Files
Prince Canuma 17397da70c format
2026-03-18 17:40:05 +01:00

280 lines
9.0 KiB
Python

"""Sampling operations for Video VAE (upsampling/downsampling)."""
from typing import Tuple, Union
import mlx.core as mx
import mlx.nn as nn
from mlx_video.models.ltx_2.video_vae.convolution import CausalConv3d, PaddingModeType
class SpaceToDepthDownsample(nn.Module):
"""Space-to-depth downsampling with 3x3 conv and skip connection.
PyTorch-compatible implementation:
1. Apply 3x3 conv: in_channels -> out_channels // prod(stride)
2. Space-to-depth on conv output: channels * prod(stride)
3. Space-to-depth on input with group averaging for skip connection
4. Add skip connection
"""
def __init__(
self,
dims: int,
in_channels: int,
out_channels: int,
stride: Union[int, Tuple[int, int, int]],
spatial_padding_mode: PaddingModeType = PaddingModeType.ZEROS,
):
super().__init__()
if isinstance(stride, int):
stride = (stride, stride, stride)
self.stride = stride
self.dims = dims
self.out_channels = out_channels
# Calculate channels
multiplier = stride[0] * stride[1] * stride[2]
self.group_size = in_channels * multiplier // out_channels
conv_out_channels = out_channels // multiplier
# 3x3 convolution (not 1x1)
self.conv = CausalConv3d(
in_channels=in_channels,
out_channels=conv_out_channels,
kernel_size=3,
stride=1,
padding=1,
spatial_padding_mode=spatial_padding_mode,
)
def _space_to_depth(self, x: mx.array) -> mx.array:
"""Rearrange: b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w"""
b, c, d, h, w = x.shape
st, sh, sw = self.stride
# Reshape to group spatial elements
x = mx.reshape(x, (b, c, d // st, st, h // sh, sh, w // sw, sw))
# Permute: (B, C, D', st, H', sh, W', sw) -> (B, C, st, sh, sw, D', H', W')
x = mx.transpose(x, (0, 1, 3, 5, 7, 2, 4, 6))
# Reshape to combine channels
new_c = c * st * sh * sw
new_d = d // st
new_h = h // sh
new_w = w // sw
x = mx.reshape(x, (b, new_c, new_d, new_h, new_w))
return x
def __call__(self, x: mx.array, causal: bool = True) -> mx.array:
b, c, d, h, w = x.shape
st, sh, sw = self.stride
# Temporal padding for causal mode
if st == 2:
# Duplicate first frame for padding
x = mx.concatenate([x[:, :, :1, :, :], x], axis=2)
d = d + 1
# Pad if necessary to make dimensions divisible by stride
pad_d = (st - d % st) % st
pad_h = (sh - h % sh) % sh
pad_w = (sw - w % sw) % sw
if pad_d > 0 or pad_h > 0 or pad_w > 0:
x = mx.pad(x, [(0, 0), (0, 0), (0, pad_d), (0, pad_h), (0, pad_w)])
# Skip connection: space-to-depth on input, then group mean
x_in = self._space_to_depth(x)
# Reshape for group mean: (b, c*prod(stride), d, h, w) -> (b, out_channels, group_size, d, h, w)
b2, c2, d2, h2, w2 = x_in.shape
x_in = mx.reshape(x_in, (b2, self.out_channels, self.group_size, d2, h2, w2))
x_in = mx.mean(x_in, axis=2) # (b, out_channels, d, h, w)
# Conv branch: apply conv then space-to-depth
x_conv = self.conv(x, causal=causal)
x_conv = self._space_to_depth(x_conv)
# Add skip connection
return x_conv + x_in
class DepthToSpaceUpsample(nn.Module):
def __init__(
self,
dims: int,
in_channels: int,
stride: Union[int, Tuple[int, int, int]],
residual: bool = False,
out_channels_reduction_factor: int = 1,
spatial_padding_mode: PaddingModeType = PaddingModeType.ZEROS,
):
super().__init__()
if isinstance(stride, int):
stride = (stride, stride, stride)
self.stride = stride
self.dims = dims
self.residual = residual
self.out_channels_reduction_factor = out_channels_reduction_factor
# Calculate output channels
multiplier = stride[0] * stride[1] * stride[2]
out_channels = in_channels // out_channels_reduction_factor
self.out_channels = out_channels
# 3x3x3 convolution to prepare channels for unpacking (matches PyTorch)
self.conv = CausalConv3d(
