o
    H&ihU                     @   s   d dl Z d dlZd dlmZ ddlmZ eejds2edejjd< edejjd< edejjd< d d	l	m
Z
mZmZ d
d Zdd ZG dd dejjZG dd dZ	dddZdS )    N)_pytree   )_dummy_typeZ_CudaStreamBase
_CUDAGraph_graph_pool_handle_cuda_isCurrentStreamCapturing)r   r   r   c                   C      t  S )zReturn True if CUDA graph capture is underway on the current CUDA stream, False otherwise.

    If a CUDA context does not exist on the current device, returns False without initializing the context.
    )r    r	   r	   <C:\wamp64\www\opt\env\Lib\site-packages\torch/cuda/graphs.pyis_current_stream_capturing   s   r   c                   C   r   )zReturn an opaque token representing the id of a graph memory pool.

    See :ref:`Graph memory management<graph-memory-management>`.

    .. warning::
        This API is in beta and may change in future releases.
    )r   r	   r	   r	   r
   graph_pool_handle   s   r   c                       sv   e Zd ZdZ fddZd fdd	Z fdd	Z fd
dZ fddZ fddZ	 fddZ
 fddZ  ZS )	CUDAGraphzrWrapper around a CUDA graph.

    .. warning::
        This API is in beta and may change in future releases.
    c                    s   t  | S N)super__new__)cls	__class__r	   r
   r   2   s   zCUDAGraph.__new__Nglobalc                    s   t  j||d dS )a  Begin capturing CUDA work on the current stream.

        Typically, you shouldn't call ``capture_begin`` yourself.
        Use :class:`~torch.cuda.graph` or :func:`~torch.cuda.make_graphed_callables`,
        which call ``capture_begin`` internally.

        Arguments:
            pool (optional): Token (returned by :func:`~torch.cuda.graph_pool_handle` or
                :meth:`other_Graph_instance.pool()<torch.cuda.CUDAGraph.pool>`) that hints this graph may share memory
                with the indicated pool.  See :ref:`Graph memory management<graph-memory-management>`.
            capture_error_mode (str, optional): specifies the cudaStreamCaptureMode for the graph capture stream.
                Can be "global", "thread_local" or "relaxed". During cuda graph capture, some actions, such as cudaMalloc,
                may be unsafe. "global" will error on actions in other threads, "thread_local" will only error for
                actions in the current thread, and "relaxed" will not error on these actions. Do NOT change this setting
                unless you're familiar with `cudaStreamCaptureMode <https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM_1g9d0535d93a214cbf126835257b16ba85>`_
        )poolcapture_error_modeN)r   capture_begin)selfr   r   r   r	   r
   r   5   s   zCUDAGraph.capture_beginc                       t    dS )aG  End CUDA graph capture on the current stream.

        After ``capture_end``, ``replay`` may be called on this instance.

        Typically, you shouldn't call ``capture_end`` yourself.
        Use :class:`~torch.cuda.graph` or :func:`~torch.cuda.make_graphed_callables`,
        which call ``capture_end`` internally.
        N)r   capture_endr   r   r	   r
   r   H   s   	zCUDAGraph.capture_endc                    r   )z,Replay the CUDA work captured by this graph.N)r   replayr   r   r	   r
   r   S      zCUDAGraph.replayc                    r   )z1Delete the graph currently held by this instance.N)r   resetr   r   r	   r
   r   W   r   zCUDAGraph.resetc                    
   t   S )zReturn an opaque token representing the id of this graph's memory pool.

        This id can optionally be passed to another graph's ``capture_begin``,
        which hints the other graph may share the same memory pool.
        )r   r   r   r   r	   r
   r   [   s   
zCUDAGraph.poolc                    r   )z/Enable debugging mode for CUDAGraph.debug_dump.)r   enable_debug_moder   r   r	   r
   r    c   s   
zCUDAGraph.enable_debug_modec                    s   t  |S )z
        Arguments:
            debug_path (required): Path to dump the graph to.

        Calls a debugging function to dump the graph if the debugging is
        enabled via CUDAGraph.enable_debug_mode()
        )r   
debug_dump)r   Z
debug_pathr   r	   r
   r!   g   s   zCUDAGraph.debug_dump)Nr   )__name__
__module____qualname____doc__r   r   r   r   r   r   r    r!   __classcell__r	   r	   r   r
   r   +   s    r   c                   @   s:   e Zd ZdZdZ			ddefddZdd Zd	d
 ZdS )grapha  Context-manager that captures CUDA work into a :class:`torch.cuda.CUDAGraph` object for later replay.

