o hZ4@sdZddlmZmZddlmZddlmZddlZddl m m Z ddl mZmZmZmZddlmZddlmZmZmZmZgd ZejeejeejeejeGd d d eZGd d d ejZdefddZdefddZ 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_RIGHTrrk/var/www/html/construction_image-detection-poc/venv/lib/python3.10/site-packages/torch/nn/attention/bias.pyr!s0 rc@seZdZdZdededefddZdejdej fd d Z dejdej fd d Z d!de ejdej fddZ e   d"dej dej dej dddedede ededej fddZed#ddZdd Zd 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_kvcCsFt|tsJ||_||_||_||kr|tjkr!tddSdSdS)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__qs zCausalBias.__init__devicereturncCsttj|j|j|tjdS)zUpper left causal biasr!dtype)torchtrilonesrrboolrr!rrr _upper_leftszCausalBias._upper_leftcCs.|j|j}tjtj|j|j|tjd|dS)zLower right causal biasr#)diagonal)rrr%r&r'r()rr!diagonal_offsetrrr _lower_rights zCausalBias._lower_rightNcCsB|dur td}|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._materializeFquerykeyvalue attn_mask dropout_p is_causalscale enable_gqacCs|rtd|j|jks|jtjkrtj|||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|||rd} tjjj|d d |d d |d d ddddd|t|j| |dd dd d St|tj|||||j|d||dStd |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)r4r5r6r7r8rF)r6return_debug_maskr7) bias cu_seqlens_q cu_seqlens_k max_seqlen_q max_seqlen_kr5custom_mask_typecompute_log_sumexpr7seqlen_kzs&       5S!