# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/emu3/modular_emu3.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_emu3.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # coding=utf-8 # Copyright 2024 HuggingFace Inc. team. All rights reserved. # # # 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. import math from functools import cached_property, partial from typing import Callable, List, Optional, Tuple, Union import torch import torch.nn as nn import torch.nn.functional as F from ...activations import ACT2FN from ...cache_utils import Cache, DynamicCache, StaticCache from ...generation import GenerationMixin from ...modeling_attn_mask_utils import AttentionMaskConverter from ...modeling_flash_attention_utils import FlashAttentionKwargs from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...utils import ( LossKwargs, add_start_docstrings, add_start_docstrings_to_model_forward, can_return_tuple, is_torch_flex_attn_available, logging, replace_return_docstrings, ) from ...utils.deprecation import deprecate_kwarg from .configuration_emu3 import Emu3Config, Emu3TextConfig, Emu3VQVAEConfig if is_torch_flex_attn_available(): from torch.nn.attention.flex_attention import BlockMask from ...integrations.flex_attention import make_flex_block_causal_mask logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = "Emu3Config" class Emu3RMSNorm(nn.Module): def __init__(self, hidden_size, eps=1e-6): """ Emu3RMSNorm is equivalent to T5LayerNorm """ super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states): input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) return self.weight * hidden_states.to(input_dtype) def extra_repr(self): return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" class Emu3MLP(nn.Module): def __init__(self, config): super().__init__() self.config = config self.hidden_size = config.hidden_size self.intermediate_size = config.intermediate_size self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias) self.act_fn = ACT2FN[config.hidden_act] def forward(self, x): down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) return down_proj 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 torch.cat((-x2, x1), dim=-1) def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): """Applies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. position_ids (`torch.Tensor`, *optional*): Deprecated and unused. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. """ cos = cos.unsqueeze(unsqueeze_dim) sin = sin.unsqueeze(unsqueeze_dim) q_embed = (q * cos) + (rotate_half(q) * sin) k_embed = (k * cos) + (rotate_half(k) * sin) return q_embed, k_embed def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: """ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) """ batch, num_key_value_heads, slen, head_dim = hidden_states.shape if n_rep == 1: return hidden_states hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) def eager_attention_forward( module: nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: Optional[torch.Tensor], scaling: float, dropout: float = 0.0, **kwargs, ): key_states = repeat_kv(key, module.num_key_value_groups) value_states = repeat_kv(value, module.num_key_value_groups) attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling if attention_mask is not None: causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] attn_weights = attn_weights + causal_mask attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = torch.matmul(attn_weights, value_states) attn_output = attn_output.transpose(1, 2).contiguous() return attn_output, attn_weights class Emu3Attention(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__(self, config: Emu3Config, layer_idx: int): super().__init__() self.config = config self.layer_idx = layer_idx self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads self.scaling = self.head_dim**-0.5 self.attention_dropout = config.attention_dropout self.is_causal = True self.q_proj = nn.Linear( config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias ) self.k_proj = nn.Linear( config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias ) self.v_proj = nn.Linear( config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias ) self.o_proj = nn.Linear( config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias ) def forward( self, hidden_states: torch.Tensor, position_embeddings: Tuple[torch.Tensor, torch.Tensor], attention_mask: Optional[torch.Tensor], past_key_value: Optional[Cache] = None, cache_position: Optional[torch.LongTensor] = None, **kwargs: Unpack[FlashAttentionKwargs], ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: input_shape = hidden_states.shape[:-1] hidden_shape = (*input_shape, -1, self.head_dim) query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2) key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2) value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2) cos, sin = position_embeddings query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) attention_interface: Callable = eager_attention_forward if self.config._attn_implementation != "eager": if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False): logger.warning_once( "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to " 'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' ) else: attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, **kwargs, ) attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights class Emu3DecoderLayer(nn.Module): def __init__(self, config: Emu3Config, layer_idx: int): super().__init__() self.hidden_size = config.hidden_size self.self_attn = Emu3Attention(config=config, layer_idx=layer_idx) self.mlp = Emu3MLP(config) self.input_layernorm = Emu3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = Emu3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.dropout = nn.Dropout(config.attention_dropout) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1, query_sequence_length, key_sequence_length)` if default attention is used. 