在使用支持的cuda接口时，需要将API名称中的cuda变换为NPU形式才能使用：torch.cuda.***变换为torch_npu.npu.***。举例如下。
torch.cuda.current_device --> torch_npu.npu.current_device
API名称 |
NPU形式名称 |
是否支持 |
限制与说明 |
|---|---|---|---|
torch.cuda.StreamContext |
否 |
||
torch.cuda.can_device_access_peer |
否 |
||
torch.cuda.current_blas_handle |
torch_npu.npu.current_blas_handle |
否 |
|
torch.cuda.current_device |
torch_npu.npu.current_device |
是 |
|
torch.cuda.DoubleTensor |
torch_npu.npu.DoubleTensor |
||
torch.cuda.ShortTensor |
torch_npu.npu.ShortTensor |
||
torch.cuda.CharTenso |
torch_npu.npu.CharTensor |
||
torch.cuda.ByteTensor |
torch_npu.npu.ByteTensor |
||
torch.cuda.current_stream |
torch_npu.npu.current_stream |
||
torch.cuda.default_stream |
torch_npu.npu.default_stream |
是 |
|
torch.cuda.device |
torch_npu.npu.device |
是 |
|
torch.cuda.device_count |
torch_npu.npu.device_count |
是 |
|
torch.cuda.device_of |
torch_npu.npu.device_of |
是 |
|
torch.cuda.get_arch_list |
否 |
||
torch.cuda.get_device_capability |
否 |
||
torch.cuda.get_device_name |
torch_npu.npu.get_device_name |
是 |
|
torch.cuda.get_device_properties |
否 |
||
torch.cuda.get_gencode_flags |
否 |
||
torch.cuda.get_sync_debug_mode |
否 |
||
torch.cuda.stream |
torch_npu.npu.stream |
是 |
|
torch.cuda.init |
torch_npu.npu.init |
是 |
|
torch.cuda.ipc_collect |
否 |
||
torch.cuda.is_available |
torch_npu.npu.is_available |
是 |
|
torch.cuda.is_initialized |
torch_npu.npu.is_initialized |
是 |
|
torch.cuda.memory_usage |
torch_npu.npu.memory_usage |
否 |
|
torch.cuda.set_device |
torch_npu.npu.set_device |
是 |
|
torch.cuda.set_stream |
否 |
||
torch.cuda.set_sync_debug_mode |
否 |
||
torch.cuda.synchronize |
torch_npu.npu.synchronize |
是 |
|
torch.cuda.utilization |
torch_npu.npu.utilization |
是 |
|
torch.cuda.get_rng_state |
torch_npu.npu.get_rng_state |
是 |
|
torch.cuda.get_rng_state_all |
torch_npu.npu.get_rng_state_all |
是 |
|
torch.cuda.set_rng_state |
torch_npu.npu.set_rng_state |
是 |
|
torch.cuda.set_rng_state_all |
torch_npu.npu.set_rng_state_all |
是 |
|
torch.cuda.manual_seed |
torch_npu.npu.manual_seed |
是 |
|
torch.cuda.manual_seed_all |
torch_npu.npu.manual_seed_all |
是 |
|
torch.cuda.seed |
torch_npu.npu.seed |
是 |
|
torch.cuda.seed_all |
torch_npu.npu.seed_all |
是 |
|
torch.cuda.initial_seed |
torch_npu.npu.initial_seed |
是 |
|
torch.cuda.comm.broadcast |
否 |
||
torch.cuda.comm.broadcast_coalesced |
否 |
||
torch.cuda.comm.reduce_add |
否 |
||
torch.cuda.comm.scatter |
否 |
||
torch.cuda.comm.gather |
否 |
||
torch.cuda.Stream |
torch_npu.npu.Stream |
是 |
|
torch.cuda.Stream.query |
torch_npu.npu.Stream.query |
||
torch.cuda.Stream.record_event |
torch_npu.npu.Stream.record_event |
是 |
|
torch.cuda.Stream.synchronize |
torch_npu.npu.Stream.synchronize |
是 |
|
