# coding=utf-8 # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. # # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX # and OPT implementations in this library. It has been modified from its # original forms to accommodate minor architectural differences compared # to GPT-NeoX and OPT used by the Meta AI team that trained the model. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """TF 2.0 Idefics model.""" from __future__ import annotations from dataclasses import dataclass from typing import List, Optional, Tuple, Union import tensorflow as tf from ... import TFPreTrainedModel from ...activations_tf import get_tf_activation from ...modeling_tf_outputs import ModelOutput from ...modeling_tf_utils import ( TFCausalLanguageModelingLoss, TFModelInputType, keras_serializable, shape_list, unpack_inputs, ) from ...tf_utils import invert_attention_mask, scaled_dot_product_attention from ...utils import ( add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings, ) from .configuration_idefics import IdeficsConfig from .perceiver_tf import TFIdeficsPerceiverResampler from .vision_tf import TFIdeficsVisionTransformer logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = "IdeficsConfig" @dataclass class TFIdeficsBaseModelOutputWithPast(ModelOutput): """ Base class for Idefics model's outputs that may also contain a past key/values (to speed up sequential decoding). Args: last_hidden_state (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1, hidden_size)` is output. past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if `config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. image_hidden_states (`tuple(tf.Tensor)`, *optional*): Tuple of `tf.Tensor` (one for the output of the image embeddings, `(batch_size, num_images, sequence_length, hidden_size)`. image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver """ last_hidden_state: Optional[tf.Tensor] = None past_key_values: Optional[Tuple[Tuple[tf.Tensor]]] = None hidden_states: Optional[Tuple[tf.Tensor]] = None attentions: Optional[Tuple[tf.Tensor]] = None image_hidden_states: Optional[Tuple[tf.Tensor]] = None @dataclass class TFIdeficsCausalLMOutputWithPast(ModelOutput): """ Base class for Idefics causal language model (or autoregressive) outputs. Args: loss (`tf.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token prediction). logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. image_hidden_states (`tuple(tf.Tensor)`, *optional*): Tuple of `tf.Tensor` (one for the output of the image embeddings, `(batch_size, num_images, sequence_length, hidden_size)`. image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver """ loss: Optional[tf.Tensor] = None logits: Optional[tf.Tensor] = None past_key_values: Optional[List[tf.Tensor]] = None hidden_states: Optional[Tuple[tf.Tensor]] = None attentions: Optional[Tuple[tf.Tensor]] = None image_hidden_states: Optional[Tuple[tf.Tensor]] = None def expand_inputs_for_generation( input_ids, expand_size=1, is_encoder_decoder=False, attention_mask=None, encoder_outputs=None, **model_kwargs, ): expanded_return_idx = tf.reshape(tf.repeat(tf.range(tf.shape(input_ids)[0]), expand_size), [-1]) input_ids = tf.gather(input_ids, expanded_return_idx) model_kwargs["pixel_values"] = model_kwargs.get("pixel_values", None) model_kwargs["image_encoder_embeddings"] = model_kwargs.get("image_encoder_embeddings", None) model_kwargs["perceiver_embeddings"] = model_kwargs.get("perceiver_embeddings", None) model_kwargs["image_attention_mask"] = model_kwargs.get("image_attention_mask", None) if "token_type_ids" in model_kwargs: token_type_ids = model_kwargs["token_type_ids"] model_kwargs["token_type_ids"] = tf.gather(token_type_ids, expanded_return_idx) if attention_mask is not None: model_kwargs["attention_mask"] = tf.gather(attention_mask, expanded_return_idx) if model_kwargs["image_attention_mask"] is not None: model_kwargs["image_attention_mask"] = tf.gather(model_kwargs["image_attention_mask"], expanded_return_idx) if model_kwargs["pixel_values"] is not None: model_kwargs["pixel_values"] = tf.gather(model_kwargs["pixel_values"], expanded_return_idx) elif model_kwargs["image_encoder_embeddings"] is not None: model_kwargs["image_encoder_embeddings"] = tf.gather( model_kwargs["image_encoder_embeddings"], expanded_return_idx ) elif model_kwargs["perceiver_embeddings"] is not None: model_kwargs["perceiver_embeddings"] = tf.gather(model_kwargs["perceiver_embeddings"], expanded_return_idx) return input_ids, model_kwargs def update_model_kwargs_for_generation(outputs, model_kwargs): # must have this key set to at least None if "past_key_values" in outputs: model_kwargs["past_key_values"] = outputs.past_key_values else: model_kwargs["past_key_values"] = None # update token_type_ids with last value if "token_type_ids" in model_kwargs: token_type_ids = model_kwargs["token_type_ids"] model_kwargs["token_type_ids"] = tf.concat([token_type_ids, token_type_ids[:, -1:, ...]], axis=-1) # update attention masks if "attention_mask" in model_kwargs: attention_mask = model_kwargs["attention_mask"] model_kwargs["attention_mask"] = tf.concat( [attention_mask, tf.ones_like(attention_mask[:, -1:, ...])], axis=-1 ) if "image_attention_mask" in model_kwargs: image_attention_mask = model_kwargs["image_attention_mask"] last_mask = image_attention_mask[:, -1:, ...] model_kwargs["image_attention_mask"] = last_mask # Get the precomputed image_hidden_states model_kwargs["image_hidden_states"] = outputs.image_hidden_states return model_kwargs def prepare_inputs_for_generation(input_ids, past_key_values=None, **kwargs): token_type_ids = kwargs.get("token_type_ids", None) # only last token for inputs_ids if past is defined in kwargs if past_key_values is not None: input_ids = input_ids[:, -1:] if token_type_ids is not None: token_type_ids = token_type_ids[:, -1:] attention_mask = kwargs.get("attention_mask", None) position_ids = kwargs.get("position_ids", None) if attention_mask is not None and position_ids is None: # create position_ids on the fly for batch generation position_ids = tf.math.cumsum(tf.cast(attention_mask, dtype=tf.int64), axis=-1) - 1 position_ids = tf.where(attention_mask == 0, 1, position_ids) if past_key_values is not None: position_ids = position_ids[:, -1:] pixel_values = kwargs.get("pixel_values", None) image_encoder_embeddings = kwargs.get("image_encoder_embeddings", None) perceiver_embeddings = kwargs.get("perceiver_embeddings", None) image_attention_mask = kwargs.get("image_attention_mask", None) interpolate_pos_encoding = kwargs.get("interpolate_pos_encoding", False) return { "input_ids": input_ids, "past_key_values": past_key_values, "use_cache": kwargs.get("use_cache"), "position_ids": position_ids, "attention_mask": attention_mask, "token_type_ids": token_type_ids, "pixel_values": pixel_values, "image_encoder_embeddings": image_encoder_embeddings, "perceiver_embeddings": perceiver_embeddings, "image_attention_mask": image_attention_mask, "interpolate_pos_encoding": interpolate_pos_encoding, } def freeze_model(model, module_exceptions=[]): mapping = { "LayerNorm": tf.keras.layers.LayerNormalization, "Dense": tf.keras.layers.Dense, "Embedding": tf.keras.layers.Embedding, } module_exceptions_mapped = [mapping[m] for m in module_exceptions] if not hasattr(model, "layers"): model.trainable = False # It is just a layer return model for layer in model.layers: if module_exceptions and any(isinstance(layer, t) for t in module_exceptions_mapped): layer.trainable = True # Explicitly setting it to true to avoid any mistakes else: layer.trainable = False return model class TFIdeficsDecoupledEmbedding(tf.keras.layers.Embedding): """ Implements a decoupling of parameters to allow freezing (or not) a subset of the embeddings. In practise, the regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `num_additional_embeddings` > 0, then it will create `num_additional_embeddings` additional parameters that are always trained. If `num_additional_embeddings=0`, then the module defaults back to the regular behavior of `tf.keras.layers.Embedding`. """ def __init__( self, num_embeddings, num_additional_embeddings, embedding_dim, partially_freeze: Optional[bool] = False, dtype=None, **kwargs, ) -> None: """ Args: num_embeddings (`int`): Size of the dictionary of embeddings num_additional_embeddings (`int`): Number of additional embeddings. Only useful when you `partially_freeze=True`. embedding_dim (`int`): The size of each embedding vector partially_freeze: (`bool`, *optional*, defaults to `False`): If `True`, the regular `weight` will be frozen. `additional_weight` is never frozen. Note: there are a lot of other parameters to initialize a standard `tf.keras.layers.Embedding` such as `mask_zero`, `input_length` or `embeddings_initializer`. We are not supporting these. """ super().__init__( input_dim=num_embeddings, output_dim=embedding_dim, dtype=dtype, **kwargs, ) self.num_embeddings = num_embeddings self.num_additional_embeddings = num_additional_embeddings self.partially_freeze = partially_freeze if partially_freeze: self.trainable = False if self.num_additional_embeddings > 0: self.additional_embedding = tf.keras.layers.Embedding( input_dim=self.num_additional_embeddings, output_dim=embedding_dim, dtype=dtype, name="additional_embedding", ) def call(self, input_ids): """ we have 2 embeddings, with different indices - one pretrained self.weight and another self.additional_embedding.weight that is being trained. in order to make a lookup of the input ids, we: 1. find out the indices of the entries belonging to the 2nd embedding 2. extract those values while subtracting the size of the first embedding (num_embeddings), since the 2nd embedding starts from 0 and not num_embeddings 3. perform the 2nd embedding lookup 4. now we handle the 1st embedding, we overwrite indices belonging to the 2nd embedding with a padding index 5. perform the 1st embedding lookup 6. now we overwrite the values in the 1st embedding lookup with the values of the 2nd embedding lookup note: for the 1st embedding lookup we could have looked up only the low indices and not do the padding, but then we have to create a new tensor and populate it with 2 tensors that are spread out across various indices - i.e. not a simple concat - I haven't benchmarked the complex case if it's any faster, given that seqlens are usually relatively short it's probably not faster or if faster not by much - but might be a good idea to measure. """ if self.num_additional_embeddings == 0: return super().call(input_ids) # Clone so that we don't modify the original input_ids later on input_ids = tf.identity(input_ids) additional_vocab_indices = tf.where(input_ids >= self.num_embeddings) input_ids_additional_vocab = tf.gather_nd(input_ids, additional_vocab_indices) additional_embeddings = self.additional_embedding(input_ids_additional_vocab - self.num_embeddings) # for successful lookup replace input_ids with 0, the results of these will be discarded anyway input_ids = tf.tensor_scatter_nd_update( input_ids, additional_vocab_indices, # tensor filled with 0, having the same length as additional_vocab_indices tf.zeros(tf.shape(additional_vocab_indices)[0], dtype=input_ids.dtype), ) full_vector = super().call(input_ids) # overwrite the records with high indices full_vector = tf.tensor_scatter_nd_update(full_vector, additional_vocab_indices, additional_embeddings) return full_vector def extra_repr(self) -> str: return "num_embeddings={}, num_additional_embeddings={}, embedding_dim={}, partially_freeze={}".format( self.num_embeddings, self.num_additional_embeddings, self.output_dim, self.partially_freeze, ) class TFIdeficsDecoupledLinear(tf.keras.layers.Layer): """ Implements a decoupling of parameters to allow freezing (or not) a subset of the parameters. In practise, the regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `out_additional_features` > 0, then it will create `out_additional_features * in_features` additional parameters that are always trained. If `out_additional_features=0`, then the module defaults back to the regular behavior of `tf.keras.layers.Dense`. """ def __init__( self, in_features: int, out_features: int, out_additional_features: int = 0, bias: bool = True, partially_freeze: bool = True, **kwargs, ) -> None: """ out_additional_features: int. Number of additional trainable dimensions. Only makes sense when `partially_freeze=True`. partially_freeze: bool. If True, the regular `weight` will be frozen and extra parameters (if any) will be trainable. If False, default to the regular behavior of tf.keras.layers.Dense. """ super().