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(torch.FloatTensor)`, *optional*): cached past key and value projection states cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence kwargs (`dict`, *optional*): Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code into the model """ residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # Self Attention hidden_states, self_attn_weights = 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, cache_position=cache_position, position_embeddings=position_embeddings, **kwargs, ) hidden_states = residual + self.dropout(hidden_states) # Fully Connected residual = hidden_states hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = residual + self.dropout(hidden_states) outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) return outputs class Emu3VQVAEVectorQuantizer(nn.Module): """ A module for vector quantization using learned embedding vectors. This module implements the quantization process similar to te one described in the VQ-VAE (Vector Quantized Variational AutoEncoder) paper. It quantizes continuous input vectors into discrete codebook vectors, which are learned during training. Current implementation improves over previous ones by avoiding costly matrix multiplications and allowing for post-hoc remapping of indices. """ def __init__(self, config: Emu3VQVAEConfig): super().__init__() self.embedding = nn.Embedding(config.codebook_size, config.embed_dim) self.embedding.weight.data.uniform_(-1.0 / config.codebook_size, 1.0 / config.codebook_size) def forward(self, hidden_state: torch.Tensor): batch_size, temporal, channels, height, width = hidden_state.shape hidden_state = hidden_state.permute(0, 1, 3, 4, 2).contiguous() hidden_state_flattened = hidden_state.view(-1, channels) # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z hidden_state_sum = torch.sum(hidden_state_flattened**2, dim=1, keepdim=True) embedding_sum = torch.sum(self.embedding.weight**2, dim=1) # "bd,dn->bn", distances = 2 * torch.matmul(hidden_state_flattened, self.embedding.weight.transpose(0, 1)) distances = hidden_state_sum + embedding_sum - distances min_encoding_indices = torch.argmin(distances, dim=1) min_encoding_indices = min_encoding_indices.view(batch_size, temporal, height, width) return min_encoding_indices class Emu3VQVAEEncoderConvDownsample(nn.Module): def __init__(self, in_channels): super().__init__() self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0) def forward(self, hidden_states): # no asymmetric padding in torch conv, must do it ourselves hidden_states = F.pad(hidden_states, pad=(0, 1, 0, 1), mode="constant", value=0) hidden_states = self.conv(hidden_states) return hidden_states class Emu3VQVAEEncoderConvUpsample(nn.Module): def __init__(self, in_channels): super().__init__() self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1) def forward(self, hidden_states): hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest") hidden_states = self.conv(hidden_states) return hidden_states class Emu3VQVAEConv3d(nn.Module): def __init__( self, in_channel: int, out_channel: int, kernel_size: Tuple[int], stride: Tuple[int], ): super().__init__() padding_sizes = [one_kernel - one_stride for one_kernel, one_stride in zip(kernel_size[1:], stride[1:])] self.padding = () for pad_size in padding_sizes[::-1]: self.padding += (pad_size // 2 + pad_size % 2, pad_size // 2) self.padding += (2, 0) self.conv = nn.Conv3d( in_channel, out_channel, kernel_size, stride=stride, ) def forward(self, hidden_states: torch.Tensor): hidden_states = F.pad(hidden_states, self.padding) hidden_states = self.conv(hidden_states) return hidden_states class Emu3VQVAESpatialNorm(nn.Module): def __init__( self, in_channels: int, out_channels: int, ): super().__init__() self.norm_layer = nn.GroupNorm( num_channels=out_channels, num_groups=32, eps=1e-6, affine=True, ) self.conv_y = nn.Conv2d( in_channels, out_channels, kernel_size=1, stride=1, padding=0, ) self.conv_b = nn.Conv2d( in_channels, out_channels, kernel_size=1, stride=1, padding=0, ) def forward(self, hidden_states: torch.Tensor, quant_states: torch.Tensor): quant_states = F.interpolate(quant_states, size=hidden_states.shape[-2:], mode="nearest") hidden_states = self.norm_layer(hidden_states) hidden_states = hidden_states * self.conv_y(quant_states) + self.conv_b(quant_states) return hidden_states class Emu3VQVAETemporalUpsample(nn.Module): def __init__( self, in_channel: int, out_channel: int, ): super().__init__() self.conv = Emu3VQVAEConv3d( in_channel, out_channel, kernel_size=(3, 3, 3), stride=(1, 1, 1), ) def forward(self, hidden_states: torch.Tensor): batch_size, channels, temporal, height, width = hidden_states.shape hidden_states = hidden_states.permute(0, 1, 3, 4, 2).contiguous().view(batch_size, -1, temporal) hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest") hidden_states = hidden_states.view(batch_size, channels, height, width, -1).permute(0, 1, 4, 2, 3).contiguous() hidden_states = self.conv(hidden_states) return hidden_states class Emu3VQVAETemporalDownsample(nn.Module): def __init__( self, in_channel: int, out_channel: int, ): super().__init__() self.conv = Emu3VQVAEConv3d( in_channel, out_channel, kernel_size=(4, 3, 3), stride=(2, 1, 1), ) def forward(self, hidden_states: torch.Tensor): hidden_states = self.conv(hidden_states) return hidden_states class Emu3VQVAETemporalResnetBlock(nn.Module): def __init__( self, in_channels, out_channels=None, ): super().__init__() self.in_channels = in_channels self.out_channels = in_channels if out_channels is None else out_channels self.norm1 = nn.BatchNorm3d(in_channels) self.conv1 = Emu3VQVAEConv3d( in_channels, out_channels, kernel_size=(3, 3, 3), stride=(1, 1, 1), ) self.norm2 = nn.BatchNorm3d(out_channels) self.conv2 = Emu3VQVAEConv3d( out_channels, out_channels, kernel_size=(3, 3, 3), stride=(1, 1, 1), ) if self.in_channels != self.out_channels: self.nin_shortcut = nn.Conv3d( in_channels, out_channels, kernel_size=1, stride=1, padding=0, ) def forward(self, hidden_states): residual = hidden_states hidden_states = self.norm1(hidden_states) hidden_states *= torch.sigmoid(hidden_states) hidden_states = self.conv1(hidden_states) hidden_states = self.norm2(hidden_states) hidden_states *= torch.sigmoid(hidden_states) hidden_states = self.conv2(hidden_states) if self.in_channels != self.out_channels: residual = self.nin_shortcut(residual) return residual + hidden_states class Emu3VQVAEResnetBlock(nn.Module): def __init__( self, in_channels: int, out_channels: Optional[int] = None, quant_channels: Optional[int] = None, ): super().