import math from keras.src import constraints from keras.src import initializers from keras.src import ops from keras.src import regularizers from keras.src.api_export import keras_export from keras.src.backend.config import is_flash_attention_enabled from keras.src.layers.activations.softmax import Softmax from keras.src.layers.core.einsum_dense import EinsumDense from keras.src.layers.layer import Layer from keras.src.layers.regularization.dropout import Dropout @keras_export("keras.layers.GroupQueryAttention") class GroupedQueryAttention(Layer): """Grouped Query Attention layer. This is an implementation of grouped-query attention introduced by [Ainslie et al., 2023](https://arxiv.org/abs/2305.13245). Here `num_key_value_heads` denotes number of groups, setting `num_key_value_heads` to 1 is equivalent to multi-query attention, and when `num_key_value_heads` is equal to `num_query_heads` it is equivalent to multi-head attention. This layer first projects `query`, `key`, and `value` tensors. Then, `key` and `value` are repeated to match the number of heads of `query`. Then, the `query` is scaled and dot-producted with `key` tensors. These are softmaxed to obtain attention probabilities. The value tensors are then interpolated by these probabilities and concatenated back to a single tensor. Args: head_dim: Size of each attention head. num_query_heads: Number of query attention heads. num_key_value_heads: Number of key and value attention heads. dropout: Dropout probability. use_bias: Boolean, whether the dense layers use bias vectors/matrices. flash_attention: If `None`, the layer attempts to use flash attention for faster and more memory-efficient attention computations when possible. This behavior can be configured using `keras.config.enable_flash_attention()` or `keras.config.disable_flash_attention()`. kernel_initializer: Initializer for dense layer kernels. bias_initializer: Initializer for dense layer biases. kernel_regularizer: Regularizer for dense layer kernels. bias_regularizer: Regularizer for dense layer biases. activity_regularizer: Regularizer for dense layer activity. kernel_constraint: Constraint for dense layer kernels. bias_constraint: Constraint for dense layer kernels. seed: Optional integer to seed the dropout layer. Call arguments: query: Query tensor of shape `(batch_dim, target_seq_len, feature_dim)`, where `batch_dim` is batch size, `target_seq_len` is the length of target sequence, and `feature_dim` is dimension of feature. value: Value tensor of shape `(batch_dim, source_seq_len, feature_dim)`, where `batch_dim` is batch size, `source_seq_len` is the length of source sequence, and `feature_dim` is dimension of feature. key: Optional key tensor of shape `(batch_dim, source_seq_len, feature_dim)`. If not given, will use `value` for both `key` and `value`, which is most common case. attention_mask: A boolean mask of shape `(batch_dim, target_seq_len, source_seq_len)`, that prevents attention to certain positions. The boolean mask specifies which query elements can attend to which key elements, where 1 indicates attention and 0 indicates no attention. Broadcasting can happen for the missing batch dimensions and the head dimension. return_attention_scores: A boolean to indicate whether the output should be `(attention_output, attention_scores)` if `True`, or `attention_output` if `False`. Defaults to `False`. training: Python boolean indicating whether the layer should behave in training mode (adding dropout) or in inference mode (no dropout). Will go with either using the training mode of the parent layer/model or `False` (inference) if there is no parent layer. use_causal_mask: A boolean to indicate whether to apply a causal mask to prevent tokens from attending to future tokens (e.g., used in a decoder Transformer). Returns: attention_output: Result of the computation, of shape `(batch_dim, target_seq_len, feature_dim)`, where `target_seq_len` is for target sequence length and `feature_dim` is the query input last dim. attention_scores: (Optional) attention coefficients of shape `(batch_dim, num_query_heads, target_seq_len, source_seq_len)`. """ def __init__( self, head_dim, num_query_heads, num_key_value_heads, dropout=0.0, use_bias=True, flash_attention=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, seed=None, **kwargs, ): super().