automatic/pipelines/omnigen2/models/transformers/block_lumina2.py

65 lines
2.2 KiB
Python

# Copyright 2024 Alpha-VLLM Authors and The HuggingFace 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.
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import RMSNorm
from diffusers.models.embeddings import Timesteps, TimestepEmbedding
# Makes timestep_scale configurable
# Omnigen 2 uses timestep_scale=1000
class Lumina2CombinedTimestepCaptionEmbedding(nn.Module):
def __init__(
self,
hidden_size: int = 4096,
text_feat_dim: int = 2048,
frequency_embedding_size: int = 256,
norm_eps: float = 1e-5,
timestep_scale: float = 1.0,
) -> None:
super().__init__()
self.time_proj = Timesteps(
num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0.0, scale=timestep_scale
)
self.timestep_embedder = TimestepEmbedding(
in_channels=frequency_embedding_size, time_embed_dim=min(hidden_size, 1024)
)
self.caption_embedder = nn.Sequential(
RMSNorm(text_feat_dim, eps=norm_eps),
nn.Linear(text_feat_dim, hidden_size, bias=True),
)
self._initialize_weights()
def _initialize_weights(self):
nn.init.trunc_normal_(self.caption_embedder[1].weight, std=0.02)
nn.init.zeros_(self.caption_embedder[1].bias)
def forward(
self, timestep: torch.Tensor, text_hidden_states: torch.Tensor, dtype: torch.dtype
) -> Tuple[torch.Tensor, torch.Tensor]:
timestep_proj = self.time_proj(timestep).to(dtype=dtype)
time_embed = self.timestep_embedder(timestep_proj)
caption_embed = self.caption_embedder(text_hidden_states)
return time_embed, caption_embed