__init__() self.in_channels = in_channels out_channels = in_channels if out_channels is None else out_channels self.out_channels = out_channels self.quant_channels = quant_channels if quant_channels is None: self.norm1 = nn.GroupNorm(num_channels=in_channels, num_groups=32, eps=1e-6, affine=True) self.norm2 = nn.GroupNorm(num_channels=out_channels, num_groups=32, eps=1e-6, affine=True) else: self.norm1 = Emu3VQVAESpatialNorm(quant_channels, in_channels) self.norm2 = Emu3VQVAESpatialNorm(quant_channels, out_channels) self.conv1 = nn.Conv2d( in_channels, out_channels, kernel_size=3, stride=1, padding=1, ) self.conv2 = nn.Conv2d( out_channels, out_channels, kernel_size=3, stride=1, padding=1, ) if self.in_channels != self.out_channels: self.nin_shortcut = nn.Conv2d( in_channels, out_channels, kernel_size=1, stride=1, padding=0, ) def forward(self, hidden_states: torch.Tensor, quant_channels: Optional[torch.Tensor] = None): norm_args = () if self.quant_channels is None else (quant_channels,) residual = hidden_states hidden_states = self.norm1(hidden_states, *norm_args) hidden_states *= torch.sigmoid(hidden_states) hidden_states = self.conv1(hidden_states) hidden_states = self.norm2(hidden_states, *norm_args) hidden_states *= torch.sigmoid(hidden_states) hidden_states = self.conv2(hidden_states) if self.in_channels != self.out_channels: residual = self.nin_shortcut(residual) return residual + hidden_states class Emu3VQVAEAttentionBlock(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__(self, config: Emu3VQVAEConfig): super().__init__() self.config = config self.embed_dim = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.embed_dim // self.num_heads if self.head_dim * self.num_heads != self.embed_dim: raise ValueError( f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:" f" {self.num_heads})." ) self.scale = self.head_dim**-0.5 self.dropout = config.attention_dropout self.is_causal = False self.k_proj = nn.Linear(self.embed_dim, self.embed_dim) self.v_proj = nn.Linear(self.embed_dim, self.embed_dim) self.q_proj = nn.Linear(self.embed_dim, self.embed_dim) self.out_proj = nn.Linear(self.embed_dim, self.embed_dim) # for compatibility with the attention interface self.num_key_value_groups = 1 def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = False, ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]: """Input shape: Batch x Time x Channel""" batch_size, seq_length, embed_dim = hidden_states.shape queries = self.q_proj(hidden_states) keys = self.k_proj(hidden_states) values = self.v_proj(hidden_states) queries = queries.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2) keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2) values = values.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2) attention_interface: Callable = eager_attention_forward if self.config._attn_implementation != "eager": if self.config._attn_implementation == "sdpa" and output_attentions: logger.warning_once( "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to " 'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' ) else: attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] attn_output, attn_weights = attention_interface( self, queries, keys, values, attention_mask, is_causal=self.is_causal, scaling=self.scale, dropout=0.0 if not self.training else self.dropout, ) attn_output = attn_output.reshape(batch_size, seq_length, embed_dim).contiguous() attn_output = self.out_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights class Emu3VQVAEGroupNorm(nn.GroupNorm): """ Same as the torch GroupNorm with the only difference that this ones accepts an optional kwarg `quant_states` which is not used. This class makes it easier to use SpatialNorm or GroupNorm without conditionals """ def __init__(self, **kwargs): super().__init__(**kwargs) def forward(self, input, quant_states=None): return F.group_norm(input, self.num_groups, self.weight, self.bias, self.eps) class Emu3VQVAEMiddleBlock(nn.Module): def __init__(self, config, in_channels, quant_channels=None): super().__init__() self.block_1 = Emu3VQVAEResnetBlock( in_channels=in_channels, out_channels=in_channels, quant_channels=quant_channels, ) self.attn_1 = Emu3VQVAEAttentionBlock(config) if quant_channels is None: self.attn_norm = Emu3VQVAEGroupNorm(num_channels=in_channels, num_groups=32, eps=1e-6, affine=True) else: self.attn_norm = Emu3VQVAESpatialNorm(quant_channels, in_channels) self.block_2 = Emu3VQVAEResnetBlock( in_channels=in_channels, out_channels=in_channels, quant_channels=quant_channels, ) def forward(self, hidden_states: torch.FloatTensor, quant_states: Optional[torch.FloatTensor] = None): hidden_states = self.block_1(hidden_states, quant_states) residual = hidden_states hidden_states = self.attn_norm(hidden_states, quant_states) batch_size, channels, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channels, height * width).transpose(1, 2) hidden_states = self.attn_1(hidden_states)[0] hidden_states = hidden_states.reshape(batch_size, height, width, channels).permute(0, 3, 1, 2) hidden_states = residual + hidden_states hidden_states = self.block_2(hidden_states, quant_states) return hidden_states class Emu3VQVAEDownBlock(nn.Module): def __init__(self, config): super().