o
    ^il                     @  s~  U d dl mZ d dlZd dlZd dlmZmZmZmZmZ d dl	m
Z
mZmZmZ d dlZd dlmZ er;d dlmZ ddlmZ g d	Zed
Ze
dZeejdske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 d2ddZd3ddZG dd dejjZ G dd dZ!edede"f f Z#de$d< e	 	!	d4d5d*d+Z%e	 	!	d4d6d.d+Z%	 	!	d4d7d1d+Z%dS )8    )annotationsN)CallableOptionaloverloadTYPE_CHECKINGUnion)	ParamSpecSelf	TypeAliasTypeVar)Tensor)_POOL_HANDLE   )_dummy_type)is_current_stream_capturinggraph_pool_handle	CUDAGraphgraphmake_graphed_callables_R_P_CudaStreamBase
_CUDAGraph_graph_pool_handle_cuda_isCurrentStreamCapturing)r   r   r   returnboolc                   C  s   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   L/var/www/html/RAG/RAG_venv/lib/python3.10/site-packages/torch/cuda/graphs.pyr   /   s   r   r   c                   C  s   t jt S )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.
    )torchcudar   r   r   r   r   r   r   8   s   r   c                      s   e Zd ZdZd'd( fddZ	
d)d* fddZd+ fddZd+ fddZd+ fddZd+ fddZ	d, fddZ
d+ fddZd- fd d!Zd. fd#d$Zd. fd%d&Z  ZS )/r   a-  Wrapper around a CUDA graph.

    Arguments:
        keep_graph (bool, optional): If ``keep_graph=False``, the
            cudaGraphExec_t will be instantiated on GPU at the end of
            ``capture_end`` and the underlying cudaGraph_t will be
            destroyed. Users who want to query or otherwise modify the
            underlying cudaGraph_t before instantiation can set
            ``keep_graph=True`` and access it via ``raw_cuda_graph`` after
            ``capture_end``. Note that the cudaGraphExec_t will not be
            instantiated at the end of ``capture_end`` in this
            case. Instead, it will be instantiated via an explicit called
            to ``instantiate`` or automatically on the first call to
            ``replay`` if ``instantiate`` was not already called. Calling
            ``instantiate`` manually before ``replay`` is recommended to
            prevent increased latency on the first call to ``replay``. It
            is allowed to modify the raw cudaGraph_t after first calling
            ``instantiate``, but the user must call ``instantiate`` again
            manually to make sure the instantiated graph has these
            changes. Pytorch has no means of tracking these changes.

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

    F
keep_graphr   r   r	   c                   s   t  | |S N)super__new__)clsr!   	__class__r   r   r$   _   s   zCUDAGraph.__new__NglobalpoolOptional[_POOL_HANDLE]capture_error_modestrNonec                   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>`_
        )r)   r+   N)r#   capture_begin)selfr)   r+   r&   r   r   r.   b   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   r1   w   s   	zCUDAGraph.capture_endc                   r0   )a$  Instantiate the CUDA graph. Will be called by
        ``capture_end`` if ``keep_graph=False``, or by ``replay`` if
        ``keep_graph=True`` and ``instantiate`` has not already been
        explicitly called. Does not destroy the cudaGraph_t returned
        by ``raw_cuda_graph``.
        N)r#   instantiater2   r&   r   r   r3      s   zCUDAGraph.instantiatec                   r0   )z,Replay the CUDA work captured by this graph.N)r#   replayr2   r&   r   r   r4         zCUDAGraph.replayc                   r0   )z1Delete the graph currently held by this instance.N)r#   resetr2   r&   r   r   r6      r5   zCUDAGraph.resetr   c                   
   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)   r2   r&   r   r   r)      s   
zCUDAGraph.poolc                   r7   )z/Enable debugging mode for CUDAGraph.debug_dump.)r#   enable_debug_moder2   r&   r   r   r8      s   
zCUDAGraph.enable_debug_mode
debug_pathc                   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/   r9   r&   r   r   r:      s   zCUDAGraph.debug_dumpintc                   r7   )a}  Returns the underlying cudaGraph_t. ``keep_graph`` must be True.

        See the following for APIs for how to manipulate this object: `Graph Managmement <https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__GRAPH.html>`_ and `cuda-python Graph Management bindings <https://nvidia.github.io/cuda-python/cuda-bindings/latest/module/runtime.html#graph-management>`_
        )r#   raw_cuda_graphr2   r&   r   r   r<         
zCUDAGraph.raw_cuda_graphc                   r7   )a  Returns the underlying cudaGraphExec_t. ``instantiate`` must have been called if ``keep_graph`` is True, or ``capture_end`` must have been called if ``keep_graph`` is False. If you call ``instantiate()`` after ``raw_cuda_graph_exec()``, the previously returned cudaGraphExec_t will be destroyed. It is your responsibility not to use this object after destruction.

