Add MAGI-1: Autoregressive Video Generation at Scale #11713
Draft
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…te attention mechanism accordingly. Updated initialization parameters and reshaping logic.
…tering and equal split ratio. Add utility functions for resizing and cropping images while preserving aspect ratio. Enhance 3D rotary positional embeddings Adds `center_grid_hw_indices` and `equal_split_ratio` parameters to the 3D rotary positional embedding function for more flexible configuration. The `center_grid_hw_indices` option centers the spatial grid indices around zero. The `equal_split_ratio` parameter provides an alternative way to divide the embedding dimension equally among the temporal and spatial axes. Updates the Magi1 VAE to utilize these new embedding features, introducing helper functions to prepare the embeddings dynamically based on input tensor dimensions.
Replaces the initial causal 3D convolution in the encoder with a standard `Conv3d` patch embedding layer. This simplifies the model and makes its input processing more consistent with Diffusion Transformer (DiT) architectures. Additionally, this change: - Removes the unused `Magi1CausalConv3d` class. - Updates the attention mechanism to use the standard `scaled_dot_product_attention`. - Sets the default for `sample_posterior` to `True` in the forward pass.
Removes the feature caching logic (`feat_cache`, `feat_idx`) from the encoder, decoder, and their sub-modules. This change significantly simplifies the forward pass implementation by removing stateful cache management. Additionally, this commit replaces the custom `Magi1RMS_norm` with a standard `nn.LayerNorm` and updates several custom causal convolution layers to use standard `nn.Linear` or `nn.Conv3d` layers.
Moves the positional embedding and dropout layers from the main autoencoder class into the decoder module. This improves encapsulation as the embedding is only used within the decoder. The decoder's forward pass is updated to apply the positional embedding and to remove the class token before the final output convolution. Additionally, `quant_conv` is renamed to `quant_linear` to accurately reflect the layer type.
Updates the `Magi1Decoder3d` from a convolutional design to a Transformer-like structure that operates on patches. This change replaces the initial convolutional and middle blocks with a linear projection layer, positional embeddings, and a class token. The logic for these components is moved from the parent `AutoencoderKLMagi1` model into the decoder for better encapsulation.
Removes several custom modules, including `Magi1ResidualBlock`, `Magi1Resample`, and `Magi1UpBlock`. Replaces the previous `Magi1MidBlock` with a more standard transformer-style `Magi1Block`. This change simplifies the overall VAE architecture by consolidating complex, specialized blocks into a more conventional design.
Replaces the custom `Magi1AttentionBlock` with the more generic `diffusers.Attention` module, combined with a new (?) `Magi1AttnProcessor2_0`. This change aligns the implementation with standard library patterns and leverages PyTorch 2.0's `scaled_dot_product_attention` for improved efficiency. The `Magi1Block` is also refactored into a more conventional transformer block structure using `Attention` and `FeedForward` modules.
Refactors the Magi1 VAE decoder to use a more standard transformer-based architecture. This change replaces the previous U-Net-like upsampling blocks with a series of standard transformer blocks, each containing self-attention and a feed-forward network. The custom rotary positional embedding logic and its helper functions have been removed, and the attention processor is simplified to work with the standard `Attention` module. This simplifies the overall model implementation.
Replaces the previous convolutional U-Net style encoder with a Vision Transformer (ViT) based implementation. This new architecture processes the input by dividing it into patches, adding positional embeddings, and then passing the sequence through a series of transformer blocks. The attention processor is also updated to support attention masks, and the model's configuration is adjusted to accommodate the new transformer-specific parameters.
Removes complex and unused parameters from the Magi1 VAE, encoder, and decoder modules. This change refactors the model to use a more standard Transformer architecture, eliminating the previous U-Net-like structure with dimension multipliers and residual blocks. The configuration is now more direct, improving clarity and maintainability.
Simplifies the initialization of the Magi1 VAE, encoder, and decoder. Reorders constructor parameters for clarity and removes unused arguments. The spatial and temporal compression ratios are now derived directly from the `patch_size` configuration, making the relationship more explicit. The pipeline is updated to use these new VAE attributes.
Simplifies the model architecture by removing the quantization and post-quantization convolution layers. This streamlines the `encode` and `decode` methods. The decoder is also updated to process the entire latent tensor at once, removing the previous frame-by-frame processing loop. Additionally, this change updates an import path for the `timm` library and renames an internal variable for consistency.