__init__(**kwargs) self.supports_masking = True self.head_dim = head_dim self.num_query_heads = num_query_heads self.num_key_value_heads = num_key_value_heads if num_query_heads % num_key_value_heads != 0: raise ValueError( "`num_query_heads` must be divisible by `num_key_value_heads`." ) self.num_repeats = num_query_heads // num_key_value_heads self.dropout = dropout self.use_bias = use_bias self._flash_attention = flash_attention or is_flash_attention_enabled() self.kernel_initializer = initializers.get(kernel_initializer) self.bias_initializer = initializers.get(bias_initializer) self.kernel_regularizer = regularizers.get(kernel_regularizer) self.bias_regularizer = regularizers.get(bias_regularizer) self.activity_regularizer = regularizers.get(activity_regularizer) self.kernel_constraint = constraints.get(kernel_constraint) self.bias_constraint = constraints.get(bias_constraint) self.seed = seed self._inverse_sqrt_head_dim = 1.0 / math.sqrt(float(self.head_dim)) self._return_attention_scores = False # Check for flash attention constraints if self._flash_attention and self.dropout > 0.0: raise ValueError( "Dropout is not supported when flash attention is enabled. " "Please set dropout to 0.0 to use flash attention." ) def build( self, query_shape, value_shape, key_shape=None, ): # Einsum variables: # b = batch size # q = query length # k = key/value length # m = model dim # u = num query heads # v = num key/value heads # h = head dim key_shape = value_shape if key_shape is None else key_shape self.feature_dim = query_shape[-1] self._query_dense = EinsumDense( "bqm,muh->bquh", output_shape=(None, self.num_query_heads, self.head_dim), bias_axes="uh" if self.use_bias else None, name="query", **self._get_common_kwargs_for_sublayer(), ) self._query_dense.build(query_shape) self._key_dense = EinsumDense( "bkm,mvh->bkvh", output_shape=(None, self.num_key_value_heads, self.head_dim), bias_axes="vh" if self.use_bias else None, name="key", **self._get_common_kwargs_for_sublayer(), ) self._key_dense.build(key_shape) self._value_dense = EinsumDense( "bkm,mvh->bkvh", output_shape=(None, self.num_key_value_heads, self.head_dim), bias_axes="vh" if self.use_bias else None, name="value", **self._get_common_kwargs_for_sublayer(), ) self._value_dense.build(value_shape) self._softmax = Softmax(axis=-1, dtype=self.dtype_policy) self._dropout_layer = Dropout( rate=self.dropout, dtype=self.dtype_policy, seed=self.seed ) self._dot_product_equation = "bquh,bkuh->buqk" self._combine_equation = "buqk,bkuh->bquh" self._output_dense = EinsumDense( "bquh,uhm->bqm", output_shape=(None, self.feature_dim), bias_axes="m" if self.use_bias else None, name="attention_output", **self._get_common_kwargs_for_sublayer(), ) self._output_dense.build( (None, None, self.num_query_heads, self.head_dim) ) def _get_common_kwargs_for_sublayer(self): common_kwargs = dict( kernel_regularizer=self.kernel_regularizer, bias_regularizer=self.bias_regularizer, activity_regularizer=self.activity_regularizer, kernel_constraint=self.kernel_constraint, bias_constraint=self.bias_constraint, dtype=self.dtype_policy, ) # Create new clone of kernel/bias initializer, so that we don't reuse # the initializer instance, which could lead to same init value since # initializer is stateless. kernel_initializer = self.kernel_initializer.__class__.from_config( self.kernel_initializer.get_config() ) bias_initializer = self.bias_initializer.