__init__() self.num_resolutions = len(config.channel_multiplier) self.num_res_blocks = config.num_res_blocks base_channels = config.base_channels channel_multiplier = config.channel_multiplier in_channel_multiplier = (1,) + tuple(channel_multiplier) self.in_channel_multiplier = in_channel_multiplier self.down = nn.ModuleList() for i_level in range(self.num_resolutions): block = nn.ModuleList() attn = nn.ModuleList() attn_norms = nn.ModuleList() block_in = base_channels * in_channel_multiplier[i_level] block_out = base_channels * channel_multiplier[i_level] for i_block in range(self.num_res_blocks): block.append( Emu3VQVAEResnetBlock( in_channels=block_in, out_channels=block_out, ) ) block_in = block_out if config.attn_resolutions is not None and i_level in config.attn_resolutions: attn.append(Emu3VQVAEAttentionBlock(config)) attn_norms.append(nn.GroupNorm(num_channels=block_in, num_groups=32, eps=1e-6, affine=True)) down = nn.Module() down.block = block down.attn = attn down.attn_norms = attn_norms if i_level != self.num_resolutions - 1: down.downsample = Emu3VQVAEEncoderConvDownsample(block_in) self.down.append(down) def forward(self, hidden_states: torch.FloatTensor): for i_level, blocks in enumerate(self.down): for i_block in range(self.num_res_blocks): hidden_states = blocks.block[i_block](hidden_states) if len(blocks.attn) > 0: residual = hidden_states hidden_states = blocks.attn_norms[i_block](hidden_states) batch_size, channels, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channels, height * width).transpose(1, 2) hidden_states = blocks.attn[i_block](hidden_states)[0] hidden_states = hidden_states.reshape(batch_size, height, width, channels).permute(0, 3, 1, 2) hidden_states = residual + hidden_states if i_level != self.num_resolutions - 1: hidden_states = blocks.downsample(hidden_states) return hidden_states class Emu3VQVAEUpBlock(nn.Module): def __init__(self, config): super().__init__() self.num_resolutions = len(config.channel_multiplier) self.num_res_blocks = config.num_res_blocks quant_channels = config.embed_dim block_in = config.base_channels * config.channel_multiplier[-1] self.up = nn.ModuleList() for i_level in reversed(range(self.num_resolutions)): block = nn.ModuleList() attn = nn.ModuleList() attn_norms = nn.ModuleList() block_out = config.base_channels * config.channel_multiplier[i_level] for i_block in range(self.num_res_blocks + 1): block.append( Emu3VQVAEResnetBlock( in_channels=block_in, out_channels=block_out, quant_channels=quant_channels, ) ) block_in = block_out if i_level in config.attn_resolutions: attn.append(Emu3VQVAEAttentionBlock(config)) attn_norms.append(Emu3VQVAESpatialNorm(quant_channels, block_in)) up = nn.Module() up.block = block up.attn = attn up.attn_norms = attn_norms if i_level != 0: up.upsample = Emu3VQVAEEncoderConvUpsample(block_in) self.up.insert(0, up) def forward(self, hidden_states: torch.FloatTensor, quant_states: torch.FloatTensor): for i_level, blocks in enumerate(self.up[::-1]): for i_block in range(self.num_res_blocks + 1): hidden_states = blocks.block[i_block](hidden_states, quant_states) if len(blocks.attn) > 0: residual = hidden_states hidden_states = blocks.attn_norms[i_block](hidden_states, quant_states) batch_size, channels, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channels, height * width).transpose(1, 2) hidden_states = blocks.attn[i_block](hidden_states)[0] hidden_states = hidden_states.reshape(batch_size, height, width, channels).permute(0, 3, 1, 2) hidden_states = residual + hidden_states if i_level != len(self.up) - 1: hidden_states = blocks.upsample(hidden_states) return hidden_states class Emu3VQVAEEncoder(nn.Module): def __init__(self, config): super().__init__() base_channels = config.base_channels in_channels = config.in_channels double_latent = config.double_latent latent_channels = config.latent_channels channel_multiplier = config.channel_multiplier out_channels = 2 * latent_channels if double_latent else latent_channels block_in = base_channels * channel_multiplier[-1] self.conv_in = torch.nn.Conv2d(in_channels, base_channels, kernel_size=3, stride=1, padding=1) self.down_block = Emu3VQVAEDownBlock(config) self.middle_block = Emu3VQVAEMiddleBlock(config, block_in) self.norm_out = torch.nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True) self.conv_out = torch.nn.Conv2d( block_in, out_channels, kernel_size=3, stride=1, padding=1, ) temporal_down_blocks = int(math.log2(config.temporal_downsample_factor)) self.time_conv = nn.ModuleList() self.time_res_stack = nn.ModuleList() for i in range(temporal_down_blocks): conv = Emu3VQVAETemporalDownsample(out_channels, out_channels) self.time_conv.append(conv) for _ in range(config.num_res_blocks): time_res_conv = Emu3VQVAETemporalResnetBlock( in_channels=out_channels, out_channels=out_channels, ) self.time_res_stack.append(time_res_conv) def forward(self, pixel_values: torch.LongTensor): temporal_dim = pixel_values.shape[1] pixel_values = pixel_values.reshape(-1, *pixel_values.shape[2:]) # downsampling & middle hidden_states = self.conv_in(pixel_values) hidden_states = self.down_block(hidden_states) hidden_states = self.middle_block(hidden_states) # end hidden_states = self.norm_out(hidden_states) hidden_states *= torch.sigmoid(hidden_states) hidden_states = self.conv_out(hidden_states) hidden_states = hidden_states.reshape(-1, temporal_dim, *hidden_states.shape[1:]) hidden_states = hidden_states.permute(0, 2, 1, 3, 4) # temporal convs for conv in self.time_conv: hidden_states = conv(hidden_states) hidden_states *= torch.sigmoid(hidden_states) for layer in self.time_res_stack: hidden_states = layer(hidden_states) hidden_states = hidden_states.permute(0, 2, 1, 3, 4) return hidden_states class Emu3VQVAEDecoder(nn.Module): def __init__(self, config: Emu3VQVAEConfig): super().__init__() quant_channels = config.embed_dim block_in = config.base_channels * config.channel_multiplier[-1] self.time_res_stack = nn.ModuleList() for _ in range(config.num_res_blocks): time_res_conv = Emu3VQVAETemporalResnetBlock( in_channels=config.latent_channels, out_channels=config.latent_channels ) self.time_res_stack.append(time_res_conv) temp_upsample_block_num = int(math.log2(config.temporal_downsample_factor)) self.time_conv = nn.ModuleList() for i in range(temp_upsample_block_num): conv = Emu3VQVAETemporalUpsample(config.latent_channels, config.latent_channels) self.time_conv.append(conv) self.conv_in = nn.Conv2d( config.latent_channels, block_in, kernel_size=3, stride=1, padding=1, ) self.middle_block = Emu3VQVAEMiddleBlock(config, block_in, quant_channels=quant_channels) self.up_block = Emu3VQVAEUpBlock(config) block_in = config.base_channels * config.channel_multiplier[0] self.norm_out = Emu3VQVAESpatialNorm(quant_channels, block_in) self.conv_out = nn.Conv2d( block_in, config.out_channels, kernel_size=3, stride=1, padding=1, ) def