__init__(**kwargs) self.out_additional_features = out_additional_features self.partially_freeze = partially_freeze self.in_features = in_features self.out_features = out_features self.use_bias = bias if out_additional_features > 0: self.additional_fc = tf.keras.layers.Dense( units=out_additional_features, use_bias=bias, name="additional_fc" ) def call(self, inputs: tf.Tensor) -> tf.Tensor: output = tf.linalg.matmul(a=inputs, b=self.weight, transpose_b=True) if self.bias is not None: output = tf.nn.bias_add(output, self.bias) if self.out_additional_features > 0: additional_features = self.additional_fc(inputs) output = tf.concat([output, additional_features], axis=-1) return output def get_config(self): config = super().get_config() config.update( { "in_features": self.in_features, "out_features": self.out_features, "out_additional_features": self.out_additional_features, "bias": self.bias is not None, "partially_freeze": self.partially_freeze, } ) return config def extra_repr(self) -> str: """Overwriting `nn.Linear.extra_repr` to include new parameters.""" return "in_features={}, out_features={}, out_additional_features={}, bias={}, partially_freeze={}".format( self.in_features, self.out_features, self.out_additional_features, self.bias is not None, self.partially_freeze, ) @classmethod def from_config(cls, config): return cls(**config) def build(self, input_shape=None): if self.built: return self.built = True self.weight = self.add_weight( shape=(self.out_features, self.in_features), trainable=not self.partially_freeze, name="weight" ) if self.use_bias: self.bias = self.add_weight(shape=(self.out_features,), trainable=not self.partially_freeze, name="bias") else: self.bias = None if getattr(self, "additional_fc", None) is not None: with tf.name_scope(self.additional_fc.name): self.additional_fc.build(self.in_features) def _make_causal_mask(input_ids_shape, dtype, past_key_values_length=0): """ Make causal mask used for bi-directional self-attention, supporting both static and dynamic shapes. """ bsz, tgt_len = input_ids_shape # Create a matrix where only the lower triangle and diagonal are filled with zeros (causal mask) mask = tf.fill((tgt_len, tgt_len), tf.dtypes.as_dtype(dtype).min) mask_cond = tf.range(tgt_len) mask = tf.where(mask_cond[:, None] >= mask_cond[None, :], 0.0, mask) if past_key_values_length > 0: mask = tf.concat([tf.zeros((tgt_len, past_key_values_length), dtype=dtype), mask], axis=-1) if bsz is None: # When batch size is dynamic, expand and tile # so we can compile a functional model mask = tf.expand_dims(mask, 0) mask = tf.expand_dims(mask, 0) # shape: (1, 1, tgt_len, tgt_len + past_key_values_length) mask = tf.tile(mask, [bsz, 1, 1, 1]) else: # When batch size is static, directly use broadcast_to mask = tf.broadcast_to(mask[None, None, :, :], (bsz, 1, tgt_len, tgt_len + past_key_values_length)) return mask def _expand_mask(mask, dtype, tgt_len=None): """ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. """ bsz, src_len = shape_list(mask) tgt_len = tgt_len if tgt_len is not None else src_len expanded_mask = tf.expand_dims(tf.expand_dims(mask, 1), 1) expanded_mask = tf.broadcast_to(expanded_mask, [bsz, 1, tgt_len, src_len]) inverted_mask = 1.0 - tf.cast(expanded_mask, dtype) return tf.where( tf.cast(inverted_mask, bool), tf.fill(dims=shape_list(inverted_mask), value=tf.float32.min), inverted_mask ) class TFIdeficsRMSNorm(tf.keras.layers.Layer): def __init__(self, hidden_size, eps=1e-6, **kwargs): """ TFIdeficsRMSNorm is equivalent to T5LayerNorm """ super().__init__(**kwargs) self.hidden_size = hidden_size self.variance_epsilon = eps def build(self, input_shape): if self.built: return self.built = True self.weight = self.add_weight(name="weight", shape=[self.hidden_size], initializer="ones") super().build(input_shape) def call(self, hidden_states): variance = tf.math.reduce_mean(tf.math.square(tf.cast(hidden_states, tf.float32)), axis=-1, keepdims=True) hidden_states = hidden_states * tf.math.rsqrt(variance + self.variance_epsilon) # convert into half-precision if necessary if self.weight.dtype in [tf.float16, tf.bfloat16]: hidden_states = tf.cast(hidden_states, self.weight.dtype) return self.weight * hidden_states class TFIdeficsEmbedding(tf.keras.layers.Layer): def __init__(self, dim, max_position_embeddings=2048, base=10000, **kwargs): super().__init__(**kwargs) self.dim = dim self.max_position_embeddings = max_position_embeddings self.base = base self.inv_freq = tf.constant( 1.0 / (self.base ** (tf.range(start=0, limit=self.dim, delta=2, dtype=tf.float32) / self.dim)) ) def _compute_cos_sin(self, seq_len): t = tf.range(seq_len, dtype=self.inv_freq.dtype) freqs = tf.einsum("i, j -> ij", t, self.inv_freq) # Outer multiplication emb = tf.concat((freqs, freqs), axis=-1) return tf.cos(emb), tf.sin(emb) def call(self, x, seq_len=None): # x: [bs, num_attention_heads, seq_len, head_size] if seq_len is None: seq_len = shape_list(x)[2] return self._compute_cos_sin(seq_len=seq_len) def rotate_half(x): """Rotates half the hidden dims of the input.""" x1 = x[..., : x.shape[-1] // 2] x2 = x[..., x.shape[-1] // 2 :] return tf.concat((-x2, x1), axis=-1) def apply_rotary_pos_emb(q, k, cos, sin, position_ids): cos = tf.gather(cos, position_ids) # [seq_len, dim] -> [batch_size, 1, seq_len, head_dim] sin = tf.gather(sin, position_ids) cos = tf.expand_dims(cos, 1) sin = tf.expand_dims(sin, 1) q_embed = (q * cos) + (rotate_half(q) * sin) k_embed = (k * cos) + (rotate_half(k) * sin) return q_embed, k_embed class TFIdeficsMLP(tf.keras.layers.Layer): def __init__( self, hidden_size: int, intermediate_size: int, hidden_act: str, **kwargs, ): super().__init__(**kwargs) self.gate_proj = tf.keras.layers.Dense(intermediate_size, use_bias=False, name="gate_proj") self.down_proj = tf.keras.layers.Dense(hidden_size, use_bias=False, name="down_proj") self.up_proj = tf.keras.layers.Dense(intermediate_size, use_bias=False, name="up_proj") self.act_fn = get_tf_activation(hidden_act) self.intermediate_size = intermediate_size self.hidden_size = hidden_size def call(self, x): return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "gate_proj", None) is not None: with tf.name_scope(self.gate_proj.name): self.gate_proj.build(self.hidden_size) if getattr(self, "down_proj", None) is not None: with tf.name_scope(self.down_proj.name): self.down_proj.build(self.intermediate_size) if getattr(self, "up_proj", None) is not None: with tf.name_scope(self.up_proj.name): self.up_proj.build(self.hidden_size) class TFIdeficsAttention(tf.keras.layers.Layer): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__( self, hidden_size: int, num_heads: int, dropout: float = 0.0, is_cross_attention: bool = False, config: IdeficsConfig = None, qk_layer_norms: bool = False, **kwargs, ): super().