        See the following for APIs for how to manipulate this object: `Graph Execution <https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__GRAPH__EXEC.html>`_ and `cuda-python Graph Execution bindings <https://nvidia.github.io/cuda-python/cuda-bindings/latest/module/runtime.html#graph-execution>`_
        )r#   raw_cuda_graph_execr2   r&   r   r   r>      r=   zCUDAGraph.raw_cuda_graph_exec)F)r!   r   r   r	   )Nr(   )r)   r*   r+   r,   r   r-   r   r-   r   r   )r9   r,   r   r-   )r   r;   )__name__
__module____qualname____doc__r$   r.   r1   r3   r4   r6   r)   r8   r:   r<   r>   __classcell__r   r   r&   r   r   D   s    	
r   c                   @  sD   e Zd ZU dZdZded< 			ddddZdddZdddZdS )r   a  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
    NOptional[torch.cuda.Stream]default_capture_streamr(   
cuda_graphr   r)   r*   streamr+   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'   rG   r   r    Streamr)   capture_streamrI   
stream_ctxrH   r+   )r/   rH   r)   rI   r+   r   r   r   __init__   s   

zgraph.__init__r   r-   c                 C  sJ   t j  t jjjrt  t j  | j	
  | jj| jd| ji d S )Nr+   )r   r    synchronizecompilerconfigforce_cudagraph_gcgccollectempty_cacherL   	__enter__rH   r.   r)   r+   r2   r   r   r   rU      s   




zgraph.__enter__argsobjectc                 G  s   | j   | jj|  d S r"   )rH   r1   rL   __exit__)r/   rV   r   r   r   rX     s   
zgraph.__exit__)NNr(   )rH   r   r)   r*   rI   rF   r+   r,   r?   )rV   rW   r   r-   )	rA   rB   rC   rD   rG   __annotations__rM   rU   rX   r   r   r   r   r      s   
 
