U d@sUdZddlmZddlmZmZmZmZmZddl Z ddl m Z ddl m Z m Z mZmZmZmZgZeeed<ee ZGd d d ZGd d d e jjZGd dde jjZdS)zkAutograd functions for stream-aware CUDA copy. It is used to overlap copy and computation on the same GPU. )deque)DequeListOptionalTupleSequenceN)Tensor)AbstractStreamcurrent_stream get_device record_stream use_stream wait_stream__all__c@seZdZUeed<eed<dS)Context prev_stream next_streamN)__name__ __module__ __qualname__r __annotations__rrH/tmp/pip-unpacked-wheel-ua33x9lu/torch/distributed/pipeline/sync/copy.pyrs rc@sJeZdZdZeeeeedddZeee e e e dfdddZ d S) Copyz#Copies tensors on specific streams.ctxrrreturnc Gs||_||_g}tt|}t|jt|V|D]J}t|rt|jt|dd}||t ||t ||q4||q4W5QRXW5QRXt |S)NTZ non_blocking) rrr r rtorch is_tensortoappendr tuple)rrrinputoutputZ output_streamxyrrrforward!s      z Copy.forward.)r grad_outputrc Gs|j}|j}tt|d}tt|}t|Xt|Dt|D]4}|jt|dd}| |t ||t ||qBW5QRXW5QRXd}|t |S)N)maxlenTrNN) rrrlenr r rreversedr! appendleftr r#) rr)rr grad_inputZ input_streamr&r' grad_streamsrrrbackward:s     z Copy.backwardN rrr__doc__ staticmethodrr Tensorsr(rrrr1rrrrrs rc@sJeZdZdZeeeeedddZeee e e e dfdddZ d S) WaitzSynchronizes a stream to another stream. Place it just before you want to start an operation on the next stream, provided that all operations on the previous stream are done. rcGs(||_||_t||tdd|DS)Ncss$|]}t|r|n|VqdS)N)rr detach).0r&rrr aszWait.forward..)rrrr#)rrrr$rrrr(Ys z Wait.forward.)rr/rcGs"|j}|j}t||d}||S)Nr+)rrr)rr/rrr0rrrr1cs  z Wait.backwardNr2rrrrr6Qs r6)r3 collectionsrtypingrrrrrrrstreamr r r r rrrstrrr5rZautogradZFunctionrr6rrrrs   3