__init__(**kwargs) self.hidden_size = hidden_size self.num_heads = num_heads self.head_dim = hidden_size // num_heads self.dropout = dropout self.config = config self.is_causal = True if (self.head_dim * num_heads) != self.hidden_size: raise ValueError( f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}" f" and `num_heads`: {num_heads})." ) self.is_cross_attention = is_cross_attention self.q_proj = tf.keras.layers.Dense( num_heads * self.head_dim, use_bias=False, name="q_proj", ) self.k_proj = tf.keras.layers.Dense( num_heads * self.head_dim, use_bias=False, name="k_proj", ) self.v_proj = tf.keras.layers.Dense( num_heads * self.head_dim, use_bias=False, name="v_proj", ) self.o_proj = tf.keras.layers.Dense( hidden_size, use_bias=False, name="o_proj", ) self.rotary_emb = TFIdeficsEmbedding(self.head_dim, name="rotary_emb") self.qk_layer_norms = qk_layer_norms if self.qk_layer_norms: self.q_layer_norm = TFIdeficsRMSNorm(self.head_dim, eps=config.rms_norm_eps, name="q_layer_norm") self.k_layer_norm = TFIdeficsRMSNorm(self.head_dim, eps=config.rms_norm_eps, name="k_layer_norm") def _shape(self, tensor: tf.Tensor, seq_len: int, bsz: int): return tf.transpose(tf.reshape(tensor, (bsz, seq_len, self.num_heads, self.head_dim)), perm=[0, 2, 1, 3]) def call( self, hidden_states: tf.Tensor, key_value_states: Optional[tf.Tensor] = None, attention_mask: Optional[tf.Tensor] = None, position_ids: Optional[tf.Tensor] = None, past_key_value: Optional[Tuple[tf.Tensor]] = None, output_attentions: bool = False, use_cache: bool = False, ) -> Tuple[tf.Tensor, Optional[tf.Tensor], Optional[Tuple[tf.Tensor]]]: # if key_value_states are provided this layer is used as a cross-attention layer is_cross_attention = self.is_cross_attention or key_value_states is not None bsz, q_len, _ = shape_list(hidden_states) query_states = self._shape(self.q_proj(hidden_states), q_len, bsz) if not is_cross_attention: key_states = self._shape(self.k_proj(hidden_states), q_len, bsz) value_states = self._shape(self.v_proj(hidden_states), q_len, bsz) else: _, kv_len, _ = shape_list(key_value_states) # Note that, in this case, `kv_len` == `kv_seq_len` key_states = self._shape(self.k_proj(key_value_states), kv_len, bsz) value_states = self._shape(self.v_proj(key_value_states), kv_len, bsz) kv_seq_len = shape_list(key_states)[-2] if past_key_value is not None: kv_seq_len += shape_list(past_key_value[0])[-2] if not is_cross_attention: # Below is to allow symbolic tensors compilation if tf.is_tensor(kv_seq_len): seq_len = tf.reduce_max(kv_seq_len, q_len) else: seq_len = max(kv_seq_len, q_len) cos, sin = self.rotary_emb(value_states, seq_len) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) # [bsz, nh, t, hd] if past_key_value is not None: # reuse k, v, self_attention key_states = tf.concat([past_key_value[0], key_states], axis=2) value_states = tf.concat([past_key_value[1], value_states], axis=2) past_key_value = (key_states, value_states) if use_cache else None if self.qk_layer_norms: query_states = self.q_layer_norm(query_states) key_states = self.k_layer_norm(key_states) tf.debugging.assert_equal( tf.shape(attention_mask), [bsz, 1, q_len, kv_seq_len], message=f"Attention weights should be of size {[bsz, 1, q_len, kv_seq_len]}, but is {tf.shape(attention_mask)}", ) attn_output = scaled_dot_product_attention( query_states, key_states, value_states, attn_mask=attention_mask, # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. is_causal=self.is_causal and attention_mask is None and q_len > 1, ) tf.debugging.assert_equal( tf.shape(attn_output), [bsz, self.num_heads, q_len, self.head_dim], message=f"Attention weights should be of size {[bsz, self.num_heads, q_len, self.head_dim]}, but is {tf.shape(attn_output)}", ) attn_output = tf.reshape(tf.transpose(attn_output, perm=[0, 2, 1, 3]), (bsz, q_len, self.hidden_size)) attn_output = self.o_proj(attn_output) attn_weights = None if output_attentions: logger.warning_once( "attn_weights are not extracted in scaled_dot_product_attention. The model returns None instead" ) return attn_output, attn_weights, past_key_value def build(self, input_shape=None): if self.built: return self.built = True if self.is_cross_attention: kv_input_dim = ( self.hidden_size if not hasattr(self.config.vision_config, "embed_dim") else self.config.vision_config.embed_dim ) else: kv_input_dim = self.hidden_size if getattr(self, "o_proj", None) is not None: with tf.name_scope(self.o_proj.name): self.o_proj.build(self.num_heads * self.head_dim) if getattr(self, "q_proj", None) is not None: with tf.name_scope(self.q_proj.name): self.q_proj.build(self.hidden_size) if getattr(self, "k_proj", None) is not None: with tf.name_scope(self.k_proj.name): self.k_proj.build(kv_input_dim) if getattr(self, "v_proj", None) is not None: with tf.name_scope(self.v_proj.name): self.v_proj.build(kv_input_dim) if getattr(self, "rotary_emb", None) is not None: with tf.name_scope(self.rotary_emb.name): self.rotary_emb.build(None) class TFIdeficsDecoderLayer(tf.keras.layers.Layer): def __init__(self, config: IdeficsConfig, **kwargs): super().__init__(**kwargs) self.hidden_size = config.hidden_size self.self_attn = TFIdeficsAttention( hidden_size=self.hidden_size, num_heads=config.num_attention_heads, dropout=config.dropout, config=config, name="self_attn", ) self.mlp = TFIdeficsMLP( hidden_size=self.hidden_size, intermediate_size=config.intermediate_size, hidden_act=config.hidden_act, name="mlp", ) self.input_layernorm = TFIdeficsRMSNorm(config.hidden_size, eps=config.rms_norm_eps, name="input_layernorm") self.post_attention_layernorm = TFIdeficsRMSNorm( config.hidden_size, eps=config.rms_norm_eps, name="post_attention_layernorm" ) self.dropout = config.dropout def call( self, hidden_states: tf.Tensor, attention_mask: Optional[tf.Tensor] = None, position_ids: Optional[tf.Tensor] = None, past_key_value: Optional[Tuple[tf.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, training=False, ) -> Tuple[tf.Tensor, Optional[Tuple[tf.Tensor, tf.Tensor]]]: """ Args: hidden_states (`tf.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`tf.Tensor`, *optional*): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(tf.Tensor)`, *optional*): cached past key and value projection states """ residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # Self Attention hidden_states, self_attn_weights, present_key_value = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, ) hidden_states = tf.nn.dropout(hidden_states, rate=self.dropout) hidden_states = residual + hidden_states # Fully Connected residual = hidden_states hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = tf.nn.dropout(hidden_states, rate=self.dropout) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) return outputs def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "self_attn", None) is not None: with tf.name_scope(self.self_attn.name): self.self_attn.build(None) if getattr(self, "mlp", None) is not None: with tf.name_scope(self.mlp.name): self.mlp.build(None) if getattr(self, "input_layernorm", None) is not None: with tf.name_scope(self.input_layernorm.name): self.input_layernorm.build(None) if getattr(self, "post_attention_layernorm", None) is not None: with tf.name_scope(self.post_attention_layernorm.name): self.post_attention_layernorm.build(None) class TFIdeficsGatedCrossAttentionLayer(tf.keras.layers.Layer): def __init__(self, config: IdeficsConfig, **kwargs): super().