Updates the conversion script for the MAGI-1 VAE to correctly handle its Vision Transformer (ViT) based architecture. The state dictionary mapping is rewritten to align with the ViT structure. This includes adding logic to split the original checkpoint's combined QKV weights into separate query, key, and value tensors for the `diffusers` model. The model class and its configuration are also updated to reflect the appropriate ViT parameters, ensuring a correct conversion.
Renames the Magi autoencoder class to align with the "MAGI-1" model name. This refactoring improves consistency and clarity throughout the codebase, including documentation and tests.
Aligns the model naming with the source paper, "MAGI-1". This change refactors the model class, associated files, tests, and documentation to use the `Magi1` prefix for better clarity and consistency.
…ross multiple files
Improve compatibility by handling various PyTorch checkpoint formats. The loader now correctly extracts the state dictionary when it is nested under common keys like "model" or "state_dict". Ensure consistent loading of sharded safetensors by sorting the checkpoint files before merging them.
Modifies the `Magi1Encoder3d` to output both the mean and log variance of the latent distribution, aligning it with a standard variational autoencoder architecture. The initial linear projection layer is replaced with a convolutional patch embedding layer. Additionally, unused convolution index state variables are removed from the tiled encoding and decoding methods, simplifying the logic.
…n execution block
…ponents - Enhanced the convert_magi1_to_diffusers.py script with improved logging and error handling. - Updated key mapping dictionaries for transformer weights to ensure consistency with diffusers format. - Added detailed docstrings to classes and methods in transformer_magi1.py for better clarity. - Modified the Magi1Transformer3DModel to support new parameters and improved validation. - Adjusted test cases in test_models_transformer_magi1.py to reflect changes in parameter names. Refactor MAGI-1 conversion script and transformer model Improves the MAGI-1 conversion script by adding support for sharded checkpoints, introducing a more systematic key-mapping approach, and enhancing logging. Refactors the `Magi1Transformer3DModel` to align its architecture and configuration with the original model. This includes updating parameter names, adding input validation, and improving patch embedding and unpatchifying logic. Adds comprehensive docstrings to the model and component classes for better clarity and maintainability.
Overhauls the MAGI-1 to Diffusers conversion script to ensure correct model loading. The previous implementation used a generic key-renaming approach which was incomplete. This update replaces it with an explicit, component-by-component conversion process based on a direct analysis of the 4.5B model checkpoint. Key improvements include: - Updates the default transformer configuration to match the actual model architecture. - Rewrites the state dictionary conversion logic to handle specific parameter shapes and splits, such as for the cross-attention K/V projections. - Restructures key mappings for clarity and correctness.
Replaces the `rearrange` function with equivalent native PyTorch `permute` and `reshape` operations. This change removes the external `einops` library dependency, simplifying the model's environment.
|
Hi @yiyixuxu, @a-r-r-o-w. Is this model supposed to be placed in |
|
I can help review the VAE code, if you'd like. |
|
Please! Also its parameter conversion script. |
…ses for clarity Refactor: Rename VAE components for clarity Renames the attention processor and transformer block classes to be more specific to the VAE architecture. This improves code readability by making the purpose of these components more explicit.
# Copyright 2025 The Sand AI Team 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.
import collections
import math
from itertools import repeat
from functools import lru_cache
from typing import List, Literal, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint
from ...configuration_utils import ConfigMixin, register_to_config
from ...loaders import FromOriginalModelMixin
from ...utils import logging
from ...utils.accelerate_utils import apply_forward_hook
from ..activations import get_activation
from ..modeling_outputs import AutoencoderKLOutput
from ..modeling_utils import ModelMixin
from .vae import DecoderOutput, DiagonalGaussianDistribution
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
def ndgrid(*tensors) -> Tuple[torch.Tensor, ...]:
"""generate N-D grid in dimension order.
The ndgrid function is like meshgrid except that the order of the first two input arguments are switched.
That is, the statement
[X1,X2,X3] = ndgrid(x1,x2,x3)
produces the same result as
[X2,X1,X3] = meshgrid(x2,x1,x3)
This naming is based on MATLAB, the purpose is to avoid confusion due to torch's change to make
torch.meshgrid behaviour move from matching ndgrid ('ij') indexing to numpy meshgrid defaults of ('xy').