__class__.from_config( self.bias_initializer.get_config() ) common_kwargs["kernel_initializer"] = kernel_initializer common_kwargs["bias_initializer"] = bias_initializer return common_kwargs def call( self, query, value, key=None, query_mask=None, value_mask=None, key_mask=None, attention_mask=None, return_attention_scores=False, training=None, use_causal_mask=False, ): self._return_attention_scores = return_attention_scores if key is None: key = value attention_mask = self._compute_attention_mask( query, value, query_mask=query_mask, value_mask=value_mask, key_mask=key_mask, attention_mask=attention_mask, use_causal_mask=use_causal_mask, ) query = self._query_dense(query) key = self._key_dense(key) value = self._value_dense(value) key = ops.repeat( key, self.num_repeats, axis=2 ) # (batch_dim, source_seq_len, query_heads, head_dim) value = ops.repeat( value, self.num_repeats, axis=2 ) # (batch_dim, source_seq_len, query_heads, head_dim) output, scores = self._compute_attention( query, key, value, attention_mask=attention_mask, training=training, ) output = self._output_dense( output ) # (batch_dim, target_seq_len, feature_dim) if return_attention_scores: return output, scores return output def _compute_attention_mask( self, query, value, query_mask=None, value_mask=None, key_mask=None, attention_mask=None, use_causal_mask=False, ): """Computes the attention mask, using the Keras masks of the inputs. * The `query`'s mask is reshaped from [B, T] to [B, T, 1]. * The `value`'s mask is reshaped from [B, S] to [B, 1, S]. * The `key`'s mask is reshaped from [B, S] to [B, 1, S]. The `key`'s mask is ignored if `key` is `None` or if `key is value`. * If `use_causal_mask=True`, then the causal mask is computed. Its shape is [1, T, S]. All defined masks are merged using a logical AND operation (`&`). In general, if the `query` and `value` are masked, then there is no need to define the `attention_mask`. Args: query: Projected query tensor of shape `(B, T, N, key_dim)`. key: Projected key tensor of shape `(B, T, N, key_dim)`. value: Projected value tensor of shape `(B, T, N, value_dim)`. attention_mask: a boolean mask of shape `(B, T, S)`, that prevents attention to certain positions. use_causal_mask: A boolean to indicate whether to apply a causal mask to prevent tokens from attending to future tokens (e.g., used in a decoder Transformer). Returns: attention_mask: a boolean mask of shape `(B, T, S)`, that prevents attention to certain positions, based on the Keras masks of the `query`, `key`, `value`, and `attention_mask` tensors, and the causal mask if `use_causal_mask=True`. """ auto_mask = None if query_mask is not None: query_mask = ops.cast(query_mask, "bool") # defensive casting # B = batch size, T = max query length auto_mask = ops.expand_dims(query_mask, -1) # shape is [B, T, 1] if value_mask is not None: value_mask = ops.cast(value_mask, "bool") # defensive casting # B = batch size, S == max value length mask = ops.expand_dims(value_mask, -2) # shape is [B, 1, S] auto_mask = mask if auto_mask is None else auto_mask & mask if key_mask is not None: key_mask = ops.cast(key_mask, "bool") # defensive casting # B == batch size, S == max key length == max value length mask = ops.expand_dims(key_mask, -2) # shape is [B, 1, S] auto_mask = mask if auto_mask is None else auto_mask & mask if use_causal_mask: # the shape of the causal mask is [1, T, S] mask = self._compute_causal_mask(query, value) auto_mask = mask if auto_mask is None else auto_mask & mask if auto_mask is not None: # merge attention_mask & automatic mask, to shape [B, T, S] attention_mask = ( auto_mask if attention_mask is None else ops.cast(attention_mask, bool) & auto_mask ) return attention_mask def _compute_causal_mask(self, query, value=None): """Computes a causal mask (e.g., for masked self-attention layers). For example, if query and value both contain sequences of length 4, this function returns a boolean tensor equal to: ``` [[[True, False, False, False], [True, True, False, False], [True, True, True, False], [True, True, True, True]]] ``` Args: query: query tensor of shape `(B, T, ...)`. value: value tensor of shape `(B, S, ...)