forward(self, hidden_states: torch.Tensor, quant_states: torch.Tensor): hidden_quant_states = torch.cat((hidden_states, quant_states), dim=0) hidden_quant_states = hidden_quant_states.permute(0, 2, 1, 3, 4) # temporal convs for layer in self.time_res_stack: hidden_quant_states = layer(hidden_quant_states) for layer in self.time_conv: hidden_quant_states = layer(hidden_quant_states) hidden_quant_states *= torch.sigmoid(hidden_quant_states) hidden_quant_states = hidden_quant_states.permute(0, 2, 1, 3, 4) hidden_states, quant_states = torch.chunk(hidden_quant_states, 2, dim=0) hidden_states = hidden_states.reshape(-1, *hidden_states.shape[2:]) quant_states = quant_states.reshape(-1, *quant_states.shape[2:]) hidden_states = self.conv_in(hidden_states) # middle & upsampling hidden_states = self.middle_block(hidden_states, quant_states) hidden_states = self.up_block(hidden_states, quant_states) hidden_states = self.norm_out(hidden_states, quant_states) hidden_states *= torch.sigmoid(hidden_states) hidden_states = self.conv_out(hidden_states) return hidden_states EMU3_VQ_START_DOCSTRING = r""" This model inherits from [`PreTrainedModel`]. 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 PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`Emu3VQVAEConfig`]): 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 [`~PreTrainedModel.from_pretrained`] method to load the model weights. """ @add_start_docstrings( """The VQ-VAE model used in Emu3 for encoding/decoding images into discrete tokens. This model follows the "Make-a-scene: Scene-based text-to-image generation with human priors" paper from [ Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman](https://arxiv.org/abs/2203.13131). """, EMU3_VQ_START_DOCSTRING, ) class Emu3VQVAE(PreTrainedModel): config_class = Emu3VQVAEConfig base_model_prefix = "emuvideovq" main_input_name = "pixel_values" _supports_sdpa = True _supports_flash_attn_2 = True _supports_flex_attn = True _no_split_modules = [ "Emu3VQVAETemporalResnetBlock", "Emu3VQVAEAttentionBlock", "Emu3VQVAEResnetBlock", "Emu3VQVAEVectorQuantizer", ] def _init_weights(self, module): if isinstance(module, (nn.Conv2d, nn.Conv3d)): nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu") elif isinstance(module, nn.Linear): nn.init.kaiming_uniform_(module.weight, a=math.sqrt(5)) if module.bias is not None: fan_in, _ = nn.init._calculate_fan_in_and_fan_out(module.weight) bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0 nn.init.uniform_(module.bias, -bound, bound) elif isinstance(module, (nn.BatchNorm2d, nn.BatchNorm3d, nn.GroupNorm)): nn.init.constant_(module.weight, 1) nn.init.constant_(module.bias, 0) def __init__(self, config: Emu3VQVAEConfig): super().__init__(config) self.config = config self.encoder = Emu3VQVAEEncoder(config) self.decoder = Emu3VQVAEDecoder(config) self.quantize = Emu3VQVAEVectorQuantizer(config) self.vision_spatial_factor = 2 ** (len(config.channel_multiplier) - 1) self.quant_conv = Emu3VQVAEConv3d( config.latent_channels, config.embed_dim, kernel_size=(3, 1, 1), stride=(1, 1, 1) ) self.post_quant_conv = Emu3VQVAEConv3d( config.embed_dim, config.latent_channels, kernel_size=(3, 1, 1), stride=(1, 1, 1) ) self.spatial_scale_factor = 2 ** (len(config.channel_multiplier) - 1) self.eval() # Emu3's VQ model is frozen self.post_init() def encode(self, pixel_values: torch.Tensor, image_sizes: torch.Tensor): is_image = pixel_values.ndim == 4 if is_image: temporal = self.config.temporal_downsample_factor batch_size, channels, height, width = pixel_values.shape pixel_values = pixel_values.unsqueeze(1).repeat(1, temporal, 1, 1, 1) else: batch_size, temporal, channels, height, width = pixel_values.shape hidden_states = self.encoder(pixel_values) # b t c h w -> b c t h w hidden_states = hidden_states.permute(0, 2, 1, 3, 4) hidden_states = self.quant_conv(hidden_states) # b c t h w -> b t c h w hidden_states = hidden_states.permute(0, 2, 1, 3, 4) codes = self.quantize(hidden_states) image_tokens = codes.squeeze(1) if is_image else codes image_tokens = [ single_image[: int(size[0] / self.vision_spatial_factor), : int(size[1] / self.vision_spatial_factor)] for single_image, size in zip(image_tokens, image_sizes) ] return image_tokens def decode(self, hidden_states: torch.Tensor): is_image = hidden_states.ndim == 3 if is_image: hidden_states = hidden_states.unsqueeze(1) batch_size, temporal, height, width = hidden_states.shape quant = self.quantize.embedding(hidden_states.flatten()) channels = quant.shape[-1] quant = quant.view(batch_size, temporal, height, width, channels).permute(0, 4, 1, 2, 3).contiguous() post_quant = self.post_quant_conv(quant) quant = quant.permute(0, 2, 1, 3, 4) post_quant = post_quant.permute(0, 2, 1, 3, 4) video = self.decoder(post_quant, quant) video = video.reshape( batch_size, temporal * self.config.temporal_downsample_factor, self.config.out_channels, height * self.spatial_scale_factor, width * self.spatial_scale_factor, ) return video[:, 0] if is_image else video class Emu3ImageVocabularyMapping: """ A class for mapping discrete image tokens from VQGAN to BPE tokens. """ def __init__(self, vocab_map): self.vocab_map = vocab_map self.eol_token_id = vocab_map.get("<|extra_200|>") self.image_token_id = vocab_map.get("") @cached_property def image_tokens(self): return sorted([val for name, val in self.vocab_map.items() if name.startswith("<|visual token")]) @cached_property def image_tokens_str(self): return sorted([name for name, val in self.vocab_map.items() if name.startswith("<|visual token")]) @cached_property def img2bpe(self): return {int(token[-8:-2]): self.vocab_map[token] for token in self.image_tokens_str} @cached_property def bpe2img(self): return {v: k for k, v in self.img2bpe.items()} @cached_property def bpe2img_mapping_tensor(self): mapping = torch.zeros(max(self.bpe2img.keys()) + 1, dtype=torch.int) for k, v in self.bpe2img.items(): mapping[k] = v return mapping @cached_property def img2bpe_mapping_tensor(self): mapping = torch.zeros(max(self.img2bpe.keys()) + 1, dtype=torch.int) for k, v in self.img2bpe.items(): mapping[k] = v return mapping def convert_img2bpe(self, img_batch: List[torch.Tensor]) -> torch.Tensor: device = img_batch.device eol_row = torch.ones((img_batch.shape[0], 1), dtype=torch.int) * self.eol_token_id img_tokens = self.img2bpe_mapping_tensor[img_batch.to("cpu")] img_tokens = torch.cat([img_tokens, eol_row], dim=-1) return img_tokens.to(device) def convert_bpe2img(self, img_batch: torch.Tensor) -> torch.Tensor: device = img_batch.device img_batch = img_batch[..., :-1] # remove last row of EOL tokens img_tokens = self.bpe2img_mapping_tensor[img_batch.to("cpu")] return img_tokens.to(device) EMU3_START_DOCSTRING = r""" This model inherits from [`PreTrainedModel`]. 