Description

closes #479

Add peer access control for DeviceMemoryResource

This PR implements peer access control for DeviceMemoryResource, allowing memory pool allocations to be accessed from devices other than the owner device.

Overview

Memory pools created by DeviceMemoryResource can now grant peer access permissions to other GPUs, enabling multi-GPU workflows where allocations on one device need to be accessed by kernels running on another device.

Key Changes

1. New peer_accessible_by property

• Get/set which devices can access allocations from a memory pool
• Returns a tuple of sorted device IDs that currently have peer access
• Setter accepts a sequence of Device objects or device IDs
• Automatically excludes the owner device from the peer access list
• Uses incremental updates (only changes permissions for devices being added/removed)

Example usage:

dmr = DeviceMemoryResource(0)
dmr.peer_accessible_by = [1, 2]  # Grant access to devices 1 and 2
assert dmr.peer_accessible_by == (1, 2)
dmr.peer_accessible_by = []      # Revoke all peer access

2. Implementation details

• Uses cuMemPoolSetAccess to modify peer access permissions at runtime
• Properly allocates C arrays for passing CUmemAccessDesc structures to the driver API
• Memory-safe with try/finally blocks ensuring allocated memory is freed
• Supports both granting (CU_MEM_ACCESS_FLAGS_PROT_READWRITE) and revoking (CU_MEM_ACCESS_FLAGS_PROT_NONE) access in a single transaction

3. Debugging support

• Added DMR_mempool_get_access() helper function to probe actual access permissions via cuMemPoolGetAccess
• Useful for debugging and verifying expected access states

4. Workaround for driver bug (nvbug 5698116)

• Added cleanup in __dealloc__ to reset peer access before destroying memory pools
• Works around an issue where recycled memory pool handles inherit peer access state from previous handles

Technical Notes

• Required adding malloc/free imports from libc.stdlib for C array allocation
• The implementation properly handles the conversion from Python collections to C structures required by the low-level CUDA driver API
• No trailing whitespace in source files (adheres to project code style)

Testing

This feature is designed to work with the existing multi-GPU test suite and enables peer access scenarios for buffer copy operations between devices.