"""
try:
return torch.meshgrid(*tensors, indexing='ij')
except TypeError:
# old PyTorch < 1.10 will follow this path as it does not have indexing arg,
# the old behaviour of meshgrid was 'ij'
return torch.meshgrid(*tensors)
def freq_bands(
num_bands: int, temperature: float = 10000.0, step: int = 2, device: Optional[torch.device] = None
) -> torch.Tensor:
exp = torch.arange(0, num_bands, step, dtype=torch.int64, device=device).to(torch.float32) / num_bands
bands = 1.0 / (temperature**exp)
return bands
def pixel_freq_bands(
num_bands: int, max_freq: float = 224.0, linear_bands: bool = True, device: Optional[torch.device] = None
):
if linear_bands:
bands = torch.linspace(1.0, max_freq / 2, num_bands, dtype=torch.float32, device=device)
else:
bands = 2 ** torch.linspace(0, math.log(max_freq, 2) - 1, num_bands, dtype=torch.float32, device=device)
return bands * torch.pi
def build_fourier_pos_embed(
feat_shape: List[int],
bands: Optional[torch.Tensor] = None,
num_bands: int = 64,
max_res: int = 224,
temperature: float = 10000.0,
linear_bands: bool = False,
include_grid: bool = False,
in_pixels: bool = True,
ref_feat_shape: Optional[List[int]] = None,
dtype: torch.dtype = torch.float32,
device: Optional[torch.device] = None,
center_imgidx=True,
) -> List[torch.Tensor]:
"""
Args:
feat_shape: Feature shape for embedding.
bands: Pre-calculated frequency bands.
num_bands: Number of frequency bands (determines output dim).
max_res: Maximum resolution for pixel based freq.
temperature: Temperature for non-pixel freq.
linear_bands: Linear band spacing for pixel based freq.
include_grid: Include the spatial grid in output.
in_pixels: Output in pixel freq.
ref_feat_shape: Reference feature shape for resize / fine-tune.
dtype: Output dtype.
device: Output device.
Returns:
"""
if bands is None:
if in_pixels:
bands = pixel_freq_bands(num_bands, float(max_res), linear_bands=linear_bands, device=device)
else:
bands = freq_bands(num_bands, temperature=temperature, step=1, device=device)
else:
if device is None:
device = bands.device
if dtype is None:
dtype = bands.dtype
if in_pixels:
t = [torch.linspace(-1.0, 1.0, steps=s, device=device, dtype=torch.float32) for s in feat_shape]
else:
if center_imgidx:
t = [
torch.arange(s, device=device, dtype=torch.int64).to(torch.float32) - (s - 1) / 2
if len(feat_shape) == 2 or i != 0
else torch.arange(s, device=device, dtype=torch.int64).to(torch.float32)
for i, s in enumerate(feat_shape)
]
else:
t = [torch.arange(s, device=device, dtype=torch.int64).to(torch.float32) for s in feat_shape]
if ref_feat_shape is not None:
assert len(feat_shape) == len(ref_feat_shape), 'shape must be in same dimension'
# eva's scheme for resizing rope embeddings (ref shape = pretrain)
t = [x / f * r for x, f, r in zip(t, feat_shape, ref_feat_shape)]
grid = torch.stack(ndgrid(t), dim=-1)
grid = grid.unsqueeze(-1)
pos = grid * bands
pos_sin, pos_cos = pos.sin().to(dtype=dtype), pos.cos().to(dtype)
out = [grid, pos_sin, pos_cos] if include_grid else [pos_sin, pos_cos]
return out
def rot(x):
return torch.stack([-x[..., 1::2], x[..., ::2]], -1).reshape(x.shape)
def apply_rot_embed(x: torch.Tensor, sin_emb, cos_emb):
if sin_emb.ndim == 3:
return x * cos_emb.unsqueeze(1).expand_as(x) + rot(x) * sin_emb.unsqueeze(1).expand_as(x)
# import ipdb; ipdb.set_trace()
return x * cos_emb + rot(x) * sin_emb
def build_rotary_pos_embed(
feat_shape: List[int],
bands: Optional[torch.Tensor] = None,
dim: int = 64,
max_res: int = 224,
temperature: float = 10000.0,
linear_bands: bool = False,
in_pixels: bool = True,
ref_feat_shape: Optional[List[int]] = None,
dtype: torch.dtype = torch.float32,
device: Optional[torch.device] = None,
center_imgidx=True,
):
"""
Args:
feat_shape: Spatial shape of the target tensor for embedding.
bands: Optional pre-generated frequency bands
dim: Output dimension of embedding tensor.
max_res: Maximum resolution for pixel mode.
temperature: Temperature (inv freq) for non-pixel mode
linear_bands: Linearly (instead of log) spaced bands for pixel mode
in_pixels: Pixel vs language (inv freq) mode.
dtype: Output dtype.
device: Output device.
