a hq4@sdZddlmZmZddlmZddlmZddlZddl m m Z ddl mZmZmZmZddlmZddlmZmZmZmZgd ZejeejeejeejeGd d d eZGd d d ejZedddZedddZ 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@seZdZdZeZeZdS)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_RIGHTrrE/var/www/auris/lib/python3.9/site-packages/torch/nn/attention/bias.pyr"s0rc seZdZdZeeedddZejej dddZ ejej ddd Z de ejej dd d Z edej ej ej deee eeej d ddZedfdd ZddZZS)raN 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_kvcCs>t|tsJ||_||_||_||kr:|tjkr:tddS)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) isinstancerrrrrr)selfrrrrrr__init__xs zCausalBias.__init__)devicereturncCsttj|j|j|tjdS)zUpper left causal biasr!Zdtype)torchtrilonesrrboolrr!rrr _upper_leftszCausalBias._upper_leftcCs.|j|j}tjtj|j|j|tjd|dS)zLower right causal biasr#)Zdiagonal)rrr$r%r&r')rr!Zdiagonal_offsetrrr _lower_rights zCausalBias._lower_rightNcCsB|durtd}|jtjkr(||S|jtjkr>||SdS)a Materializes the causal bias into a tensor form. Depending on the variant, this method generates either an upper-left or lower-right triangular matrix to represent the causal bias. Args: device (Optional[torch.device]): The device on which to create the tensor. Defaults to CPU. Returns: torch.Tensor: The materialized bias tensor. Ncpu)r$r!rrrr)rr*r(rrr _materializes     zCausalBias._materializeF) querykeyvalue attn_mask dropout_p is_causalscale enable_gqar"cCs|r td|j|jks$|jtjkr>tj|||d|d||dS|jtjkrt |||d|||t |||d|||}t |r:| dddk} | d} t | |} | rtjj|dd| ddf}tjj|dd| ddf}tjj|dd| ddf}tjjj||||dd| d d} t| | St|rd} t|||rZd} tjjj|d d |d d |d d ddddd|t|j| |dd dd d St|tj|||||j|d||dSntd |jdS)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)r1r2r3r4r5rF)r3Zreturn_debug_maskr4) ZbiasZ cu_seqlens_qZ cu_seqlens_kZ max_seqlen_qZ max_seqlen_kr2Zcustom_mask_typecompute_log_sumexpr4Zseqlen_kzrrrrrr-srcGs*t|dksJd|\}}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. r9z+causal_lower_right only supports 2D tensors)rUrrrrVrrrrNsr)!renumrrtypingrwarningsrr$Ztorch.nn.functionalr?r@r<Ztorch.backends.cudarrrr Ztorch.nn.attentionr Ztorch.nn.attention._utilsr r r r__all__Z_dynamoZallow_in_graphrrPrrrrrrrs$       5W!