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 PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`Emu3Config`]): 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 [`~PreTrainedModel.from_pretrained`] method to load the model weights. """ @add_start_docstrings( "The bare emu3 Model outputting raw hidden-states without any specific head on top.", EMU3_START_DOCSTRING, ) class Emu3PreTrainedModel(PreTrainedModel): config_class = Emu3Config base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = [ "Emu3DecoderLayer", ] _skip_keys_device_placement = ["past_key_values", "causal_mask"] _supports_flash_attn_2 = True _supports_sdpa = True _supports_quantized_cache = True _supports_cache_class = True _supports_static_cache = True _supports_param_buffer_assignment = False _supports_flex_attn = True def _init_weights(self, module): std = self.config.get_text_config().initializer_range if isinstance(module, Emu3VQVAE): module.apply(module._init_weights) elif isinstance(module, (nn.Linear, nn.Conv2d)): module.weight.data.normal_(mean=0.0, std=std) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=std) if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_() class Emu3RotaryEmbedding(nn.Module): def __init__(self, config: Emu3Config, device=None): super().__init__() # BC: "rope_type" was originally "type" if hasattr(config, "rope_scaling") and config.rope_scaling is not None: self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) else: self.rope_type = "default" self.max_seq_len_cached = config.max_position_embeddings self.original_max_seq_len = config.max_position_embeddings self.config = config self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device) self.register_buffer("inv_freq", inv_freq, persistent=False) self.original_inv_freq = self.inv_freq @torch.no_grad() @dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope) def forward(self, x, position_ids): inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) position_ids_expanded = position_ids[:, None, :].float() device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): # Force float32 freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() * self.attention_scaling sin = emb.sin() * self.attention_scaling return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) EMU3_TEXT_INPUTS_DOCSTRING = r""" Args: input_ids (`torch.LongTensor` 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 (`torch.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 `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 (`torch.LongTensor` 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 (`Cache`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. Has to be an instance of [`~cache_utils.Cache`] instance, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). The model will output the same cache type that is fed as input. If no `past_key_values` are passed, the legacy cache format will be returned. If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` 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. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length. """ @add_start_docstrings( "The bare Emu3Text Model outputting raw hidden-states without any specific head on top.", EMU3_START_DOCSTRING, ) class Emu3TextModel(Emu3PreTrainedModel): """ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Emu3TextDecoderLayer`] Args: config: Emu3TextConfig """ def __init__(self, config: Emu3Config): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) self.layers = nn.ModuleList( [Emu3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.norm = Emu3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.rotary_emb = Emu3RotaryEmbedding(config=config) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, value): self.embed_tokens = value @can_return_tuple @add_start_docstrings_to_model_forward(EMU3_TEXT_INPUTS_DOCSTRING) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, **flash_attn_kwargs: Unpack[FlashAttentionKwargs], ) -> BaseModelOutputWithPast: 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 if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if self.gradient_checkpointing and self.training and use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." ) use_cache = False # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache if not isinstance(past_key_values, (type(None), Cache)): raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.") if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if use_cache and past_key_values is None: past_key_values = DynamicCache() if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 cache_position = torch.arange( past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) if position_ids is None: position_ids = cache_position.unsqueeze(0) causal_mask = self._update_causal_mask( attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions ) hidden_states = inputs_embeds # create position embeddings to be shared across the decoder layers position_embeddings = self.rotary_emb(hidden_states, position_ids) # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None for decoder_layer in self.layers[: self.config.num_hidden_layers]: if output_hidden_states: all_hidden_states += (hidden_states,) if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( partial(decoder_layer.