__init__(**kwargs) self.hidden_size = config.hidden_size self.cross_attn = TFIdeficsAttention( hidden_size=self.hidden_size, num_heads=config.num_attention_heads, is_cross_attention=True, dropout=config.dropout, config=config, qk_layer_norms=config.qk_layer_norms, name="cross_attn", ) self.mlp = TFIdeficsMLP( hidden_size=self.hidden_size, intermediate_size=config.intermediate_size, hidden_act=config.hidden_act, name="mlp", ) self.input_layernorm = TFIdeficsRMSNorm(config.hidden_size, eps=config.rms_norm_eps, name="input_layernorm") self.post_attention_layernorm = TFIdeficsRMSNorm( config.hidden_size, eps=config.rms_norm_eps, name="post_attention_layernorm" ) self.config = config.dropout self.act_cross_attn = tf.keras.activations.tanh self.act_dense = tf.keras.activations.tanh self.alpha_initializer = config.alpha_initializer self.alpha_type = config.alpha_type self.alphas_initializer_range = config.alphas_initializer_range def build(self, input_shape): if self.built: return self.built = True if self.alpha_initializer == "zeros": if self.alpha_type == "vector": self.alpha_cross_attn = self.add_weight( shape=(1, 1, self.hidden_size), initializer="zeros", trainable=True, name="alpha_cross_attn" ) self.alpha_dense = self.add_weight( shape=(1, 1, self.hidden_size), initializer="zeros", trainable=True, name="alpha_dense" ) elif self.alpha_type == "float": self.alpha_cross_attn = self.add_weight( shape=(1,), initializer="zeros", trainable=True, name="alpha_cross_attn" ) self.alpha_dense = self.add_weight(shape=(1,), initializer="zeros", trainable=True, name="alpha_dense") else: raise ValueError(f"Unknown value for `alpha_type` ({self.alpha_type})") elif self.alpha_initializer == "ones": if self.alpha_type == "vector": self.alpha_cross_attn = self.add_weight( shape=(1, 1, self.hidden_size), initializer="ones", trainable=True, name="alpha_cross_attn" ) self.alpha_dense = self.add_weight( shape=(1, 1, self.hidden_size), initializer="ones", trainable=True, name="alpha_dense" ) elif self.alpha_type == "float": self.alpha_cross_attn = self.add_weight( shape=(1,), initializer="ones", trainable=True, name="alpha_cross_attn" ) self.alpha_dense = self.add_weight(shape=(1,), initializer="ones", trainable=True, name="alpha_dense") else: raise ValueError(f"Unknown value for `alpha_type` ({self.alpha_type})") elif self.alpha_initializer in {"normal", "gaussian", "random"}: if self.alpha_type == "vector": self.alpha_cross_attn = self.add_weight( shape=(1, 1, self.hidden_size), initializer=tf.keras.initializers.RandomNormal(mean=0.0, stddev=self.alphas_initializer_range), trainable=True, name="alpha_cross_attn", ) self.alpha_dense = self.add_weight( shape=(1, 1, self.hidden_size), initializer=tf.keras.initializers.RandomNormal(mean=0.0, stddev=self.alphas_initializer_range), trainable=True, name="alpha_dense", ) elif self.alpha_type == "float": self.alpha_cross_attn = self.add_weight( shape=(1,), initializer=tf.keras.initializers.RandomNormal(mean=0.0, stddev=self.alphas_initializer_range), trainable=True, name="alpha_type", ) self.alpha_dense = self.add_weight( shape=(1,), initializer=tf.keras.initializers.RandomNormal(mean=0.0, stddev=self.alphas_initializer_range), trainable=True, name="alpha_dense", ) else: raise ValueError(f"Unknown value for `alpha_type` ({self.alpha_type})") else: raise NotImplementedError(f"Alpha initialization scheme {self.alpha_initializer} not yet implemented!") if not (hasattr(self, "alpha_cross_attn") and hasattr(self, "alpha_dense")): raise ValueError("Alpha parameters not initialized correctly!") with tf.name_scope(self.cross_attn.name): self.cross_attn.build(None) with tf.name_scope(self.mlp.name): self.mlp.build(None) with tf.name_scope(self.input_layernorm.name): self.input_layernorm.build(None) with tf.name_scope(self.post_attention_layernorm.name): self.post_attention_layernorm.build(None) super().build(input_shape) def call( self, hidden_states: tf.Tensor, attention_mask: Optional[tf.Tensor] = None, image_hidden_states: Optional[tf.Tensor] = None, image_attention_mask: Optional[tf.Tensor] = None, cross_attention_gate: Optional[tf.Tensor] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, past_key_value: Optional[Tuple[tf.Tensor]] = None, ) -> Tuple[tf.Tensor, Optional[Tuple[tf.Tensor, tf.Tensor]]]: """ Args: hidden_states (`tf.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`tf.Tensor`, *optional*): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(tf.Tensor)`, *optional*): cached past key and value projection states no_images (`bool`, *optional*, defaults to `False`): If `True` the vision part is ignored """ if image_hidden_states is None: raise ValueError( "`image_hidden_states` is required for Idefics cross attention module which are visual features to be" " conditioned on." ) if cross_attention_gate is None: raise ValueError( "`cross_attention_gate` is required for Idefics cross attention module to zero-out the cross-attention hidden_states attending to no images." ) if past_key_value is not None: raise NotImplementedError("Past key value states are not implemented for Idefics cross attention module.") residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # Self Attention hidden_states, self_attn_weights, present_key_value = self.cross_attn( hidden_states=hidden_states, key_value_states=image_hidden_states, attention_mask=image_attention_mask, output_attentions=output_attentions, ) hidden_states = tf.nn.dropout(hidden_states, rate=self.config) mask = tf.cast(cross_attention_gate == 0, dtype=hidden_states.dtype) # Expand dimensions of mask to match hidden_states mask = tf.expand_dims(mask, -1) hidden_states = tf.where( tf.broadcast_to(mask, tf.shape(hidden_states)) == 1, tf.zeros_like(hidden_states), hidden_states ) # when there are no images the model is used in pure language mode # gate = 0 if no_images else 1 hidden_states = residual + self.act_cross_attn(self.alpha_cross_attn) * hidden_states # Fully Connected residual = hidden_states hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = tf.nn.dropout(hidden_states, rate=self.config) hidden_states = residual + self.act_dense(self.alpha_dense) * hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) return outputs LLAMA_START_DOCSTRING = r""" This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a TensorFlow [tf.keras.layers.Layer](https://www.tensorflow.org/api_docs/python/tf/keras/layers/Layer) subclass. Use