˲j5dZddlmZmZddlmZddlmZddlZddl m cm Z ddl mZmZmZmZddlmZddlmZmZmZmZgd ZejeejeejeejeGd d eZGd d ejZdefdZdefdZ dS)zCDefines bias subclasses that work with scaled_dot_product_attention)autoIntEnum)Optional)warnN)can_use_efficient_attentioncan_use_flash_attentionis_flash_attention_available SDPAParams)_raise_kernel_warnings)_calculate_scale_input_requires_grad_postprocess_flash_output_validate_sdpa_input)causal_upper_leftcausal_lower_right CausalVariant CausalBiasc:eZdZdZeZeZdS)ra# Enum for causal variants used in attention mechanisms. Defines two types of causal biases: `UPPER_LEFT`: Represents upper-left triangular bias for standard causal attention. The equivalent pytorch code for constructing this bias is: .. code-block:: python torch.tril(torch.ones(size, dtype=torch.bool)) For instance, with `shape=(3,4)`, the materialized bias tensor will be: .. code-block:: text [[1, 0, 0, 0], [1, 1, 0, 0], [1, 1, 1, 0]] `LOWER_RIGHT`: Represents lower-right triangular bias, the include values are aligned to the lower right corner of the matrix. The equivalent pytorch code for constructing this bias is: .. code-block:: python diagonal_offset = size[1] - size[0] torch.tril( torch.ones(size, dtype=torch.bool), diagonal=diagonal_offset, ) For instance, with `shape=(3,4)`, the materialized bias tensor will be: .. code-block:: text [[1, 1, 0, 0], [1, 1, 1, 0], [1, 1, 1, 1]] Note that these variants are equivalent to each other when the sequence lengths of the query and key/value tensors are equal since the triangular matrix is square. .. warning:: This enum is a prototype and subject to change. N)__name__ __module__ __qualname____doc__r UPPER_LEFT LOWER_RIGHTHC:\Projects\Historical\venv\Lib\site-packages\torch/nn/attention/bias.pyrr!s1..`J$&&KKKrrc8eZdZdZdededefdZdejdej fdZ dejdej fd Z dde ejdej fd Z e ddej dej dej dddedede ededej fdZeddZdZd S)ra  A bias representing causal attention patterns. For an overview of the bias structure, see the :class:`CausalVariant` enum. This class is used for defining causal (triangular) attention biases. For construing the bias, there exist two factory functions: :func:`causal_upper_left` and :func:`causal_lower_right`. Example: .. code-block:: python from torch.nn.attention.bias import causal_lower_right bsz, num_heads, seqlen_q, seqlen_kv, head_dim = 32, 8, 4, 12, 8 # Create a lower-right causal bias attn_bias = causal_lower_right(seqlen_q, seqlen_kv) q = torch.randn(bsz, num_heads, seqlen_q, head_dim, device="cuda", dtype=torch.float16) k = torch.randn(bsz, num_heads, seqlen_kv, head_dim, device="cuda", dtype=torch.float16) v = torch.randn(bsz, num_heads, seqlen_kv, head_dim, device="cuda", dtype=torch.float16) out = F.scaled_dot_product_attention(q, k, v, attn_bias) .. warning:: This class is a prototype and subject to change. variant seq_len_q seq_len_kvct|tsJ||_||_||_||kr!|tjkrt ddSdSdS)a Initializes the CausalBias instance with a specified variant and sequence lengths. Args: variant (CausalVariant): The type of causal bias to use (either UPPER_LEFT or LOWER_RIGHT). seq_len_q (int): The sequence length of the query tensor. seq_len_kv (int): The sequence length of the key/value tensor. Raises a warning if the LOWER_RIGHT variant is used with seq_len_q > seq_len_kv, as it may produce NaNs. zTLower right causal bias will produce NaNs in the output when seq_len_q > seq_len_kv!N) isinstancerrr r!rr)selfrr r!s r__init__zCausalBias.__init__qst'=11111 "$ z ! !g1J&J&J f      " !&J&Jrdevicereturnc~tjtj|j|j|tjS)zUpper left causal biasr&dtype)torchtrilonesr r!boolr$r&s r _upper_leftzCausalBias._upper_lefts2z Jt~tvUZ X X X   rc|j|jz }tjtj|j|j|tj|S)zLower right causal biasr))diagonal)r!r r+r,r-r.)r$r&diagonal_offsets r _lower_rightzCausalBias._lower_rightsQ/DN:z Jej   %     rNc|tjd}|jtjkr||S|jtjkr||SdS)a Materializes the causal bias into a tensor form. Depending on the variant, this method generates either an upper-left or lower-right triangular matrix to represent the causal bias. Args: device (Optional[torch.device]): The device on which to create the tensor. Defaults to CPU. Returns: torch.Tensor: The materialized bias tensor. Ncpu)r+r&rrrr0rr4r/s r _materializezCausalBias._materializesh >\%((F <=3 3 3##F++ + \]6 6 6$$V,, ,7 6rFquerykeyvalue attn_mask dropout_p is_causalscale enable_gqac|rtd|j|jks|jtjkrt