    See :ref:`CUDA Graphs <cuda-graph-semantics>` for a general introduction,
    detailed use, and constraints.

    Arguments:
        cuda_graph (torch.cuda.CUDAGraph): Graph object used for capture.
        pool (optional): Opaque token (returned by a call to :func:`~torch.cuda.graph_pool_handle()` or
            :meth:`other_Graph_instance.pool()<torch.cuda.CUDAGraph.pool>`) hinting this graph's capture
            may share memory from the specified pool. See :ref:`Graph memory management<graph-memory-management>`.
        stream (torch.cuda.Stream, optional): If supplied, will be set as the current stream in the context.
            If not supplied, ``graph`` sets its own internal side stream as the current stream in the context.
        capture_error_mode (str, optional): specifies the cudaStreamCaptureMode for the graph capture stream.
            Can be "global", "thread_local" or "relaxed". During cuda graph capture, some actions, such as cudaMalloc,
            may be unsafe. "global" will error on actions in other threads, "thread_local" will only error for
            actions in the current thread, and "relaxed" will not error on actions. Do NOT change this setting
            unless you're familiar with `cudaStreamCaptureMode <https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM_1g9d0535d93a214cbf126835257b16ba85>`_

    .. note::
        For effective memory sharing, if you pass a ``pool`` used by a previous capture and the previous capture
        used an explicit ``stream`` argument, you should pass the same ``stream`` argument to this capture.

    .. warning::
        This API is in beta and may change in future releases.

    .. _cudaStreamCaptureMode:
        https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html#group__CUDART__STREAM_1g9d0535d93a214cbf126835257b16ba85
    Nr   r   c                 C   sr   | j jd u rtj | j _|d u rdn|f| _|d ur|n| j j| _| jd us)J tj| j| _|| _	|| _
d S )Nr	   )r   default_capture_streamtorchcudaStreamr   Zcapture_streamstream
stream_ctx
cuda_graphr   )r   r.   r   r,   r   r	   r	   r
   __init__   s   

zgraph.__init__c                 C   s@   t j  t  t j  | j  | jj	| j
d| ji d S )Nr   )r)   r*   synchronizegcZcollectZempty_cacher-   	__enter__r.   r   r   r   r   r	   r	   r
   r2      s   



zgraph.__enter__c                 C   s   | j   | j||| d S r   )r.   r   r-   __exit__)r   exc_type	exc_value	tracebackr	   r	   r
   r3      s   
zgraph.__exit__)NNr   )	r"   r#   r$   r%   r(   strr/   r2   r3   r	   r	   r	   r
   r'   r   s    
r'      Fc           %         sl  t  rt  rtdd}t| tsd}| f} |f}g  t| |D ]L\}}t|t jjrUt	|j
dkrBt	|jdkrBt	|jdksFJ dtdd | D sUJ dtj| } t| td	d |D snJ d
q"dd  D }dd | D  fddtt	| D }	dd tt	| D }
dd tt	| D }t }t j  t jt j D t| ||	D ]5\}}}t|D ])}t|| }t jjtdd |D tdd |D tdd |D d|d}q~~qW d   n1 sw   Y  t j  g }g }t| ||
D ]6\}}}t jj||d || }W d   n	1 s*w   Y  t|\}}|t| || qg }g }tt|	t|t|tD ]w\}}}}tdd |D }t jj||d( t jjtdd |D tdd |D tdd |D d|d}W d   n	1 sw   Y  g }d}|D ]}|jr|||  |d7 }q|d qt|}|| || qWtt|}tt|}dd } g }!t | D ]E\}"}| |
|" ||" |" ||" ||" |	|" ||" ||" ||" 	}#t|t jjr$dd }$|$||j!|#|j"|_"|!| q|!|# q|r2|!d S t|!S ) a  Accept callables (functions or :class:`nn.Module<torch.nn.Module>`\ s) and returns graphed versions.

    Each graphed callable's forward pass runs its source callable's
    forward CUDA work as a CUDA graph inside a single autograd node.