Returns:
"""
sin_emb, cos_emb = build_fourier_pos_embed(
feat_shape,
bands=bands,
num_bands=dim // (len(feat_shape) * 2),
max_res=max_res,
temperature=temperature,
linear_bands=linear_bands,
in_pixels=in_pixels,
ref_feat_shape=ref_feat_shape,
device=device,
dtype=dtype,
center_imgidx=center_imgidx,
)
num_spatial_dim = 1
# this would be much nicer as a .numel() call to torch.Size(), but torchscript sucks
for x in feat_shape:
num_spatial_dim *= x
sin_emb = sin_emb.reshape(num_spatial_dim, -1).repeat_interleave(2, -1)
cos_emb = cos_emb.reshape(num_spatial_dim, -1).repeat_interleave(2, -1)
return sin_emb, cos_emb
class Magi1Mlp(nn.Module):
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class Magi1ManualLayerNorm(nn.Module):
def __init__(self, normalized_shape, eps=1e-5, elementwise_affine=True):
super(Magi1ManualLayerNorm, self).__init__()
self.normalized_shape = normalized_shape
self.eps = eps
self.elementwise_affine = elementwise_affine
def forward(self, x):
mean = x.mean(dim=-1, keepdim=True)
std = x.std(dim=-1, keepdim=True, unbiased=False)
x_normalized = (x - mean) / (std + self.eps)
return x_normalized
@lru_cache(maxsize=50)
def cache_rotary_emb(feat_shape, device='cuda', dim=64, dtype=torch.bfloat16, max_res=512, ref_feat_shape=(4, 16, 16)):
return build_rotary_pos_embed(
feat_shape=feat_shape,
dim=dim,
max_res=max_res,
in_pixels=False,
ref_feat_shape=ref_feat_shape,
device=device,
dtype=dtype,
)
class Magi1Attention(nn.Module):
def __init__(
self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.0, proj_drop=0.0, ln_in_attn=False, use_rope=False
):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop_rate = attn_drop
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
if ln_in_attn:
self.qkv_norm = Magi1ManualLayerNorm(head_dim, elementwise_affine=False)
else:
self.qkv_norm = nn.Identity()
self.use_rope = use_rope
def forward(self, x, feat_shape=None):
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
qkv = self.qkv_norm(qkv)
q, k, v = qkv.chunk(3, dim=2)
if self.use_rope:
assert feat_shape is not None
q, k, v = qkv.chunk(3, dim=2)
rope_emb = cache_rotary_emb(feat_shape=feat_shape, dim=C // self.num_heads, device=x.device, dtype=x.dtype)
sin_emb = rope_emb[0].unsqueeze(0).unsqueeze(2)
cos_emb = rope_emb[1].unsqueeze(0).unsqueeze(2)
print(q.shape, sin_emb.shape)
q[:, 1:, :] = apply_rot_embed(q[:, 1:, :], sin_emb, cos_emb).bfloat16()
k[:, 1:, :] = apply_rot_embed(k[:, 1:, :], sin_emb, cos_emb).bfloat16()
from flash_attn import flash_attn_func
x = flash_attn_func(q, k, v, dropout_p=self.attn_drop_rate)
# TODO: Remove flash_attn_func and use F.scaled_dot_product_attention
# x = F.scaled_dot_product_attention(q, k, v, dropout_p=self.attn_drop_rate)
else:
from flash_attn import flash_attn_qkvpacked_func
x = flash_attn_qkvpacked_func(qkv=qkv.bfloat16(), dropout_p=self.attn_drop_rate)
# TODO: Remove flash_attn_qkvpacked_func and use F.scaled_dot_product_attention
# x = F.scaled_dot_product_attention(q.bfloat16(), k.bfloat16(), v.bfloat16(), dropout_p=self.attn_drop_rate)
# x = v
x = x.reshape(B, N, C)
# import ipdb; ipdb.set_trace()
x = self.proj(x)
x = self.proj_drop(x)
return x
class Magi1Block(nn.Module):
def __init__(
self,
dim,
num_heads,
mlp_ratio=4.0,
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
ln_in_attn=False,
use_rope=False,
):
super().__init__()
if not ln_in_attn:
self.norm1 = norm_layer(dim)
else:
self.norm1 = nn.Identity()
self.attn = Magi1Attention(
dim,
num_heads=num_heads,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop=attn_drop,
proj_drop=drop,
ln_in_attn=ln_in_attn,
use_rope=use_rope,
)
self.drop_path = nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Magi1Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