__call__, **flash_attn_kwargs), hidden_states, causal_mask, position_ids, past_key_values, output_attentions, use_cache, cache_position, position_embeddings, ) else: layer_outputs = decoder_layer( hidden_states, attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, position_embeddings=position_embeddings, **flash_attn_kwargs, ) hidden_states = layer_outputs[0] 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,) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values if use_cache else None, hidden_states=all_hidden_states, attentions=all_self_attns, ) def _update_causal_mask( self, attention_mask: torch.Tensor, input_tensor: torch.Tensor, cache_position: torch.Tensor, past_key_values: Cache, output_attentions: bool = False, ): if self.config._attn_implementation == "flash_attention_2": if attention_mask is not None and (attention_mask == 0.0).any(): return attention_mask return None if self.config._attn_implementation == "flex_attention": if isinstance(attention_mask, torch.Tensor): attention_mask = make_flex_block_causal_mask(attention_mask) if isinstance(attention_mask, BlockMask): return attention_mask # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail # to infer the attention mask. past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 using_static_cache = isinstance(past_key_values, StaticCache) # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: if AttentionMaskConverter._ignore_causal_mask_sdpa( attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens, is_training=self.training, ): return None dtype, device = input_tensor.dtype, input_tensor.device sequence_length = input_tensor.shape[1] if using_static_cache: target_length = past_key_values.get_max_cache_shape() else: target_length = ( attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length + 1 ) # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position( attention_mask, sequence_length=sequence_length, target_length=target_length, dtype=dtype, device=device, cache_position=cache_position, batch_size=input_tensor.shape[0], ) if ( self.config._attn_implementation == "sdpa" and attention_mask is not None and attention_mask.device.type in ["cuda", "xpu"] and not output_attentions ): # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. # Details: https://github.com/pytorch/pytorch/issues/110213 min_dtype = torch.finfo(dtype).min causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) return causal_mask @staticmethod def _prepare_4d_causal_attention_mask_with_cache_position( attention_mask: torch.Tensor, sequence_length: int, target_length: int, dtype: torch.dtype, device: torch.device, cache_position: torch.Tensor, batch_size: int, **kwargs, ): """ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. device (`torch.device`): The device to place the 4D attention mask on. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. """ if attention_mask is not None and attention_mask.dim() == 4: # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. causal_mask = attention_mask else: min_dtype = torch.finfo(dtype).min causal_mask = torch.full( (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device ) if sequence_length != 1: causal_mask = torch.triu(causal_mask, diagonal=1) causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) if attention_mask is not None: causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit mask_length = attention_mask.shape[-1] padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to( causal_mask.device ) padding_mask = padding_mask == 0 causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( padding_mask, min_dtype ) return causal_mask class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... class Emu3ForCausalLM(Emu3PreTrainedModel, GenerationMixin): _tied_weights_keys = ["lm_head.weight"] _tp_plan = {"lm_head": "colwise_rep"} _pp_plan = {"lm_head": (["hidden_states"], ["logits"])} config_class = Emu3TextConfig def __init__(self, config): super().__init__(config) self.model = Emu3TextModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() 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 @can_return_tuple @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep") @add_start_docstrings_to_model_forward(EMU3_TEXT_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="Emu3TextConfig") def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, logits_to_keep: Union[int, torch.Tensor] = 0, **kwargs: Unpack[KwargsForCausalLM], ) -> CausalLMOutputWithPast: r""" labels (`torch.LongTensor` 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]`. logits_to_keep (`int` or `torch.Tensor`, *optional*): If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension. This is useful when using packed tensor format (single dimension for batch and sequence length). Returns: Example: ```python >>> from transformers import Emu3Processor, Emu3ForConditionalGeneration >>> import torch >>> import requests >>> from PIL import Image >>> model = Emu3ForCausalLM.from_pretrained("BAAI/Emu3-Chat-hf", torch_dtype=torch.bfloat16) >>> processor = Emu3Processor.from_pretrained("BAAI/Emu3-Chat-hf") >>> inputs = processor(text=["Can you write me a poem about winter."], return_tensors="pt").to(model.device) >>> generated_ids = model.generate(**inputs, max_new_tokens=100, do_sample=False) >>> processor.batch_decode(generated_ids, skip_special_tokens=True)[0] ```""" 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 ) # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs: BaseModelOutputWithPast = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, cache_position=cache_position, **kwargs, ) hidden_states = outputs.last_hidden_state # Only compute necessary logits, and do not upcast them to float if we are not computing the loss slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]) loss = None if labels is not None: loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) EMU3_INPUTS_DOCSTRING = r""" Args: input_ids (`torch.LongTensor` 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) pixel_values (`torch.FloatTensor` of shape `(batch_size, max_num_images, max_num_tiles, channels, image_size, image_size)): The tensors corresponding to the input images. Pixel values can be obtained using [`AutoImageProcessor`]. See [`Emu3ImageProcessor.__call__`] for details ([]`Emu3Processor`] uses [`Emu3ImageProcessor`] for processing images). image_sizes (`torch.LongTensor` of shape `(batch_size, 2)`): The sizes of the images in the batch, being (height, width) for each image. Image sizes can be obtained using [`AutoImageProcessor`]. See [`Emu3ImageProcessor.