    The graphed callable's forward pass also appends
    a backward node to the autograd graph. During backward, this node runs the
    callable's backward work as a CUDA graph.

    Therefore, each graphed callable should be a drop-in replacement for its source callable
    in an autograd-enabled training loop.

    See :ref:`Partial-network capture<partial-network-capture>` for detailed use and constraints.

    If you pass a tuple of several callables, their captures will use the same memory pool.
    See :ref:`Graph memory management<graph-memory-management>` for when this is appropriate.

    Arguments:
        callables (torch.nn.Module or Python function, or tuple of these): Callable or callables to graph.
            See :ref:`Graph memory management<graph-memory-management>` for when passing a tuple of callables
            is appropriate.  If you pass a tuple of callables, their order in the tuple must be the same order
            they'll run in the live workload.
        sample_args (tuple of Tensors, or tuple of tuples of Tensors): Samples args for each callable.
            If a single callable was passed, ``sample_args`` must be a single tuple of argument Tensors.
            If a tuple of callables was passed, ``sample_args`` must be tuple of tuples of argument Tensors.
        num_warmup_iters (int): The number of warmup iterations. Currently, ``DataDistributedParallel`` needs
            11 iterations for warm up. Default: ``3``.
        allow_unused_input (bool): If False, specifying inputs that were not used when computing outputs
            (and therefore their grad is always zero) is an error. Defaults to False.

    .. note::
        The ``requires_grad`` state of each Tensor in ``sample_args`` must match the state
        that's expected for the corresponding real input in the training loop.

    .. warning::
        This API is in beta and may change in future releases.

    .. warning::
        ``sample_args`` for each callable must contain only Tensors. Other types are not allowed.

    .. warning::
        Returned callables do not support higher order differentiation (e.g., double backward).

    .. warning::
        In any :class:`~torch.nn.Module` passed to :func:`~make_graphed_callables`, only parameters
        may be trainable. Buffers must have ``requires_grad=False``.

    .. warning::
        After you pass a :class:`torch.nn.Module` through :func:`~make_graphed_callables`,
        you may not add or remove any of that Module's parameters or buffers.

    .. warning::
        :class:`torch.nn.Module`\s passed to :func:`~torch.cuda.make_graphed_callables` must not have module hooks
        registered on them at the time they are passed. However, registering hooks on modules *after* passing them
        through :func:`~torch.cuda.make_graphed_callables` is allowed.

    .. warning::
        When running a graphed callable, you must pass its arguments in the same order and format
        they appeared in that callable's ``sample_args``.

    .. warning::
        The automatic mixed precision is supported in :func:`~torch.cuda.make_graphed_callables` only with disabled
        caching. The context manager `torch.cuda.amp.autocast()` must have `cache_enabled=False`.
    z_make_graphed_callables does not support the autocast caching. Please set `cache_enabled=False`.FTr   zModules must not have hooks registered at the time they are passed. However, registering hooks on modules after passing them through make_graphed_callables is allowed.c                 s   s    | ]}|j d u V  qdS )FNrequires_grad.0br	   r	   r
   	<genexpr>      z)make_graphed_callables.<locals>.<genexpr>zIn any :class:`~torch.nn.Module` passed to :func:`~make_graphed_callables`, only parameters may be trainable. All buffers must have ``requires_grad=False``.c                 s   s    | ]	}t |tjV  qd S r   )
isinstancer)   ZTensor)r<   argr	   r	   r
   r>         zfIn the beta API, sample_args for each callable must contain only Tensors. Other types are not allowed.c                 S   s   g | ]}t |qS r	   )len)r<   argsr	   r	   r
   
