stormlog.cuda_native_debug

CUDA-native allocator history capture and attribution helpers.

Functions

build_cuda_tensor_attribution_index([...])	Build a best-effort index from CUDA storage pointers to live tensors.
build_snapshot_allocation_attribution(...)	Cross-reference allocator addresses against live tensor storage pointers.
capture_cuda_snapshot_artifacts(output_dir, *)	Capture the current CUDA allocator snapshot and write debug artifacts.
cuda_memory_history([device, ...])	Context manager that records CUDA allocator history for a block.
cuda_memory_history_supported()	Return whether the current PyTorch runtime exposes CUDA history APIs.
start_cuda_memory_history([device, ...])	Enable CUDA allocator history recording for the selected device.
stop_cuda_memory_history([device])	Disable CUDA allocator history recording for the selected device.
write_cuda_snapshot_artifacts(output_dir, ...)	Write snapshot, attribution, and best-effort visualization artifacts.

stormlog.cuda_native_debug.build_cuda_tensor_attribution_index(device=None, *, skip_gc=False)

Build a best-effort index from CUDA storage pointers to live tensors.

Parameters:

• device (int | torch.device | None)

• skip_gc (bool)

Return type:

dict[str, Any]

stormlog.cuda_native_debug.build_snapshot_allocation_attribution(snapshot, tensor_index)

Cross-reference allocator addresses against live tensor storage pointers.

Parameters:

• snapshot (Any)

• tensor_index (dict[str, Any])

Return type:

dict[str, Any]

stormlog.cuda_native_debug.capture_cuda_snapshot_artifacts(output_dir, *, device=None, history_recorded)

Capture the current CUDA allocator snapshot and write debug artifacts.

Parameters:

• output_dir (Path)

• device (int | torch.device | None)

• history_recorded (bool)

Return type:

list[str]

stormlog.cuda_native_debug.cuda_memory_history(device=None, trace_alloc_max_entries=100000)

Context manager that records CUDA allocator history for a block.

Parameters:

• device (int | torch.device | None)

• trace_alloc_max_entries (int)

Return type:

Iterator[None]

stormlog.cuda_native_debug.cuda_memory_history_supported()

Return whether the current PyTorch runtime exposes CUDA history APIs.

Return type:

bool

stormlog.cuda_native_debug.start_cuda_memory_history(device=None, trace_alloc_max_entries=100000)

Enable CUDA allocator history recording for the selected device.

Parameters:

• device (int | torch.device | None)

• trace_alloc_max_entries (int)

Return type:

None

stormlog.cuda_native_debug.stop_cuda_memory_history(device=None)

Disable CUDA allocator history recording for the selected device.

Parameters:

device (int | torch.device | None)

Return type:

None

stormlog.cuda_native_debug.write_cuda_snapshot_artifacts(output_dir, snapshot, tensor_index, *, history_recorded, device=None)

Write snapshot, attribution, and best-effort visualization artifacts.

Parameters:

• output_dir (Path)

• snapshot (Any)

• tensor_index (dict[str, Any])

• history_recorded (bool)

• device (int | torch.device | None)

Return type:

list[str]