__call__`] for details ([]`Emu3Processor`] uses [`Emu3ImageProcessor`] for processing images). attention_mask (`torch.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 `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 (`torch.LongTensor` 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 (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. Has to be an instance of [`~cache_utils.Cache`] instance, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the legacy cache format will be returned. If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` 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. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length. """ class Emu3ForConditionalGeneration(Emu3PreTrainedModel, GenerationMixin): _tied_weights_keys = ["text_model.lm_head.weight"] _supports_static_cache = False # `get_image_tokens()`, called when `pixel_values` is passed, is not compileable def __init__(self, config): super().__init__(config) self.text_model = Emu3ForCausalLM._from_config(config.text_config) self.vqmodel = Emu3VQVAE(config.vq_config) self.vocabulary_mapping = Emu3ImageVocabularyMapping(config.vocabulary_map) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.text_model.get_input_embeddings() def set_input_embeddings(self, value): self.text_model.set_input_embeddings(value) def get_image_tokens(self, pixel_values: torch.FloatTensor, image_sizes: torch.LongTensor): """ Tokenizes images into discrete tokens with VQGAN module. Converts obtained image tokens into BPE tokens and wraps with "boi" and "eoi" special tokens. Args: pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`): The tensors corresponding to the input images. image_sizes (`torch.LongTensor` of shape `(batch_size, 2)`): The sizes of the images in the batch, being (height, width) for each image. """ image_tokens_list = self.vqmodel.encode(pixel_values, image_sizes) bpe_tokens_list = [self.vocabulary_mapping.convert_img2bpe(tokens).flatten() for tokens in image_tokens_list] bpe_tokens = torch.cat(bpe_tokens_list) return bpe_tokens @torch.no_grad def decode_image_tokens(self, image_tokens: torch.LongTensor, height: int, width: int): """ Decodes generated image tokens from language model to continuous pixel values with VQGAN module via upsampling. Args: image_tokens (`torch.LongTensor` of shape `(batch_size, num_of_tokens)`): The tensors corresponding to the input images. height (`int`): Height of the generated image before upsampling. width (`int`): Width of the generated image before upsampling. """ sequences = image_tokens[:, :-3].view(-1, height, width + 1) image_tokens = self.vocabulary_mapping.convert_bpe2img(sequences) image = self.vqmodel.decode(image_tokens) return image @can_return_tuple @add_start_docstrings_to_model_forward(EMU3_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: Optional[torch.LongTensor] = None, pixel_values: Optional[torch.FloatTensor] = None, image_sizes: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, labels: Optional[torch.LongTensor] = None, logits_to_keep: Union[int, torch.Tensor] = 0, ) -> CausalLMOutputWithPast: r""" labels (`torch.LongTensor` 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]`. logits_to_keep (`int` or `torch.Tensor`, *optional*): If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension. This is useful when using packed tensor format (single dimension for batch and sequence length). Returns: Example: ```python >>> from transformers import Emu3Processor, Emu3ForConditionalGeneration >>> import torch >>> import requests >>> from PIL import Image >>> model = Emu3ForConditionalGeneration.from_pretrained("BAAI/Emu3-Chat-hf", torch_dtype=torch.bfloat16) >>> processor = Emu3Processor.from_pretrained("BAAI/Emu3-Chat-hf") >>> conversation = [ ... { ... "role": "system", ... "content": [ ... {"type": "text", "text": "You are a helpful assistant."}, ... ], ... }, ... { ... "role": "user", ... "content": [ ... {"type": "image"}, ... {"type": "text", "text": "Please describe the image."}, ... ], ... }, ... ] >>> prompt = processor.apply_chat_template(conversation, add_generation_prompt=True) >>> image = Image.open(requests.get("https://www.ilankelman.org/stopsigns/australia.jpg", stream=True).raw) >>> inputs = processor(images=[image], text=[prompt], return_tensors="pt").to(model.device, torch.bfloat16) >>> generated_ids = model.generate(**inputs, max_new_tokens=100, do_sample=False) >>> processor.batch_decode(generated_ids, skip_special_tokens=True)[0] ```""" 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 ) if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError( "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" ) if pixel_values is not None and inputs_embeds is not None: raise ValueError( "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one" ) if pixel_values is not None: image_tokens = self.get_image_tokens(pixel_values, image_sizes) special_image_mask = input_ids == self.vocabulary_mapping.image_token_id image_tokens = image_tokens.to(input_ids.device, input_ids.dtype) input_ids = input_ids.masked_scatter(special_image_mask, image_tokens) # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) return self.text_model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, cache_position=cache_position, logits_to_keep=logits_to_keep, ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, cache_position=None, position_ids=None, use_cache=True, pixel_values=None, **kwargs, ): # Overwritten -- in specific circumstances we don't want to forward image inputs to the model model_inputs = super().prepare_inputs_for_generation( input_ids, past_key_values=past_key_values, attention_mask=attention_mask, inputs_embeds=inputs_embeds, cache_position=cache_position, position_ids=position_ids, pixel_values=pixel_values, use_cache=use_cache, **kwargs, ) if cache_position[0] != 0: model_inputs["pixel_values"] = None return model_inputs __all__ = ["Emu3ForConditionalGeneration", "Emu3ForCausalLM", "Emu3TextModel", "Emu3PreTrainedModel", "Emu3VQVAE"]