r   torch.nn.Module.r
   _ModuleOrCallable   F	callablessample_argstuple[Tensor, ...]num_warmup_itersr;   allow_unused_inputr)   r*   c                 C     d S r"   r   r]   r^   r`   ra   r)   r   r   r   r        r   tuple[_ModuleOrCallable, ...]tuple[tuple[Tensor, ...], ...]c                 C  rb   r"   r   rc   r   r   r   r     rd   7Union[_ModuleOrCallable, tuple[_ModuleOrCallable, ...]]9Union[tuple[Tensor, ...], tuple[tuple[Tensor, ...], ...]]c           )        s  t  rt  rtdd}t| ts$d}| f} tttdf |f}nttttdf df |}g  t	| |D ]N\}}t|t j
jrlt|jdkrYt|jdkrYt|jdks]J dtdd | D slJ d	t jjj| }	 t|	 td
d |	D sJ dq9dd  D }
dd | D  fddtt| D }dd tt| D }dd tt| D }|du rt n|}t j  t jt j \ t	| ||D ]M\}}}d\}}}t|D ]4}t jj|| }tdd |D }t|dkrt jj|tdd |D tdd |D d|d}q|||fD ]}~q qW d   n	1 s1w   Y  t j  g }g }t	| ||D ]8\}}}t jj||d || }W d   n	1 sbw   Y  t jj |\}}|t| || qEg }g }t	t!|t!|t!|D ]\}}}tdd |D } tdd |D }d}t|dkrt jj||d! t jj|tdd |D tdd | D d|d}W d   n	1 sw   Y  g }!d}"|D ]}#|#j"r|dur|!||"  |"d7 }"q|!d qt|!}!||  ||! q|#  |#  d:d/d0}$g }%t$| D ]F\}&}|$||& ||& |& |
|& ||& ||& ||& ||& ||& 	}'t|t j
jrhd;d8d9}(|(||j%|'|j&|_&|%| q(|%|' q(|rv|%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.
        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>`.
    .. 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`.FT.r   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   r   )rl   argr   r   r   rn         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)rl   rV   r   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   )rp   r   nnModuletuple
parameters)rl   cr   r   r   rt     s    c                   s   g | ]
} | |  qS r   r   rl   iflatten_sample_argsper_callable_module_paramsr   r   rt     s    c                 S     g | ]}t j qS r   r   r    r   rl   _r   r   r   rt         c                 S  r   r   r   r   r   r   r   rt     r   N)NNNc                 s      | ]}|j r|V  qd S r"   ri   rl   or   r   r   rn     ro   c                 s  r   r"   ri   rz   r   r   r   rn     s    
c                 s  s     | ]}|j rt|V  qd S r"   rj   r   
empty_liker   r   r   r   rn     s    
)outputsinputsgrad_outputsonly_inputsallow_unused)r)   c                 s  s$    | ]}|j rt|nd V  qd S r"   r   r   r   r   r   rn         
c                 s  r   r"   ri   r   r   r   r   rn     ro   c                 s  r   r"   ri   rz   r   r   r   rn     ro   c                 s  s    | ]	}|d ur|V  qd S r"   r   r   r   r   r   rn     rr      	fwd_graphr   	bwd_graphmodule_paramstuple[torch.nn.Parameter, ...]len_user_argsr;   output_unflatten_spectorch.utils._pytree.TreeSpecstatic_input_surfacer_   static_outputsstatic_grad_outputstuple[Optional[Tensor], ...]static_grad_inputsr   Callable[..., object]c	           
        s:   G 	fdddt jj d fdd}	|	S )	Nc                      sD   e Zd ZedfddZeejjjd fd
dZ	dS )zOmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.GraphedctxrW   r   r   r   r_   c                   s`   t D ]}|  ||  kr| ||  q   tts'J tdd D S )Nc                 s  s    | ]}|  V  qd S r"   detachr   r   r   r   rn     s    zjmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphed.forward.<locals>.<genexpr>)rangedata_ptrcopy_r4   rp   rw   )r   r   r{   )r   r   r   r   r   r   forward  s   zWmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphed.forwardgradsc                   sr   t |t ks
J t|D ]\}}|d ur$| | kr$|| q   tts0J tdd D S )Nc                 s  s$    | ]}|d ur|  n|V  qd S r"   r   rk   r   r   r   rn   %  r   zkmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphed.backward.<locals>.<genexpr>)rs   zipr   r   r4   rp   rw   )r   r   ggrad)r   r   r   r   r   backward  s   
zXmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.Graphed.backwardN)r   rW   r   r   r   r_   )r   rW   r   r   r   r_   )
rA   rB   rC   staticmethodr   r   autogradfunctiononce_differentiabler   r   )r   r   r   r   r   r   r   r   r   Graphed  s    	r   	user_argsrW   r   c                    s0   t jjj|  } jt|  }t jj|S r"   )r   utils_pytreearg_tree_leavesapplyrw   tree_unflatten)r   flatten_user_argsout)r   r   r   r   r   functionalized)  s   zVmake_graphed_callables.<locals>.make_graphed_autograd_function.<locals>.functionalized)r   rW   r   rW   )r   r   Function)
r   r   r   r   r   r   r   r   r   r   r   )
r   r   r   r   r   r   r   r   r   r   r   make_graphed_autograd_function  s   $z>make_graphed_callables.<locals>.make_graphed_autograd_functionfuncrZ   graph_training_stater   graphedCallable[_P, _R]orig_fwdc                   s   d	 fdd}|S )
Nr   _P.argsuser_kwargs	_P.kwargsr   r   c                    s&    j kr| i |S | i |S r"   )training)r   r   r   r   r   r   r   r   new_fwdJ  s   
zEmake_graphed_callables.<locals>.make_graphed_forward.<locals>.new_fwd)r   r   r   r   r   r   r   )r   r   r   r   r   r   r   r   make_graphed_forwardD  s   z4make_graphed_callables.<locals>.make_graphed_forward)r   r   r   r   r   r   r   r;   r   r   r   r_   r   r_   r   r   r   r_   r   r   )
r   rZ   r   r   r   r   r   r   r   r   )'r   is_autocast_enabledis_autocast_cache_enabledRuntimeErrorrp   rw   typingcastr   r   ru   rv   rs   _backward_hooks_forward_hooks_forward_pre_hooksallbuffersr   r   r   appendr   r   r    rN   rI   rJ   tree_leavesr   r   r   tree_flattenreversedrj   reverse	enumerater   r   ))r]   r^   r`   ra   r)   just_one_callable_sample_argsry   rV   flatten_argper_callable_len_user_args"per_callable_static_input_surfaces
fwd_graphs
bwd_graphsmempoolr   r   grad_inputsr   outputs_gradr   vper_callable_static_outputs"per_callable_output_unflatten_specr   func_outputsflatten_outputsspec per_callable_static_grad_outputsper_callable_static_grad_inputsr   r   r   r   grad_idxrq   r   retr{   r   r   r   r|   r   r   %  s  I







3
)r   r   r@   )r\   FN)r]   r[   r^   r_   r`   r;   ra   r   r)   r*   r   r[   )r]   re   r^   rf   r`   r;   ra   r   r)   r*   r   re   )r]   rg   r^   rh   r`   r;   ra   r   r)   r*   r   rg   )&
__future__r   rR   r   r   r   r   r   r   typing_extensionsr   r	   r
   r   r   r   
torch.cudar   _utilsr   __all__r   r   hasattr_C__dict__torch._Cr   r   r   r   r   r   r   rW   r[   rY   r   r   r   r   r   <module>   sN    	

	tV	