host‑copy signatures must now accept PinnedHost
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About eitamring/gocudrv
All releases →Related context
Related tools
Beta — feedback welcome: [email protected]