o p i@sddlZddlZddlmZddlmZddlmZddlmZddl m Z ddl m Z gZ dd d Zd d Zd dZddZGdddZddZddZddZdddZdS)N)_C_ops)core)check_variable_and_dtype)EagerParamBase) LayerHelper)in_dynamic_or_pir_modecCsltr t||St|dgddtd it}|j|j|jd}|j dd|gid|gid|id|S) aI This OP takes in the Tensor :attr:`x` and change it to output with aclFormat with int value. This API is only used for Ascend NPU. Args: x(Tensor): An input N-D Tensor with data type bool, float16, float32, float64, int32, int64, int16, int8, uint8. format(int): Storage data format of the output in aclFormat, default value is -1. Returns: Tensor: A Tensor with acl storage format on Ascend NPU. Examples: .. code-block:: python >>> # doctest: +REQUIRES(env:NPU) >>> import paddle >>> paddle.device.set_device('npu') >>> x = paddle.ones(shape=[6]) >>> y = paddle.incubate._npu_identity(x, 3) # ACL_FORMAT_NC1HWC0 = 3 >>> print(y.shape) [1, 1, 1, 1, 16] x) boolZint8Zuint8Zint16Zint32Zint64Zfloat16Zfloat32Zfloat64 npu_identity)dtype stop_gradientoutformat)typeZinputsZoutputsattrsN)r ) rrr rrlocalsZ"create_variable_for_type_inferencer r Z append_op)r rhelperrrj/home/app/PaddleOCR-VL/.venv_paddleocr/lib/python3.10/site-packages/paddle/incubate/tensor/manipulation.py _npu_identitys&  rcCsjt|trt|j}t|j|jfi|}|||}||fSt|tjr3t j }|||}||fSdS)zt Helper to create a new destination tensor and call 'func(dst, src)' which is either offload or reload. N) isinstancercopydeepcopy__dict__shaper paddleZTensorreager) src_tensorfuncstate new_paramtaskZ new_varbaserrr_load_reload_implZs      r#cCs8tj}tr tStdd|DrdStS)zf Constructs a new AsyncLoad object. It is used to load/reload data asynchronously on GPU. cs|]}t|VqdSNrZis_compiled_with_custom_device.0devrrr r  z$create_async_load..N)rdeviceget_all_custom_device_typeis_compiled_with_xpur XpuAsyncLoadanyZ AsyncLoad)custom_devicesrrrcreate_async_loadjs r2cCstS)zf Constructs a new AsyncLoad object. It is used to load/reload data asynchronously on XPU. )rr/rrrrcreate_xpu_async_loadzsr3c@s(eZdZdZddZddZddZdS) _NoopAsyncTasku(A dummy Task for sync‐fallback on XPU.cCsdS)NTrselfrrr is_completedz_NoopAsyncTask.is_completedcCdSr%rr5rrrcpu_waitr8z_NoopAsyncTask.cpu_waitcCr9r%rr5rrrxpu_waitr8z_NoopAsyncTask.xpu_waitN)__name__ __module__ __qualname____doc__r7r:r;rrrrr4s  r4cCsto t|do |j}tj}tdd|Do't|do'|j|v}|s,|r@| }tj | t d}|t fSt||jS)a Loads the source tensor into the destination tensor asynchronously. Args: src_tensor (EagerParamBase|paddle.Tensor): The source tensor. async_load (core.AsyncLoad): The AsyncLoad object. Returns: tuple: A tuple containing two elements: - dest_tensor (EagerParamBase|paddle.Tensor): The destination tensor. - task (Task): The task that loads the source tensor into the destination tensor. placecsr$r%r&r'rrrr*r+z async_offload..r@)rr.hasattrr@Z is_xpu_placer,r-r0Zcustom_device_typecpu to_tensornumpyZCPUPlacer4r#Zoffload)r async_loadZ is_xpu_tensorr1Zis_custom_tensorZ host_tensorrrrr async_offloads$    rGcCsNtr!t|dr!|jr!|}tj|tdd}|tfSt ||j S)a Reloads the source tensor into the destination tensor asynchronously. Args: src_tensor (EagerParamBase|paddle.Tensor): The source tensor. async_load (core.AsyncLoad): The AsyncLoad object. Returns: tuple: A tuple containing two elements: - dest_tensor (EagerParamBase|paddle.Tensor): The destination tensor. - task (Task): The task that reloads the source tensor into the destination tensor. r@rrA) rr.rBr@Z is_cpu_placerErDZXPUPlacer4r#reload)rrFZarrZxpurrr async_reloads  rIcCsRt|jdks Jdt|jdksJd|j|jks Jd||||||S)ac Offloading the source tensor into the destination tensor asynchronously with offset and size customized. Args: src_tensor (EagerParamBase|paddle.Tensor): The source tensor. dst_tensor (EagerParamBase|paddle.Tensor): The destination tensor. src_offset (int): The element offset of the source tensor. dst_offset (int): The element offset of the destination tensor. offload_size (int): The size of the data to be loaded. async_loader (core.AsyncLoad): The AsyncLoad object. Returns: task (Task): The task that operates partial offloading. zOnly support 1-D tensorzOnly support same dtype)lenrr Zoffload_with_offset)rZ dst_tensorZ src_offsetZ dst_offsetZ offload_sizeZ async_loaderrrrasync_offload_with_offsets  rLTrJcCsX|rtd|itjtj|dStdditjtjgdS)z# Enable activation offload ZFLAGS_offload_retry_timesrN)rZ set_flagsrZregister_offload_callbackZ!set_skip_offload_callback_tensors parametersZclear_offload_callback)modelenableZ retry_timesrrrenable_activation_offloads  rP)r)TrJ)rrrZ paddle.baserZpaddle.base.data_feederrZpaddle.base.frameworkrZpaddle.base.layer_helperrZpaddle.frameworkr__all__rr#r2r3r4rGrIrLrPrrrrs$       < &