torch.cuda.Stream.wait_event |
torch_npu.npu.Stream.wait_event |
||
torch.cuda.Stream.wait_stream |
torch_npu.npu.Stream.wait_stream |
是 |
|
torch.cuda.ExternalStream |
否 |
||
torch.cuda.ExternalStream.query |
否 |
||
torch.cuda.ExternalStream.record_event |
否 |
||
torch.cuda.ExternalStream.synchronize |
否 |
||
torch.cuda.ExternalStream.wait_event |
否 |
||
torch.cuda.ExternalStream.wait_stream |
否 |
||
torch.cuda.Event |
torch_npu.npu.Event |
是 |
|
torch.cuda.Event.elapsed_time |
torch_npu.npu.Event.elapsed_time |
是 |
|
torch.cuda.Event.from_ipc_handle |
否 |
||
torch.cuda.Event.ipc_handle |
否 |
||
torch.cuda.Event.query |
torch_npu.npu.Event.query |
是 |
|
torch.cuda.Event.record |
torch_npu.npu.Event.record |
是 |
|
torch.cuda.Event.synchronize |
torch_npu.npu.Event.synchronize |
||
torch.cuda.Event.wait |
torch_npu.npu.Event.wait |
是 |
|
torch.cuda.graph_pool_handle |
否 |
||
torch.cuda.CUDAGraph |
否 |
||
torch.cuda.CUDAGraph.capture_begin |
否 |
||
torch.cuda.CUDAGraph.capture_end |
否 |
||
torch.cuda.CUDAGraph.pool |
否 |
||
torch.cuda.CUDAGraph.replay |
否 |
||
torch.cuda.CUDAGraph.reset |
否 |
||
torch.cuda.graph |
否 |
||
torch.cuda.make_graphed_callables |
否 |
||
torch.cuda.empty_cache |
torch_npu.npu.empty_cache |
是 |
|
torch.cuda.list_gpu_processes |
否 |
||
torch.cuda.mem_get_info |
否 |
||
torch.cuda.memory_stats |
torch_npu.npu.memory_stats |
是 |
|
torch.cuda.memory_stats_as_nested_dict |
torch_npu.npu.memory_stats_as_nested_dict |
||
torch.cuda..reset_accumulated_memory_stats |
torch_npu.npu.reset_accumulated_memory_stats |
||
torch.cuda.memory_summary |
torch_npu.npu.memory_summary |
是 |
|
torch.cuda.memory_snapshot |
torch_npu.npu.memory_snapshot |
是 |
|
torch.cuda.memory_allocated |
torch_npu.npu.memory_allocated |
是 |
|
torch.cuda.max_memory_allocated |
torch_npu.npu.max_memory_allocated |
是 |
|
torch.cuda.reset_max_memory_allocated |
torch_npu.npu.reset_max_memory_allocated |
是 |
|
torch.cuda.memory_reserved |
torch_npu.npu.memory_reserved |
是 |
|
torch.cuda.max_memory_reserved |
torch_npu.npu.max_memory_reserved |
是 |
|
torch.cuda.set_per_process_memory_fraction |
torch_npu.npu.set_per_process_memory_fraction |
是 |
|
torch.cuda.memory_cached |
torch_npu.npu.memory_cached |
是 |
|
torch.cuda.max_memory_cached |
torch_npu.npu.max_memory_cached |
是 |
|
torch.cuda.reset_max_memory_cached |
torch_npu.npu.reset_max_memory_cached |
是 |
|
torch.cuda.reset_peak_memory_stats |
torch_npu.npu.reset_peak_memory_stats |
是 |
|
torch.cuda.caching_allocator_alloc |
torch_npu.npu.caching_allocator_alloc |
是 |
|
torch.cuda.caching_allocator_delete |
torch_npu.npu.caching_allocator_delete |
是 |
|
torch.cuda.nvtx.mark |
否 |
||
torch.cuda.nvtx.range_push |
torch_npu.npu.nvtx.range_push |
否 |
|
torch.cuda.nvtx.range_pop |
否 |