it as a regular TensorFlow Layer and refer to the TensorFlow documentation for all matter related to general usage and behavior. Parameters: config ([`IdeficsConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~TFPreTrainedModel.from_pretrained`] method to load the model weights. """ @add_start_docstrings( "The bare LLaMA Model outputting raw hidden-states without any specific head on top.", LLAMA_START_DOCSTRING, ) class TFIdeficsPreTrainedModel(TFPreTrainedModel): config_class = IdeficsConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["TFIdeficsDecoderLayer", "TFIdeficsGatedCrossAttentionLayer"] LLAMA_INPUTS_DOCSTRING = r""" Args: input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. """ @add_start_docstrings( "The bare LLaMA Model outputting raw hidden-states without any specific head on top.", LLAMA_START_DOCSTRING, ) @keras_serializable class TFIdeficsMainLayer(tf.keras.layers.Layer): """ Transformer decoder consisting of `config.num_hidden_layers` layers. Each layer is a [`IdeficsDecoderLayer`] Args: config: IdeficsConfig """ config_class = IdeficsConfig def __init__(self, config: IdeficsConfig, add_pooling_year: bool = True, **kwargs): super().__init__(**kwargs) self.config = config self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = TFIdeficsDecoupledEmbedding( num_embeddings=config.vocab_size, num_additional_embeddings=config.additional_vocab_size, embedding_dim=config.hidden_size, partially_freeze=config.freeze_text_layers, name="embed_tokens", ) self.image_size = config.vision_config.image_size self.vision_config = config.vision_config self.vision_model = TFIdeficsVisionTransformer(config.vision_config, name="vision_model") # Perceiver Resampler if config.use_resampler: perceiver_config = config.perceiver_config self.perceiver_resampler = TFIdeficsPerceiverResampler( config, config.vision_config.embed_dim, perceiver_config.resampler_depth, perceiver_config.resampler_n_heads, perceiver_config.resampler_head_dim, perceiver_config.resampler_n_latents, name="perceiver_resampler", ) self.decoder_layers = [ TFIdeficsDecoderLayer(config, name=f"layers.{i}") for i in range(config.num_hidden_layers) ] self.cross_layer_interval = config.cross_layer_interval num_cross_layers = config.num_hidden_layers // self.cross_layer_interval self.gated_cross_attn_layers = [ TFIdeficsGatedCrossAttentionLayer(config, name=f"gated_cross_attn_layers.{i}") for i in range(num_cross_layers) ] self.gradient_checkpointing = False self.norm = TFIdeficsRMSNorm(config.hidden_size, eps=config.rms_norm_eps, name="norm") self.gradient_checkpointing = False self.freeze_relevant_params(config) def freeze_relevant_params(self, config=None): if config is None: config = self.config if config.freeze_text_layers: self.freeze_text_layers(config.freeze_text_module_exceptions) if config.freeze_vision_layers: freeze_model(self.vision_model, module_exceptions=config.freeze_vision_module_exceptions) def freeze_text_layers(self, module_exceptions=[]): for module in [self.decoder_layers, self.norm]: freeze_model(module, module_exceptions=module_exceptions) def freeze_vision_layers(self, module_exceptions=[]): freeze_model(self.vision_model, module_exceptions=module_exceptions) def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, value): self.embed_tokens = value def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length): # create causal mask # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] combined_attention_mask = None # if input_shape[-1] > 1: combined_attention_mask = _make_causal_mask( input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length, ) if attention_mask is not None: # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]) combined_attention_mask = ( expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask ) return combined_attention_mask @unpack_inputs @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING) def call( self, input_ids: TFModelInputType | None = None, attention_mask: Optional[tf.Tensor] = None, position_ids: Optional[tf.Tensor] = None, past_key_values: Optional[List[tf.Tensor]] = None, inputs_embeds: Optional[tf.Tensor] = None, pixel_values: Optional[tf.Tensor] = None, image_encoder_embeddings: Optional[tf.Tensor] = None, perceiver_embeddings: Optional[tf.Tensor] = None, image_attention_mask: Optional[tf.Tensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, interpolate_pos_encoding: Optional[bool] = False, return_dict: Optional[bool] = None, training: Optional[bool] = None, ) -> Union[TFIdeficsBaseModelOutputWithPast, Tuple[tf.Tensor]]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict # retrieve input_ids and inputs_embeds if input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") elif input_ids is not None: batch_size, seq_length = shape_list(input_ids) elif inputs_embeds is not None: batch_size, seq_length, _ = shape_list(inputs_embeds) else: raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") seq_length_with_past = seq_length past_key_values_length = 0 if past_key_values is not None: past_key_values_length = shape_list(past_key_values[0][0])[2] seq_length_with_past = seq_length_with_past + past_key_values_length if attention_mask is not None and position_ids is None: # create position_ids on the fly for batch generation position_ids = tf.math.cumsum(tf.cast(attention_mask, dtype=tf.int32), axis=-1) - 1 position_ids = tf.where(attention_mask == 0, 1, position_ids) elif position_ids is None: position_ids = tf.range(past_key_values_length, seq_length + past_key_values_length, dtype=tf.int32) position_ids = tf.expand_dims(position_ids, 0) no_images = False if ( sum((int(pixel_values is None), int(image_encoder_embeddings is None), int(perceiver_embeddings is None))) != 2 ): raise ValueError( "Exactly 1 of pixel_values, image_encoder_embeddings or perceiver_embeddings has to be not-None." ) elif pixel_values is not None: no_images = tf.reduce_sum(tf.cast(pixel_values, dtype=tf.int32)) == 0 pixel_values = tf.cast(pixel_values, dtype=self.dtype) # fp16 compatibility # Below hack is because when cross-loading pytorch weights, there is an # initial forward pass with dummy input and code below is here to handle that if len(pixel_values.shape) == 4: batch_size = shape_list(pixel_values)[0] num_images = shape_list(pixel_values)[0] # pixel_values = tf.reshape(pixel_values, [batch_size * num_images, *pixel_values.shape[1:]]) elif len(pixel_values.shape) == 5: batch_size, num_images = shape_list(pixel_values)[:2] pixel_values = tf.reshape(pixel_values, [batch_size * num_images, *pixel_values.shape[2:]]) # Get sequence from the vision encoder