j|||d|d||S|jtjkr~t|||d|||t|||d|||}t|rD| ddzdk} | d} t| |} | rtjj|dd| ddzz f}tjj|dd| ddzz f}tjj|dd| ddzz f}tjj||||dd| d} t+| | St-|rd} t/|||rd} tjj|d d |d d |d d ddddd|t5|j| |d dd d St7|t j|||||j|d||Std |j)a8 Handles the logic for computing attention with the specified causal bias. Args: query (Tensor): Query tensor; shape :math:`(N, ..., L, E)`. key (Tensor): Key tensor; shape :math:`(N, ..., S, E)`. value (Tensor): Value tensor; shape :math:`(N, ..., S, Ev)`. attn_mask (CausalBias): The type of causal attention to apply. A boolean mask where a value of True indicates that the element *should* take part in attention. A float mask of the same type as query, key, value that is added to the attention score. dropout_p (float): Dropout probability; if greater than 0.0, dropout is applied is_causal (bool): If true, assumes upper left causal attention masking and errors if both attn_mask and is_causal are set. scale (optional float): Scaling factor applied prior to softmax. If None, the default value is set to :math:`\frac{1}{\sqrt{E}}`. enable_gqa (optional bool): If set to True, Grouped Query Attention (GQA) is enabled, by default it is set to False. Returns: output (Tensor): Attention output; shape :math:`(N, ..., L, Ev)`. Raises: ValueError: If the causal bias variant is not a CausalVariant type. z.CausalBias should not be used with causal=TrueNT)r<r=r>r?r@rF)r>return_debug_maskr?) bias cu_seqlens_q cu_seqlens_k max_seqlen_q max_seqlen_kr=custom_mask_typecompute_log_sumexpr?seqlen_kzr?r@ sdpa_params needs_padding og_head_sizeog_scaleoutrMs r _dispatchzCausalBias._dispatchs8F  OMNN N  9#7 7 7 M$<<<1#%     -"; ; ; UD)YPU V V V$sE4IzK'{33 D % 2 2a 7 $zz"~~ +L%@@ X!H/33EAq5::b>>TUCU?U;VWWE(-11#1sxx||a?O;O7PQQC!H/33EAq5::b>>TUCU?U;VWWEinHH"&+"I1lCCC*;77 %*"'sE::.)-&y~BBOOAq))MM!Q''OOAq))!%!%!%!%'%():%;%;'9!CYq!__% '{3335'44U\BB'#)    byO`bb rrcv|i}|tjjjkrt d|j|i|S)zjDefines the behavior of torch.nn.functional.scaled_dot_product_attention when the attn_bias is an AttnBiasNz5CausalBias only supports scaled_dot_product_attention)r+rSrTrQNotImplementedErrorra)clsfunctypesargskwargss r__torch_function__zCausalBias.__torch_function__sO >F 58&C C C%G s}d-f---rcN|SN)r7__repr__)r$s rrlzCausalBias.__repr__$s   ""++---rrk)r8FNF)rN)rrrrrr[r%r+r&Tensorr0r4rr7 staticmethodfloatr.ra classmethodrirlrrrrrVs4 #3( %, 5<     5< EL    --8EL#9-U\----( !% mm|m \m|m m  m  mmm mmm\m^...[......rrr'c|t|dks Jd|\}}ttj||S)a& Creates an upper-left triangular causal bias. This function generates a upper-left triangular matrix to represent causal attention bias with a diagonal offset set so that the inclusive values are aligned to the upper left corner of the matrix. This equivalent to the `is_causal=True` argument in `scaled_dot_product_attention`. The equivalent pytorch code for constructing this bias is: .. code-block:: python torch.tril(torch.ones(size, dtype=torch.bool)) For instance, with `shape=(3,4)`, the materialized bias tensor will be: .. code-block:: text [[1, 0, 0, 0], [1, 1, 0, 0], [1, 1, 1, 0]] Args: size: The size of the bias matrix. Returns: CausalBias: The UPPER_LEFT triangular causal bias variant. rFz*causal_upper_left only supports 2D tensors)lenrrrrRr r!s rrr(s=8 t99>>>G>>> Iz m. : F FFrc|t|dks Jd|\}}ttj||S)a: Creates a lower-right triangular causal bias. This function generates a lower-right triangular matrix to represent causal attention bias with a diagonal offset set so that the inclusive values are aligned to the lower right corner of the matrix. The equivalent pytorch code for constructing this bias is: .. code-block:: python diagonal_offset = size[1] - size[0] torch.tril( torch.ones(size, dtype=torch.bool), diagonal=diagonal_offset, ) For instance, with `shape=(3,4)`, the materialized bias tensor will be: .. code-block:: text [[1, 1, 0, 0], [1, 1, 1, 0], [1, 1, 1, 1]] Args: size: The size of the bias matrix. Returns: CausalBias: The LOWER_RIGHT triangular causal bias variant. rFz+causal_lower_right only supports 2D tensors)rrrrrrss rrrIs=> t99>>>H>>> Iz m/J G GGr)!renumrrtypingrwarningsrr+torch.nn.functionalrSrTrPtorch.backends.cudarrr r torch.nn.attentionr torch.nn.attention._utilsr r rr__all___dynamoallow_in_graphrrmrrrrrrrsII  655555 U T T 9::: 4555 8999 Z(((22222G222jO.O.O.O.O.O.O.O.dG GGGGB!H!H!H!H!H!H!Hr