pyisolate 0.10.3rc1: per-channel tensor transport + modern-Python event loop

[pollockjj]

Summary

Single squashed commit off develop carrying the two 0.10.3rc1 fixes plus the version bump. Both bugs are pre-existing in released 0.10.2 and were reproduced against the released wheel.

Per-channel tensor transport (fixes KeyError('data') in torch-free sealed/conda workers)

Serialize torch tensors before the registry's mode-bound Tensor serializer in both serialization pre-passes (_prepare_for_rpc_impl, _serialize_for_isolation_impl), so a shared-memory TensorRef is never emitted onto a JSON channel. Previously, a co-resident shared_memory (share_torch) extension pinned the process-global serializer, and a torch-free sealed/conda worker decoding the resulting TensorRef on its JSON channel crashed on the missing "data" key. CUDA-IPC routing is evaluated at call time so configured CUDA tensors are not downgraded to CPU.

Modern-Python event loop (fixes RuntimeError on Python >=3.12)

AsyncRPC.__init__ acquires the event loop without raising when constructed outside a running loop: prefer the running loop, then the thread's installed loop, and create one only as a last resort. Previously the eager asyncio.get_event_loop() raised RuntimeError("There is no current event loop") on Python >=3.12 in the synchronous host launch path.

Tests / version

• Regression tests for both fixes.
• Version bumped to 0.10.3rc1 in pyproject.toml and pyisolate/__init__.py.

Test plan

• pytest tests/test_rpc_contract.py tests/test_tensor_transport_mode_isolation.py tests/test_event_channel.py — green.
• Both bugs reproduced on the released pyisolate==0.10.2 wheel and confirmed fixed here.

[Copilot]

Pull request overview

This PR prepares the 0.10.3rc1 release by fixing two pre-existing issues in 0.10.2: (1) tensor serialization being incorrectly pre-encoded via the process-global registry instead of deferring to the per-channel transport, and (2) AsyncRPC failing to construct on Python 3.12+ when no current event loop exists (sync host launch path). It also adds regression tests and bumps the version.

Changes:

• Defer torch tensor encoding in the RPC/host serialization pre-passes so the selected transport (JSON vs shared-memory/CUDA IPC) decides the wire format.
• Make AsyncRPC.__init__ acquire an event loop safely across “running”, “installed”, and “none installed” scenarios (Python 3.12+ compatible).
• Add regression tests for both fixes and bump version to 0.10.3rc1.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
pyisolate/_internal/rpc_serialization.py	Defers tensor handling ahead of registry serializer to preserve per-channel tensor transport selection.
pyisolate/_internal/model_serialization.py	Defers tensor handling ahead of registry serializer and reads CUDA-IPC env at call time (vs import time).
pyisolate/_internal/rpc_protocol.py	Updates AsyncRPC loop acquisition to avoid Python 3.12+ get_event_loop() failure when no loop is installed.
tests/test_tensor_transport_mode_isolation.py	New regression coverage for per-channel tensor transport and runtime CUDA-IPC env evaluation.
tests/test_tensor_shared_memory.py	New regression coverage for CPU shared-memory tensor “zero-copy” semantics.
tests/test_rpc_contract.py	New regression coverage for AsyncRPC construction behavior across event-loop states.
tests/test_event_channel.py	Switches tests to asyncio.run() to avoid reliance on deprecated/changed get_event_loop() behavior.
pyproject.toml	Version bump to 0.10.3rc1.
pyisolate/__init__.py	Version bump to 0.10.3rc1.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.