<listcomp>%  s    z*make_graphed_callables.<locals>.<listcomp>c                 S   s*   g | ]}t |tjjrt| nd qS )r	   )r@   r)   nnModuletuple
parameters)r<   cr	   r	   r
   rE   &  s    c                    s   g | ]
} | |  qS r	   r	   r<   iZflatten_sample_argsZper_callable_module_paramsr	   r
   rE   *  s    c                 S      g | ]}t j qS r	   r)   r*   r   r<   _r	   r	   r
   rE   /      c                 S   rN   r	   rO   rP   r	   r	   r
   rE   0  rR   c                 s       | ]}|j r|V  qd S r   r9   r<   or	   r	   r
   r>   ?  r?   c                 s   rS   r   r9   rK   r	   r	   r
   r>   @  r?   c                 s   s     | ]}|j rt|V  qd S r   r:   r)   Z
empty_likerT   r	   r	   r
   r>   A  s    
)outputsinputsZgrad_outputsZonly_inputsZallow_unusedN)r   c                 s   s$    | ]}|j rt|nd V  qd S r   rV   rT   r	   r	   r
   r>   d      
c                 s   rS   r   r9   rT   r	   r	   r
   r>   j  r?   c                 s   rS   r   r9   rK   r	   r	   r
   r>   k  r?   c                 s   s    | ]	}|d ur|V  qd S r   r	   rT   r	   r	   r
   r>   l  rB   r   c	           
         s8   G 	fdddt jj  fdd}	|	S )Nc                       s@   e Zd ZefddZeejjj fddZ	dS )zOmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphedc                    s`   t D ]}|  ||  kr| ||  q   tts'J tdd D S )Nc                 s   s    | ]}|  V  qd S r   detachrT   r	   r	   r
   r>     s    zjmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphed.forward.<locals>.<genexpr>)rangedata_ptrcopy_r   r@   rH   )ctxrX   rL   )	fwd_graphlen_user_argsstatic_input_surfacestatic_outputsr	   r
   forward  s   zWmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphed.forwardc                    sr   t |t ks
J t|D ]\}}|d ur$| | kr$|| q   tts0J tdd D S )Nc                 s   s$    | ]}|d ur|  n|V  qd S r   rZ   r;   r	   r	   r
   r>     rY   zkmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphed.backward.<locals>.<genexpr>)rC   zipr]   r^   r   r@   rH   )r_   Zgradsggrad)	bwd_graphstatic_grad_inputsstatic_grad_outputsr	   r
   backward  s   
zXmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphed.backwardN)
r"   r#   r$   staticmethodrd   r)   autogradfunctionZonce_differentiablerk   r	   )rh   r`   ra   ri   rj   rb   rc   r	   r
   Graphed  s    	ro   c                     s(   t j|  } jt|  }t |S r   )r   arg_tree_leavesapplyrH   Ztree_unflatten)	user_argsZflatten_user_argsout)ro   module_paramsoutput_unflatten_specr	   r
   functionalized  s   
zVmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.functionalized)r)   rm   Function)
r`   rh   rt   ra   ru   rb   rc   rj   ri   rv   r	   )
ro   rh   r`   ra   rt   ru   ri   rj   rb   rc   r
   make_graphed_autograd_function  s   $z>make_graphed_callables.<locals>.make_graphed_autograd_functionc                    s    fdd}|S )Nc                     s    j kr	|  S |  S r   )training)rr   funcgraph_training_stategraphedorig_fwdr	   r
   new_fwd  s   
zEmake_graphed_callables.<locals>.make_graphed_forward.<locals>.new_fwdr	   )r{   r|   r}   r~   r   r	   rz   r
   make_graphed_forward  s   z4make_graphed_callables.<locals>.make_graphed_forward)#r)   Zis_autocast_enabledZis_autocast_cache_enabledRuntimeErrorr@   rH   re   rF   rG   rC   Z_backward_hooksZ_forward_hooksZ_forward_pre_hooksallbuffersr   rp   appendr\   r   r*   r0   r,   r+   Ztree_leavesrm   rg   r'   Ztree_flattenreversedr:   list	enumeratery   rd   )%Z	callablesZsample_argsZnum_warmup_itersZallow_unused_inputZjust_one_callablerJ   rD   Zflatten_argZper_callable_len_user_argsZ"per_callable_static_input_surfacesZ
fwd_graphsZ
bwd_graphsZmempoolr{   rb   rQ   rW   Zgrad_inputsZper_callable_static_outputsZ"per_callable_output_unflatten_specr`   Zflatten_outputsspecZ per_callable_static_grad_outputsZper_callable_static_grad_inputsrc   rh   rt   rj   ri   Zgrad_idxrA   rx   retrL   r}   r   r	   rM   r
   make_graphed_callables   s   C



	


3r   )r8   F)r1   r)   Ztorch.utilsr   _utilsr   hasattrZ_C__dict__Ztorch._Cr   r   r   r   r   r   r'   r   r	   r	   r	   r
   <module>   s"    	GK