def forward(self, x, feat_shape=None):
x = x + self.drop_path(self.attn(self.norm1(x), feat_shape=feat_shape))
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
def to_2tuple(x):
if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
return tuple(x)
return tuple(repeat(x, 2))
class Magi1PatchEmbed(nn.Module):
"""Image to Patch Embedding"""
def __init__(self, video_size=224, video_length=16, patch_size=16, patch_length=1, in_chans=3, embed_dim=768):
super().__init__()
video_size = to_2tuple(video_size)
patch_size = to_2tuple(patch_size)
num_patches = (video_length // patch_length) * (video_size[1] // patch_size[1]) * (video_size[0] // patch_size[0])
self.video_size = video_size
self.patch_size = patch_size
self.video_length = video_length
self.patch_length = patch_length
self.num_patches = num_patches
self.proj = nn.Conv3d(
in_chans,
embed_dim,
kernel_size=(patch_length, patch_size[0], patch_size[1]),
stride=(patch_length, patch_size[0], patch_size[1]),
)
def forward(self, x):
"""
Forward pass of the PatchEmbed module.
Args:
x (torch.Tensor): Input tensor of shape (B, C, T, H, W), where
B is the batch size, C is the number of channels, T is the
number of frames, H is the height, and W is the width.
Returns:
torch.Tensor: Output tensor of shape (B, L, C'), where B is the
batch size, L is the number of tokens, and C' is the number
of output channels after flattening and transposing.
"""
B, C, T, H, W = x.shape
x = self.proj(x)
return x
def resize_pos_embed(posemb, src_shape, target_shape):
posemb = posemb.reshape(1, src_shape[0], src_shape[1], src_shape[2], -1)
posemb = posemb.permute(0, 4, 1, 2, 3)
posemb = nn.functional.interpolate(posemb, size=target_shape, mode='trilinear', align_corners=False)
posemb = posemb.permute(0, 2, 3, 4, 1)
posemb = posemb.reshape(1, target_shape[0] * target_shape[1] * target_shape[2], -1)
return posemb
class Magi1ViTEncoder(nn.Module):
"""Vision Transformer with support for patch or hybrid CNN input stage"""
def __init__(
self,
video_size=256,
video_length=16,
patch_size=8,
patch_length=4,
in_chans=3,
z_chans=4,
double_z=True,
embed_dim=768,
depth=12,
num_heads=12,
mlp_ratio=4.0,
qkv_bias=False,
qk_scale=None,
drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.0,
norm_layer=nn.LayerNorm,
with_cls_token=True,
norm_code=False,
ln_in_attn=False,
conv_last_layer=False,
use_rope=False,
use_final_proj=False,
):
super().__init__()
conv_last_layer = False # duplicate argument
# self.num_classes = num_classes
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
self.latent_size = video_size // patch_size
self.latent_length = video_length // patch_length
self.patch_embed = Magi1PatchEmbed(
video_size=video_size,
video_length=video_length,
patch_size=patch_size,
patch_length=patch_length,
in_chans=in_chans,
embed_dim=embed_dim,
)
num_patches = self.patch_embed.num_patches
self.with_cls_token = with_cls_token
if with_cls_token:
self.cls_token_nums = 1
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
else:
self.cls_token_nums = 0
self.cls_token = None
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.cls_token_nums, embed_dim))
self.pos_drop = nn.Dropout(p=drop_rate)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
self.blocks = nn.ModuleList(
[
Magi1Block(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[i],
norm_layer=norm_layer,
ln_in_attn=ln_in_attn,
use_rope=use_rope,
)
for i in range(depth)
]
)
self.norm = norm_layer(embed_dim)
self.norm_code = norm_code
self.out_channels = z_chans * 2 if double_z else z_chans
self.last_layer = nn.Linear(embed_dim, self.out_channels, bias=True)
nn.init.trunc_normal_(self.pos_embed, std=0.02)
if self.with_cls_token:
nn.init.trunc_normal_(self.cls_token, std=0.02)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