` (optional, defaults to query). Returns: mask: a boolean tensor of shape `(1, T, S)` containing a lower triangular matrix of shape `(T, S)`. """ q_seq_length = ops.shape(query)[1] v_seq_length = q_seq_length if value is None else ops.shape(value)[1] ones_mask = ops.ones((1, q_seq_length, v_seq_length), dtype="int32") row_index = ops.cumsum(ones_mask, axis=-2) col_index = ops.cumsum(ones_mask, axis=-1) return ops.greater_equal(row_index, col_index) def _compute_attention( self, query, key, value, attention_mask=None, training=None ): # Check for flash attention constraints if self._flash_attention and self._return_attention_scores: raise ValueError( "Returning attention scores is not supported when flash " "attention is enabled. Please disable flash attention to access" " attention scores." ) # Determine whether to use dot-product attention use_dot_product_attention = not ( self.dropout > 0.0 or self._return_attention_scores or (len(query.shape) != 4) ) if use_dot_product_attention: if attention_mask is not None: # Ensure attention_mask has the correct shape for broadcasting # Expected shape: [batch_size, num_heads, query_seq_len, # key_seq_len]. mask_expansion_axis = -1 * 2 - 1 len_attention_scores_shape = 4 # Only accepts 4D inputs for _ in range( len_attention_scores_shape - len(attention_mask.shape) ): attention_mask = ops.expand_dims( attention_mask, axis=mask_expansion_axis ) attention_mask = ops.cast(attention_mask, dtype="bool") # Directly compute the attention output using dot-product attention attention_output = ops.dot_product_attention( query=query, key=key, value=value, bias=None, mask=attention_mask, scale=self._inverse_sqrt_head_dim, is_causal=False, flash_attention=self._flash_attention, ) return attention_output, None # Default behavior without flash attention, with explicit attention # scores query = ops.multiply( query, ops.cast(self._inverse_sqrt_head_dim, query.dtype) ) # Take the dot product between "query" and "key" to get the raw # attention scores. scores = ops.einsum( self._dot_product_equation, query, key ) # (batch_dim, query_heads, target_seq_len, source_seq_len) scores = self._masked_softmax(scores, attention_mask=attention_mask) # This is actually dropping out entire tokens to attend to, which might # seem a bit unusual, but is taken from the original Transformer paper. if self.dropout > 0.0: scores_dropout = self._dropout_layer(scores, training=training) else: scores_dropout = scores output = ops.einsum(self._combine_equation, scores_dropout, value) return output, scores def _masked_softmax(self, scores, attention_mask=None): # Normalize the attention scores to probabilities. # scores = [B, N, T, S] if attention_mask is not None: # The expand dim happens starting from the `num_heads` dimension, # (, num_heads, ) mask_expansion_axis = -1 * 2 - 1 for _ in range(len(scores.shape) - len(attention_mask.shape)): attention_mask = ops.expand_dims( attention_mask, axis=mask_expansion_axis ) return self._softmax(scores, mask=attention_mask) def compute_output_shape( self, query_shape, value_shape, key_shape=None, ): if key_shape is None: key_shape = value_shape if query_shape[-1] != value_shape[-1]: raise ValueError( "The last dimension of `query_shape` and `value_shape` " f"must be equal, but are {query_shape[-1]}, {value_shape[-1]}. " f"Received: query_shape={query_shape}, " f"value_shape={value_shape}" ) if value_shape[1:-1] != key_shape[1:-1]: raise ValueError( "All dimensions of `value` and `key`, except the last one, " f"must be equal. Received: value_shape={value_shape} and " f"key_shape={key_shape}" ) return query_shape def get_config(self): config = { "head_dim": self.head_dim, "num_query_heads": self.num_query_heads, "num_key_value_heads": self.num_key_value_heads, "use_bias": self.use_bias, "dropout": self.dropout, "kernel_initializer": initializers.serialize( self.kernel_initializer ), "bias_initializer": initializers.serialize(self.bias_initializer), "kernel_regularizer": regularizers.serialize( self.kernel_regularizer ), "bias_regularizer": regularizers.serialize(self.bias_regularizer), "activity_regularizer": regularizers.serialize( self.activity_regularizer ), "kernel_constraint": constraints.serialize(self.kernel_constraint), "bias_constraint": constraints.serialize(self.bias_constraint), "seed": self.seed, } base_config = super().get_config() return {**base_config, **config}