&`iaddlmZmZddlmZmZmZmZddlm Z ddl m Z edZ e Gddeee Z e Gdd e Ze Gd d e eZd S) )ABCabstractmethod)AnyDictGenericTypeVar)AcceleratorContext) DeveloperAPITc(eZdZdZedZdS)DAGOperationFuturez A future representing the result of a DAG operation. This is an abstraction that is internal to each actor, and is not exposed to the DAG caller. ct)zM Wait for the future and return the result of the operation. )NotImplementedErrorselfs p/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/ray/dag/dag_operation_future.pywaitzDAGOperationFuture.waits "!N)__name__ __module__ __qualname____doc__rrrrr r s9""^"""rr ceZdZdZdZdZdS)ResolvedFuturez~ A future that is already resolved. Calling `wait()` on this will immediately return the result without blocking. c||_dS)zl Initialize a resolved future. Args: result: The result of the future. N_result)rresults r__init__zResolvedFuture.__init__"s rc|jS)zX Wait and immediately return the result. This operation will not block. rrs rrzResolvedFuture.wait+s |rN)rrrrr rrrrrrs< rrceZdZUdZiZeedfed<ededdddfdZ ededdfdZ dd e ded e fd Z de fd Z dd ZdS) GPUFuturea A future for a GPU event on a CUDA stream. This future wraps a buffer, and records an event on the given stream when it is created. When the future is waited on, it makes the current CUDA stream wait on the event, then returns the buffer. The buffer must be a GPU tensor produced by an earlier operation launched on the given stream, or it could be CPU data. Then the future guarantees that when the wait() returns, the buffer is ready on the current stream. The `wait()` does not block CPU. gpu_futuresfut_idfutreturnNc|tjvr1tj||tj|<dS)z Cache the GPU future. Args: fut_id: GPU future ID. fut: GPU future to be cached. Nr#r$pop destroy_event)r%r&s radd_gpu_futurezGPUFuture.add_gpu_futureNsH Y* * *  ! % %f - - ; ; = = =(+ f%%%rc|tjvr3tj|dSdS)z{ Remove the cached GPU future and destroy its CUDA event. Args: fut_id: GPU future ID. Nr))r%s rremove_gpu_futurezGPUFuture.remove_gpu_future[sC Y* * *  ! % %f - - ; ; = = = = = + *rbufstreamc<|%tj}||_tj|_|j|||_d|_t ||dS)an Initialize a GPU future on the given stream. Args: buf: The buffer to return when the future is resolved. fut_id: The future ID to cache the future. stream: The torch stream to record the event on, this event is waited on when the future is resolved. If None, the current stream is used. NF) r getcurrent_stream_buf create_event_eventrecord_fut_id_waitedr#r,)rr/r%r0s rr zGPUFuture.__init__es >'+--<<>>F (,..;;==  6""" "    .....rctj}|js@d|_||jt |j|j S)z Wait for the future on the current CUDA stream and return the result from the GPU operation. This operation does not block CPU. T) r r2r3r9 wait_eventr6r#r.r8r4)rr3s rrzGPUFuture.wait{sc ,/11@@BB| 6DL  % %dk 2 2 2  ' ' 5 5 5yrc&|jdSd|_dS)zE Destroy the CUDA event associated with this future. N)r6rs rr+zGPUFuture.destroy_events ;  F r)N)r'N)rrrrr$rint__annotations__ staticmethodr,r.rr rr+rrrr#r#2s  0+-Kc;&',,, ,s , , , , ,\ ,>#>$>>>\>//C//c////, c    rr#N)abcrrtypingrrrr,ray.experimental.channel.accelerator_contextr ray.util.annotationsr r r rr#rrrrDs5########............KKKKKK------ GCLL " " " " "gaj " " " ',]]]]]"3']]]]]r