nn.init.trunc_normal_(m.weight, std=0.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
return {'pos_embed', 'cls_token'}
def forward(self, x):
B = x.shape[0]
# B C T H W -> B C T/pT H/pH W//pW
x = self.patch_embed(x)
latentT, latentH, latentW = x.shape[2], x.shape[3], x.shape[4]
# B C T/pT H/pH W//pW -> B (T/pT H/pH W//pW) C
x = x.flatten(2).transpose(1, 2)
if self.with_cls_token:
cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks
x = torch.cat((cls_tokens, x), dim=1)
if latentT != self.latent_length or latentH != self.latent_size or latentW != self.latent_size:
pos_embed = resize_pos_embed(
self.pos_embed[:, 1:, :],
src_shape=(self.latent_length, self.latent_size, self.latent_size),
target_shape=(latentT, latentH, latentW),
)
pos_embed = torch.cat((self.pos_embed[:, 0:1, :], pos_embed), dim=1)
else:
pos_embed = self.pos_embed
x = x + pos_embed
x = self.pos_drop(x)
for idx, blk in enumerate(self.blocks):
x = blk(x, feat_shape=(latentT, latentH, latentW))
x = self.norm(x)
x = self.last_layer(x)
if self.with_cls_token:
x = x[:, 1:] # remove cls_token
# B L C - > B , lT, lH, lW, zC
x = x.reshape(B, latentT, latentH, latentW, self.out_channels)
# B , lT, lH, lW, zC -> B, zC, lT, lH, lW
x = x.permute(0, 4, 1, 2, 3)
if self.norm_code:
prev_dtype = x.dtype
x = x.float()
x = x / torch.norm(x, dim=1, keepdim=True)
x = x.to(prev_dtype)
return x
def freeze_pretrain(self):
# Freeze all parameters
for param in self.parameters():
param.requires_grad = False
class Magi1ViTDecoder(nn.Module):
"""Vision Transformer with support for patch or hybrid CNN input stage"""
def __init__(
self,
video_size=256,
video_length=16,
patch_size=8,
patch_length=4,
in_chans=3,
z_chans=4,
double_z=True,
embed_dim=768,
depth=12,
num_heads=12,
mlp_ratio=4.0,
qkv_bias=False,
qk_scale=None,
drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.0,
norm_layer=nn.LayerNorm,
with_cls_token=True,
norm_code=False,
ln_in_attn=False,
conv_last_layer=False,
use_rope=False,
use_final_proj=False,
):
super().__init__()
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
self.latent_size = video_size // patch_size
self.latent_length = video_length // patch_length
self.patch_size = patch_size
self.patch_length = patch_length
self.proj_in = nn.Linear(z_chans, embed_dim)
num_patches = self.latent_size * self.latent_size * self.latent_length
self.with_cls_token = with_cls_token
if with_cls_token:
self.cls_token_nums = 1
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
else:
self.cls_token_nums = 0
self.cls_token = None
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.cls_token_nums, embed_dim))
self.pos_drop = nn.Dropout(p=drop_rate)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
self.blocks = nn.ModuleList(
[
Magi1Block(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[i],
norm_layer=norm_layer,
ln_in_attn=ln_in_attn,
use_rope=use_rope,
)
for i in range(depth)
]
)
self.norm = norm_layer(embed_dim)
assert conv_last_layer == True, "Only support conv_last_layer=True"
self.unpatch_channels = embed_dim // (patch_size * patch_size * patch_length)
self.final_proj = nn.Identity()
self.final_norm = nn.Identity()
self.use_final_proj = use_final_proj
if self.use_final_proj:
self.unpatch_channels = 4
self.final_proj = nn.Linear(embed_dim, self.unpatch_channels * (patch_size * patch_size * patch_length), bias=True)
self.final_norm = norm_layer(self.unpatch_channels * (patch_size * patch_size * patch_length))
self.last_layer = nn.Conv3d(in_channels=self.unpatch_channels, out_channels=3, kernel_size=3, stride=1, padding=1)
nn.init.trunc_normal_(self.pos_embed, std=0.02)
if self.with_cls_token:
nn.init.trunc_normal_(self.cls_token, std=0.02)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
