Add vmem chunked allocator#516

Draft

mawad-amd wants to merge 24 commits into

mainfrom

muhaawad/vmem-chunked-allocator

Commits on May 12, 2026

Add chunked VMem allocator with power-of-two free lists

authored and
artulab
committed
Apply Ruff auto-fixes

github-actions[bot]
authored and
artulab
committed
Fix free_list_size_classes test: use granularity-adaptive sizes

authored and
artulab
committed
Fix import_external_tensor: tolerate hipMemSetAccess failure on DMA-BUF imports

authored and
artulab
committed
Fix import chunks: separate from shareable chunks for peer exchange

authored and
artulab
committed
Fix VA space exhaustion: reduce default VA size, add cleanup

authored and
artulab
committed
Apply Ruff auto-fixes

github-actions[bot]
authored and
artulab
committed
Fix chunk_size and VA sizing for small heaps

authored and
artulab
committed
Fix hipErrorUnknown from async GPU ops during VMem close

authored and
artulab
committed
Add GPU sync to SymmetricHeap.close() before freeing VA ranges

authored and
artulab
committed
Apply Ruff auto-fixes

github-actions[bot]
authored and
artulab
committed
Fix hipErrorUnknown: unmap peer-imported chunks before freeing VA ranges

authored and
artulab
committed
Fix import_external_tensor: use External Memory API for hipMalloc tensors

authored and
artulab
committed
Apply Ruff auto-fixes

github-actions[bot]
authored and
artulab
committed
Fix test_chunked_stats: improve GC reliability for weakref finalizer

authored and
artulab
committed
Remove non-deterministic GC assertion from test_chunked_stats

authored and
artulab
committed
Increase default VA reservation to 128 GiB

authored and
artulab
committed
Document ROCm 7.1+ requirement for hipMemImportFromShareableHandle

authored and
artulab
committed
Document ROCm 7.1+ requirement for hipMemImportFromShareableHandle

authored and
artulab
committed
Fixing import issues on vmem_chunked_allocator.py.
artulab
committed
Refactor memory drivers to decouple virtual address reservation
artulab
committed
Apply Ruff auto-fixes

github-actions[bot]
authored and
artulab
committed
feat(drivers): implement VMM local drivers, factory, and fabric VA support
artulab
committed
Refactor SymmetricHeap and VMemChunkedAllocator to use topology-aware driver-based peer setup for chunked VMM heaps, with separate local FD and fabric-handle refresh paths
artulab
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add vmem chunked allocator#516

Add vmem chunked allocator#516
mawad-amd wants to merge 24 commits into
mainfrom
muhaawad/vmem-chunked-allocator

Commits on May 12, 2026

Add chunked VMem allocator with power-of-two free lists

Apply Ruff auto-fixes

Fix free_list_size_classes test: use granularity-adaptive sizes

Fix import_external_tensor: tolerate hipMemSetAccess failure on DMA-BUF imports

Fix import chunks: separate from shareable chunks for peer exchange

Fix VA space exhaustion: reduce default VA size, add cleanup

Apply Ruff auto-fixes

Fix chunk_size and VA sizing for small heaps

Fix hipErrorUnknown from async GPU ops during VMem close

Add GPU sync to SymmetricHeap.close() before freeing VA ranges

Apply Ruff auto-fixes

Fix hipErrorUnknown: unmap peer-imported chunks before freeing VA ranges

Fix import_external_tensor: use External Memory API for hipMalloc tensors

Apply Ruff auto-fixes

Fix test_chunked_stats: improve GC reliability for weakref finalizer

Remove non-deterministic GC assertion from test_chunked_stats

Increase default VA reservation to 128 GiB

Document ROCm 7.1+ requirement for hipMemImportFromShareableHandle

Document ROCm 7.1+ requirement for hipMemImportFromShareableHandle

Fixing import issues on vmem_chunked_allocator.py.

Refactor memory drivers to decouple virtual address reservation

Apply Ruff auto-fixes

feat(drivers): implement VMM local drivers, factory, and fabric VA support

Refactor SymmetricHeap and VMemChunkedAllocator to use topology-aware driver-based peer setup for chunked VMM heaps, with separate local FD and fabric-handle refresh paths