in_channels=in_channels,
out_channels=out_channels * multiplier,
kernel_size=3,
stride=1,
padding=1,
spatial_padding_mode=spatial_padding_mode,
)
def _depth_to_space(self, x: mx.array) -> mx.array:
b, c_packed, d, h, w = x.shape
st, sh, sw = self.stride
c = c_packed // (st * sh * sw)
# (B, C*st*sh*sw, D, H, W) -> (B, C, st, sh, sw, D, H, W)
x = mx.reshape(x, (b, c, st, sh, sw, d, h, w))
# (B, C, st, sh, sw, D, H, W) -> (B, C, D, st, H, sh, W, sw)
x = mx.transpose(x, (0, 1, 5, 2, 6, 3, 7, 4))
# (B, C, D, st, H, sh, W, sw) -> (B, C, D*st, H*sh, W*sw)
x = mx.reshape(x, (b, c, d * st, h * sh, w * sw))
return x
def __call__(
self, x: mx.array, causal: bool = True, chunked_conv: bool = False
) -> mx.array:
b, c, d, h, w = x.shape
st, sh, sw = self.stride
# Compute residual path if enabled
x_residual = None
if self.residual:
# Reshape input: treat channels as spatial factors
# "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)"
x_residual = self._depth_to_space(x)
# Tile channels to match output (PyTorch .repeat() tiles, not element-repeat!)
# num_repeat = prod(stride) / out_channels_reduction_factor
num_repeat = (st * sh * sw) // self.out_channels_reduction_factor
x_residual = mx.tile(x_residual, (1, num_repeat, 1, 1, 1))
# Remove first temporal frame if temporal upsampling
if st > 1:
x_residual = x_residual[:, :, 1:, :, :]
# Use chunked mode for large tensors to reduce peak memory
if chunked_conv and d > 4:
x = self._chunked_conv_depth_to_space(x, causal)
else:
# Apply conv
x = self.conv(x, causal=causal)
# Depth to space rearrangement
x = self._depth_to_space(x)
# Remove first frame for causal temporal upsampling
if st > 1:
x = x[:, :, 1:, :, :]
# Add residual
if self.residual and x_residual is not None:
x = x + x_residual
return x
def _chunked_conv_depth_to_space(
self, x: mx.array, causal: bool = True
) -> mx.array:
"""Chunked conv + depth_to_space that processes in temporal chunks.
This reduces peak memory by avoiding the full high-channel intermediate tensor.
Instead of materializing (B, 4096, D, H, W), we process temporal chunks and
immediately apply depth_to_space.
Args:
x: Input tensor of shape (B, C, D, H, W)
causal: Whether to use causal convolutions
Returns:
Output tensor after conv + depth_to_space
"""
b, c, d, h, w = x.shape
st, sh, sw = self.stride
out_c = self.out_channels
# Output dimensions
out_d = d * st
out_h = h * sh
out_w = w * sw
# Chunk size in temporal dimension (process 4 frames at a time)
chunk_size = 4
kernel_t = 3 # Temporal kernel size
# For causal conv, we need (kernel_t - 1) frames of padding at the start
# For non-causal, we need (kernel_t - 1) // 2 on each side
if causal:
# Pad start with first frame repeated
pad_start = kernel_t - 1
pad_end = 0
else:
pad_start = (kernel_t - 1) // 2
pad_end = (kernel_t - 1) // 2
# Allocate output
outputs = []
# Process in chunks with overlap for conv kernel
t_pos = 0
while t_pos < d:
t_end = min(t_pos + chunk_size, d)
# Calculate input range with padding for kernel
in_start = max(0, t_pos - pad_start)
in_end = min(d, t_end + pad_end)
# Extract chunk
chunk = x[:, :, in_start:in_end, :, :]
# Apply conv to chunk
chunk_conv = self.conv(chunk, causal=causal)
# Apply depth_to_space
chunk_out = self._depth_to_space(chunk_conv)
# Calculate valid output range (excluding padding effects)
# Each input frame produces st output frames
out_start = (t_pos - in_start) * st
out_end = out_start + (t_end - t_pos) * st
# Extract valid portion
chunk_out = chunk_out[:, :, out_start:out_end, :, :]
outputs.append(chunk_out)
# Evaluate to free intermediate memory
mx.eval(outputs[-1])
t_pos = t_end
# Concatenate all chunks
if len(outputs) == 1:
return outputs[0]
return mx.concatenate(outputs, axis=2)