image_hidden_states = self.vision_model( pixel_values=pixel_values, interpolate_pos_encoding=interpolate_pos_encoding ).last_hidden_state elif image_encoder_embeddings is not None: batch_size, num_images, image_seq_len, image_hidden_size = shape_list(image_encoder_embeddings) image_hidden_states = tf.cast(image_encoder_embeddings, dtype=self.dtype) image_hidden_states = tf.reshape( image_hidden_states, (batch_size * num_images, image_seq_len, image_hidden_size) ) if self.config.use_resampler: if perceiver_embeddings is None: perceiver_embeddings = self.perceiver_resampler(image_hidden_states) image_seq_len, image_hidden_size = shape_list(perceiver_embeddings)[1:3] else: batch_size, num_images, image_seq_len, image_hidden_size = shape_list(perceiver_embeddings) image_hidden_states = perceiver_embeddings elif perceiver_embeddings is None: image_seq_len, image_hidden_size = shape_list(image_hidden_states)[1:3] else: raise ValueError("If `perceiver_embeddings` are passed, use_resampler should be True") image_hidden_states = tf.reshape( image_hidden_states, (batch_size, num_images * image_seq_len, image_hidden_size) ) # # Hack to use the model in full language modeling mode # image_attention_mask = tf.zeros((batch_size, seq_length, 1), dtype=tf.int32) # this is to account for the dummy inputs if pixel_values is not None and len(pixel_values.shape) == 4 and image_attention_mask is None: image_attention_mask = tf.zeros((batch_size, seq_length, 1), dtype=tf.int32) text_seq_len = shape_list(image_attention_mask)[1] image_attention_mask = tf.expand_dims(image_attention_mask, -1) image_attention_mask = tf.repeat(image_attention_mask, repeats=image_seq_len) image_attention_mask = tf.reshape(image_attention_mask, (batch_size, text_seq_len, num_images * image_seq_len)) if image_hidden_states is not None: image_batch_size, image_sequence_length, _ = shape_list(image_hidden_states) image_hidden_shape = (image_batch_size, image_sequence_length) if image_attention_mask is None: image_attention_mask = tf.ones(image_hidden_shape, dtype=tf.int32) image_attention_mask = invert_attention_mask(image_attention_mask) else: image_attention_mask = None cross_attention_gate = tf.squeeze( tf.cast(tf.reduce_any(image_attention_mask == 0, axis=-1), dtype=self.dtype), axis=1 ) if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) # embed positions if attention_mask is None: attention_mask = tf.ones((batch_size, seq_length_with_past), dtype=tf.bool) attention_mask = self._prepare_decoder_attention_mask( attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length ) hidden_states = inputs_embeds if self.gradient_checkpointing and training: if use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." ) use_cache = False # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None next_decoder_cache = () if use_cache else None for idx, decoder_layer in enumerate(self.decoder_layers): if output_hidden_states: all_hidden_states += (hidden_states,) past_key_value = past_key_values[idx] if past_key_values is not None else None def vblock( main_block, hidden_states, attention_mask, position_ids, past_key_value, image_hidden_states, image_attention_mask, cross_attention_gate, output_attentions, use_cache, layer_idx, cross_layer_interval, gated_cross_attn_layers, ): # TODO(ls): Add cross attention values to respective lists if layer_idx % cross_layer_interval == 0: xblock = gated_cross_attn_layers[layer_idx // cross_layer_interval] outputs = xblock( hidden_states, attention_mask=attention_mask, image_hidden_states=image_hidden_states, image_attention_mask=image_attention_mask, cross_attention_gate=cross_attention_gate, output_attentions=output_attentions, use_cache=use_cache, past_key_value=None, # not implemented ) hidden_states = outputs[0] layer_outputs = main_block( hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, ) return layer_outputs if self.gradient_checkpointing and training: past_key_value = None if use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." ) use_cache = False layer_outputs = tf.recompute_grad( vblock, decoder_layer, hidden_states, attention_mask, position_ids, past_key_value, image_hidden_states, image_attention_mask, output_attentions, use_cache, no_images, idx, self.cross_layer_interval, self.gated_cross_attn_layers, ) else: layer_outputs = vblock( decoder_layer, hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, image_hidden_states=image_hidden_states, image_attention_mask=image_attention_mask, cross_attention_gate=cross_attention_gate, output_attentions=output_attentions, use_cache=use_cache, layer_idx=idx, cross_layer_interval=self.cross_layer_interval, gated_cross_attn_layers=self.gated_cross_attn_layers, ) hidden_states = layer_outputs[0] if use_cache: next_decoder_cache += (layer_outputs[2 if output_attentions else 1],) if output_attentions: all_self_attns += (layer_outputs[1],) hidden_states = self.norm(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) next_cache = next_decoder_cache if use_cache else None image_hidden_states = tf.reshape( image_hidden_states, (batch_size, num_images, image_seq_len, image_hidden_size) ) if not return_dict: return tuple( v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, image_hidden_states] if v is not None ) return TFIdeficsBaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_self_attns, image_hidden_states=image_hidden_states, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "embed_tokens", None) is not None: with tf.name_scope(self.embed_tokens.name): self.embed_tokens.build(None) if getattr(self, "vision_model", None) is not None: with tf.name_scope(self.vision_model.name): self.vision_model.build(None) if getattr(self, "norm", None) is not None: with tf.name_scope(self.norm.name): self.norm.build(None) if getattr(self, "perceiver_resampler", None) is not None: with tf.name_scope(self.perceiver_resampler.name): self.perceiver_resampler.build(None) if getattr(self, "decoder_layers", None) is not None: for layer in self.decoder_layers: with tf.name_scope(layer.name): layer.build(None) if getattr(self, "gated_cross_attn_layers", None) is not None: for layer in self.gated_cross_attn_layers: with tf.name_scope(layer.name): layer.build(None) class TFIdeficsModel(TFIdeficsPreTrainedModel): def __init__(self, config: IdeficsConfig, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.model = TFIdeficsMainLayer(config, name="model") def call( self, input_ids: TFModelInputType | None = None, attention_mask: Optional[tf.Tensor] = None, position_ids: Optional[tf.Tensor] = None, past_key_values: Optional[List[tf.Tensor]] = None, inputs_embeds: Optional[tf.Tensor] = None, pixel_values: Optional[tf.Tensor] = None, image_encoder_embeddings: Optional[tf.Tensor] = None, perceiver_embeddings: Optional[tf.Tensor] = None, image_attention_mask: Optional[tf.Tensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, interpolate_pos_encoding: Optional[bool] = False, return_dict: Optional[bool] = None, training: Optional[bool] = None, ) -> Union[TFIdeficsBaseModelOutputWithPast, Tuple[tf.Tensor]]: outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, pixel_values=pixel_values, image_encoder_embeddings=image_encoder_embeddings, perceiver_embeddings=perceiver_embeddings, image_attention_mask=image_attention_mask, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, interpolate_pos_encoding=interpolate_pos_encoding, return_dict=return_dict, training=training, ) return outputs def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "model", None) is not None: with tf.name_scope(self.model.name): self.model.build(None) class TFIdeficsForVisionText2Text(TFPreTrainedModel, TFCausalLanguageModelingLoss): _keys_to_ignore_on_load_missing = [r"lm_head.weight"] _tied_weights_keys = ["model.embed_tokens.weight", "lm_head.weight"] config_class = IdeficsConfig def __init__(self, config, vision_model=None, **kwargs): super().__init__(config, **kwargs) self.model = TFIdeficsMainLayer(config, name="model") self.lm_head = TFIdeficsDecoupledLinear( config.hidden_size, config.vocab_size, config.additional_vocab_size, bias=False, partially_freeze=config.freeze_lm_head, name="lm_head", ) def get_input_embeddings(self): return self.model.embed_tokens def set_input_embeddings(self, value): self.model.embed_tokens = value def get_output_embeddings(self): return self.lm_head def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings def set_decoder(self, decoder): self.model = decoder def get_decoder(self): return self.model def tie_weights(self): """ Overwrite `transformers.modeling_utils.PreTrainedModel.tie_weights` to handle the case of IdeficsDecoupledLinear and IdeficsDecoupledEmbedding. """ output_embeddings = self.get_output_embeddings() input_embeddings = self.get_input_embeddings() if getattr(self.config, "tie_word_embeddings", True): output_embeddings.weight = input_embeddings.weight if input_embeddings.num_additional_embeddings > 0: assert output_embeddings.out_additional_features == input_embeddings.num_additional_embeddings output_embeddings.additional_fc.weight = input_embeddings.additional_embedding.weight if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"): output_embeddings.out_features = input_embeddings.num_embeddings if hasattr(output_embeddings, "out_additional_features") and hasattr( input_embeddings, "num_additional_embeddings" ): output_embeddings.out_additional_features = input_embeddings.num_additional_embeddings @unpack_inputs @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=TFIdeficsCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) def call( self, input_ids: TFModelInputType | None = None, attention_mask: Optional[tf.Tensor] = None, position_ids: Optional[tf.Tensor] = None, past_key_values: Optional[List[tf.Tensor]] = None, inputs_embeds: Optional[tf.Tensor] = None, pixel_values: Optional[tf.Tensor] = None, image_encoder_embeddings: Optional[tf.Tensor] = None, perceiver_embeddings: Optional[tf.Tensor] = None, image_attention_mask: Optional[tf.Tensor] = None, labels: Optional[tf.Tensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, interpolate_pos_encoding: Optional[bool] = False, return_dict: Optional[bool] = None, training=False, ) -> Union[TFIdeficsCausalLMOutputWithPast, Tuple[tf.Tensor]]: r""" labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. Returns: Example: ```python >> from transformers import AutoTokenizer, TFIdeficsForVisionText2Text >> model = TFIdeficsForVisionText2Text.from_pretrained("HuggingFaceM4/idefics-9b") >> tokenizer = AutoTokenizer.from_pretrained("HuggingFaceM4/idefics-9b") >> prompt = "Hey, are you consciours? Can you talk to me?" >> inputs = tokenizer(prompt, return_tensors="tf") >> # Generate >> generate_ids = model.generate(inputs.input_ids, max_length=30) >> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you." ```""" output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, pixel_values=pixel_values, image_encoder_embeddings=image_encoder_embeddings, perceiver_embeddings=perceiver_embeddings, image_attention_mask=image_attention_mask, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, interpolate_pos_encoding=interpolate_pos_encoding, return_dict=return_dict, training=training, ) hidden_states = outputs[0] logits = self.lm_head(hidden_states) loss = None if labels is not None: # Shift so that tokens < n predict n if attention_mask is not None: shift_attention_mask = attention_mask[..., 1:] shift_logits = logits[..., :-1, :][shift_attention_mask != 0] shift_labels = labels[..., 1:][shift_attention_mask != 0] else: shift_logits = logits[..., :-1, :] shift_labels = labels[..., 1:] # Flatten the tokens loss = self.hf_compute_loss( labels=tf.reshape(shift_labels, [-1]), logits=tf.reshape(shift_logits, [-1, shift_logits.shape[-1]]) ) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return TFIdeficsCausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, image_hidden_states=outputs.image_hidden_states, ) def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs): image_hidden_states = kwargs.pop("image_hidden_states", None) if image_hidden_states is not None: if self.config.use_resampler: kwargs["perceiver_embeddings"] = image_hidden_states else: kwargs["image_encoder_embeddings"] = image_hidden_states kwargs["pixel_values"] = None inputs = prepare_inputs_for_generation(input_ids, past=past, **kwargs) unwanted_kwargs = ["token_type_ids"] for kwarg in unwanted_kwargs: inputs.pop(kwarg, None) return inputs @staticmethod def _expand_inputs_for_generation( *args, **model_kwargs, ): return expand_inputs_for_generation(*args, **model_kwargs) @staticmethod def _update_model_kwargs_for_generation(outputs, model_kwargs, is_encoder_decoder): return update_model_kwargs_for_generation(outputs, model_kwargs) @staticmethod def _reorder_cache(past, beam_idx): reordered_past = () for layer_past in past: reordered_past += (tuple(tf.gather(past_state, beam_idx) for past_state in layer_past),) return reordered_past def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "model", None) is not None: with tf.name_scope(self.model.name): self.model.build(None) if getattr(self, "lm_head", None) is not None: with tf.name_scope(self.lm_head.name): self.lm_head.build(None) __all__ = ["TFIdeficsForVisionText2Text", "TFIdeficsModel", "TFIdeficsPreTrainedModel"]