diff --git a/tests/basic_correctness/test_mem.py b/tests/basic_correctness/test_mem.py index 66ca2df530..c3f66e24ab 100644 --- a/tests/basic_correctness/test_mem.py +++ b/tests/basic_correctness/test_mem.py @@ -18,6 +18,21 @@ from ..utils import create_new_process_for_each_test, requires_fp8 DEVICE_TYPE = current_platform.device_type +def mapped_usage(allocator) -> int: + """Bytes the allocator still has mapped on the device. + + `torch.accelerator.get_memory_info()` is a device-wide `cudaMemGetInfo` / + `hipMemGetInfo` query, so its delta also carries whatever else on the GPU + allocates or frees while we measure. The allocator's own bookkeeping is + process-local and exact. + """ + return sum( + data.handle[1] + for data in allocator.pointer_to_data.values() + if not data.is_asleep + ) + + def _wake_up_with_poisoned_mappings(allocator, byte_value: int = 0xA5) -> None: """Wake discarded allocations with deterministic nonzero contents.""" original_create_and_map = cumem.create_and_map @@ -82,11 +97,13 @@ def test_basic_cumem(): output = x + y + z assert torch.allclose(output, torch.ones_like(output) * 3) - free_bytes = torch.accelerator.get_memory_info()[0] + # Track the allocator's own mapped bytes, not device-wide free memory. + mapped_bytes = mapped_usage(allocator) + assert mapped_bytes >= y.nbytes + z.nbytes allocator.sleep() - free_bytes_after_sleep = torch.accelerator.get_memory_info()[0] - assert free_bytes_after_sleep > free_bytes + assert mapped_usage(allocator) == 0 allocator.wake_up() + assert mapped_usage(allocator) == mapped_bytes # they can be used together output = x + y + z @@ -105,13 +122,14 @@ def test_discard_tags(): with allocator.use_memory_pool("kv_cache"): kv = torch.ones(512, 512, device=DEVICE_TYPE) - free_bytes = torch.accelerator.get_memory_info()[0] + mapped_bytes = mapped_usage(allocator) # Discard kv_cache only — weights should remain valid allocator.discard("kv_cache") - free_bytes_after_discard = torch.accelerator.get_memory_info()[0] - assert free_bytes_after_discard > free_bytes + mapped_bytes_after_discard = mapped_usage(allocator) + assert mapped_bytes - mapped_bytes_after_discard >= kv.nbytes + assert mapped_bytes_after_discard >= weights.nbytes # Weights are still usable assert torch.allclose(weights, torch.ones_like(weights)) @@ -172,11 +190,12 @@ def test_cumem_with_cudagraph(): with torch.cuda.graph(model_graph): y = model(x) - free_bytes = torch.accelerator.get_memory_info()[0] + mapped_bytes = mapped_usage(allocator) + assert mapped_bytes >= weight.nbytes + cache.nbytes allocator.sleep() - free_bytes_after_sleep = torch.accelerator.get_memory_info()[0] - assert free_bytes_after_sleep > free_bytes + assert mapped_usage(allocator) == 0 allocator.wake_up() + assert mapped_usage(allocator) == mapped_bytes # after waking up, the content in the weight tensor # should be restored, but the content in the cache tensor