nn.init.trunc_normal_(m.weight, std=0.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
return {'pos_embed', 'cls_token'}
def forward(self, x):
B, C, latentT, latentH, latentW = x.shape # x: (B, C, latentT, latentH, latenW)
x = x.permute(0, 2, 3, 4, 1) # x: (B, latentT, latentH, latenW, C)
x = x.reshape(B, -1, C)
x = self.proj_in(x)
if self.with_cls_token:
cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks
x = torch.cat((cls_tokens, x), dim=1)
if latentT != self.latent_length or latentH != self.latent_size or latentW != self.latent_size:
pos_embed = resize_pos_embed(
self.pos_embed[:, 1:, :],
src_shape=(self.latent_length, self.latent_size, self.latent_size),
target_shape=(latentT, latentH, latentW),
)
pos_embed = torch.cat((self.pos_embed[:, 0:1, :], pos_embed), dim=1)
else:
pos_embed = self.pos_embed
x = x + pos_embed
x = self.pos_drop(x)
for idx, blk in enumerate(self.blocks):
x = blk(x, feat_shape=(latentT, latentH, latentW))
x = self.norm(x)
if self.with_cls_token:
x = x[:, 1:] # remove cls_token
# B L C - > B, lT, lH, lW, pT, pH, pW, C
if self.use_final_proj:
x = self.final_proj(x)
x = self.final_norm(x)
x = x.reshape(B, latentT, latentH, latentW, self.patch_length, self.patch_size, self.patch_size, self.unpatch_channels).contiguous()
x = x.permute(0, 7, 1, 4, 2, 5, 3, 6)
x = x.reshape(B, self.unpatch_channels, latentT * self.patch_length, latentH * self.patch_size, latentW * self.patch_size)
x = self.last_layer(x)
return x
class AutoencoderKLMagi1(ModelMixin, ConfigMixin, FromOriginalModelMixin):
@register_to_config
def __init__(
self,
ddconfig: dict,
model_type: Literal['vit', 'vit_ncthw'] = 'vit',
use_rope: bool = False,
use_final_proj: bool = False,
):
super().__init__()
if model_type == 'vit':
self.encoder = Magi1ViTEncoder(**ddconfig)
self.decoder = Magi1ViTDecoder(**ddconfig)
elif model_type == 'vit_ncthw':
raise ValueError("model_type 'vit_ncthw' not supported")
else:
raise ValueError(f"model_type {model_type} not supported")
if 'patch_length' in ddconfig:
self._temporal_downsample_factor = ddconfig['patch_length']
else:
self._temporal_downsample_factor = 1
if 'patch_size' in ddconfig:
self._spatial_downsample_factor = ddconfig['patch_size']
else:
self._spatial_downsample_factor = 8
@property
def spatial_downsample_factor(self):
return self._spatial_downsample_factor
@property
def temporal_downsample_factor(self):
return self._temporal_downsample_factor
def encode(self, x, sample_posterior=True):
"""
Encode the input video.
Args:
x (torch.Tensor): Input video tensor has shape N C T H W
Returns:
tuple: Tuple containing the quantized tensor, embedding loss, and additional information.
"""
N, C, T, H, W = x.shape
if T == 1 and self._temporal_downsample_factor > 1:
x = x.expand(-1, -1, 4, -1, -1)
x = self.encoder(x)
posterior = DiagonalGaussianDistribution(x)
if sample_posterior:
z = posterior.sample()
else:
z = posterior.mode()
return z[:, :, :1, :, :].type(x.dtype)
else:
x = self.encoder(x)
posterior = DiagonalGaussianDistribution(x)
if sample_posterior:
z = posterior.sample()
else:
z = posterior.mode()
return z.type(x.dtype)
def decode(self, x):
"""
Decode the quantized tensor.
Args:
quant (torch.Tensor): Quantized tensor.
Returns:
torch.Tensor: Decoded tensor.
"""
N, C, T, H, W = x.shape
if T == 1:
x = x.expand(-1, -1, 1, -1, -1)
x = self.decoder(x)
x = x[:, :, :1, :, :]
return x
else:
x = self.decoder(x)
return x
def forward(self, x, sample_posterior=True):
x = self.encoder(x)
posterior = DiagonalGaussianDistribution(x)
if sample_posterior:
z = posterior.sample()
else:
z = posterior.mode()
dec = self.decoder(z)
return dec, posterior
def get_last_layer(self):
"""
Get the last layer of the decoder.
