Skip to content

omnipkg v3.4.1 - IPC telemetry, Native PyTorch shared memory, and Windows ARM64

Latest
Latest

Choose a tag to compare

View all tags
@1minds3t 1minds3t released this 09 Jun 02:45
v3.4.1
0a5ed3d
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

IPC telemetry, Native PyTorch shared memory, and Windows ARM64

Welcome to omnipkg v3.4.1! This release brings deep telemetry to the daemon IPC layer, implements native PyTorch shared memory reconstructions, and paves the way for Windows ARM64 support.

  • Native PyTorch IPC Outputs: Re-engineered the CUDA fallback paths to properly extract and reconstruct TypedStorage representations from PyTorch 1.x IPC handles, allowing massive GPU tensor outputs to jump boundaries with zero extra copies.

  • Multi-Package CLI Swaps: Upgraded 8pkg swap to handle multiple packages at once with native passthrough of standard pip flags (--no-deps, --pre, --no-cache-dir).

  • IPC Dispatch Telemetry: Added the DispatchPerfLogger (OMNIPKG_DEBUG=1) to emit microsecond-resolution JSONL telemetry. Track exact encode, flush, queue-wait, and decode latencies across the client ↔ daemon boundary.

  • Warm-Worker Transport Benchmarks: Introduced benchmark_transport.py to directly compare omnipkg's managed SHM vs raw POSIX SHM, msgpack, and pickle.

  • Faster Idle Reaping: Lowered the daemon idle eviction interval from 60s to 5s, resolving stall conditions during extreme stress tests.

  • Windows ARM64 CI: Configured Azure pipelines to officially cross-compile the C dispatcher to Windows ARM64 (win_arm64) on conda-forge.

  • Documentation: Massively overhauled the README and conda-forge descriptions to center on our sub-millisecond C-dispatcher metrics, "God Mode" GPU IPC, and Rust FFI atomic swaps.

📝 Code Changes:

  • UPDATE: src/omnipkg/cli.py (103 lines changed)
  • UPDATE: src/omnipkg/dispatcher.py (2 lines changed)
  • UPDATE: src/omnipkg/isolation/worker_daemon.py (171 lines changed)
  • UPDATE: src/omnipkg/loader.py (60 lines changed)
  • UPDATE: src/omnipkg/package_meta_builder.py (2 lines changed)

🧪 Tests:

  • NEW: src/omnipkg/tests/benchmark_transport.py (1166 lines)
  • UPDATE: src/omnipkg/tests/test_loader_stress_test.py (127 lines)

📚 Documentation:

  • README.md (1635 lines)

⚙️ Configuration:

  • .github/dependabot.yml (9 lines)
  • pyproject.toml (3 lines)
  • src/omnipkg/conda-recipes/conda-forge-platforms.yml
  • src/omnipkg/conda-recipes/meta-noarch.yaml (35 lines)
  • src/omnipkg/conda-recipes/meta-platforms.yaml (29 lines)

Additional Changes:

  • perf: add IPC dispatch telemetry and transport benchmarks
  • ci(win): attempt to build win arm 64 on conda-forge
  • docs: improve conda forge descriptions
  • docs: revamp README with hardware benchmarks and fix release packaging
  • feat(isolation): implement native PyTorch IPC output reconstruction and multi-package cli swaps

19 files changed, 1786 insertions(+), 1593 deletions(-)

Assets 10
Loading