del model deadlocks on ROCm 7.2.1 + gfx1100 (Windows) — HIP allocator free path

Environment

Symptom

del on the CTranslate2 Whisper model object hangs indefinitely (no CPU/GPU activity, no error, no timeout). The hang occurs during Python reference counting cleanup when the C++ destructor chain runs.

The model loads and transcribes correctly — only teardown hangs.

Minimal Reproduction

import ctranslate2

model = ctranslate2.models.Whisper("large-v3", device="cuda", compute_type="float16")
# ... optional: run some inference ...
del model  # HANGS here, never returns

Verified that torch.cuda.empty_cache() (PyTorch's HIP cache cleanup) works fine on the same environment — only CTranslate2's model destructor deadlocks.

Trace

The destructor chain:

Python: del model
  → WhisperModel C++ destructor
    → std::unique_ptr<Layer> members destroyed
      → StorageView::~StorageView() → release() → _allocator->free(data, device)
        → CubCachingAllocator::free() → hipcub::CachingDeviceAllocator::DeviceFree()
          OR
          CudaAsyncAllocator::free() → hipFreeAsync(ptr, stream)

The hang is in one of these HIP allocator free paths (likely hipcub::DeviceFree or hipFreeAsync), where the ROCm runtime's internal stream synchronization deadlocks on gfx1100.

Workaround

Skip explicit del model in application code — let Python's GC handle cleanup when the function returns. The 4-5 GB model footprint is acceptable on modern GPUs alongside other models (e.g., Wav2Vec2 for alignment).

# DON'T: del model          # deadlocks on ROCm
# DO:    torch.cuda.empty_cache()  # works fine, frees PyTorch cache
#        # let model be GC'd at scope exit

Additional Context

• torch.cuda.empty_cache() (PyTorch HIP backend) works correctly on the same GPU — no deadlock
• del on PyTorch models (e.g., Wav2Vec2) works correctly — only CTranslate2's destructor triggers the issue
• The same CTranslate2 version on NVIDIA CUDA does not exhibit this behavior

Start with the minimal Python reproduction and trace the destructor path through StorageView::release(), CubCachingAllocator::free(), and CudaAsyncAllocator::free(). Compare the hipcub::DeviceFree and hipFreeAsync paths on ROCm 7.2.1 with gfx1100, then verify that model teardown completes without hanging while preserving the reported CUDA behavior.