Returns:
torch.Tensor: Last layer of the decoder.
"""
return self.decoder.last_layer.weight
@property
def allow_spatial_tiling(self):
return FalseThis is my modification to My coding philosophy is as follows:
I have tested this code, and it successfully outputs videos. The |
|
The test code is as follows. You can use any video for testing. from diffusers import AutoencoderKLMagi1
from diffusers.utils import export_to_video, load_video
import torch
from PIL import Image
import numpy as np
from torchvision import transforms
video_path = "curry_vs_thunder.mp4"
vae = AutoencoderKLMagi1.from_pretrained("sand-ai/MAGI-1", subfolder="ckpt/vae", torch_dtype=torch.bfloat16).to("cuda")
image_list = load_video(video_path)[:100]
transform = transforms.Compose([
transforms.Resize((256, 256)),
transforms.CenterCrop((256, 256)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
])
frame_list = []
for image in image_list:
frame_list.append(transform(image).to("cuda"))
input_video = torch.stack(frame_list).to("cuda")
output_frame_list = []
with torch.no_grad() and torch.cuda.amp.autocast():
input_video = input_video.unsqueeze(0).transpose(1, 2)
output_video, posterior = vae(input_video)
output_video = output_video.squeeze(0)
for i in range(output_video.shape[1]):
output_frame = output_video[:, i, :, :].permute(1, 2, 0).cpu().detach().numpy()
# Denormalize from [-1, 1] to [0, 1]
output_frame = (output_frame + 1.0) * 0.5
# Convert to uint8
output_frame = (output_frame.clip(0, 1) * 255).astype(np.uint8)
output_frame = Image.fromarray(output_frame)
output_frame_list.append(output_frame)
export_to_video(output_frame_list, "output.mp4", fps=25) |
if self.use_rope:
assert feat_shape is not None
q, k, v = qkv.chunk(3, dim=2)
rope_emb = cache_rotary_emb(feat_shape=feat_shape, dim=C // self.num_heads, device=x.device, dtype=x.dtype)
sin_emb = rope_emb[0].unsqueeze(0).unsqueeze(2)
cos_emb = rope_emb[1].unsqueeze(0).unsqueeze(2)
print(q.shape, sin_emb.shape)
q[:, 1:, :] = apply_rot_embed(q[:, 1:, :], sin_emb, cos_emb).bfloat16()
k[:, 1:, :] = apply_rot_embed(k[:, 1:, :], sin_emb, cos_emb).bfloat16()
q, k, v = map(lambda t: t.squeeze(2).transpose(1, 2), (q, k, v))
x = F.scaled_dot_product_attention(q, k, v, dropout_p=self.attn_drop_rate)
x = x.transpose(1, 2)
else:
q, k, v = map(lambda t: t.squeeze(2).transpose(1, 2), (q, k, v))
x = F.scaled_dot_product_attention(q, k, v, dropout_p=self.attn_drop_rate)
x = x.transpose(1, 2)Solved. |
|
Thanks for the suggestions.
Right, I forgot to replace At first, I was trying to port positional embedding-related calculations, but then I realized that the shared models' vae doesn't use RoPE by default. Rather than Rather than Rather than In order to use I removed training-related parts such as I have also been trying to follow WDYT? |
|
I think it would be good to make the code follow the style of |
Dynamically resize positional embeddings in the encoder and decoder to match the input's latent shape using trilinear interpolation. This change allows the model to process inputs of varying height, width, and frame count in a single forward pass, removing the previous iterative encoding logic.
|
I was only focusing on the decoding part because inference is the priority. |
…toencoderKLMagi1.decode
Introduces a `ManualLayerNorm` class to provide an explicit, self-contained layer normalization. The VAE attention processor is updated to use this new manual normalization, removing the previous dependency.
Updates the `Magi1VAEAttnProcessor2_0` to remove the additive residual connection. This change simplifies the attention block's forward pass.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Thanks for the opportunity to fix #11519!
Original repo: https://github.com/SandAI-org/MAGI-1
AutoencoderKLMagi1Magi1Transformer3DModelMAGI-1-Diffusers
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR.