Reduce LTX2 VAE VRAM consumption (#12028)

* causal_video_ae: Remove attention ResNet

This attention_head_dim argument does not exist on this constructor so
this is dead code. Remove as generic attention mid VAE conflicts with
temporal roll.

* ltx-vae: consoldate causal/non-causal code paths

* ltx-vae: add cache rolling adder

* ltx-vae: use cached adder for resnet

* ltx-vae: Implement rolling VAE

Implement a temporal rolling VAE for the LTX2 VAE.

Usually when doing temporal rolling VAEs you can just chunk on time relying
on causality and cache behind you as you go. The LTX VAE is however
non-causal.

So go whole hog and implement per layer run ahead and backpressure between
the decoder layers using recursive state beween the layers.

Operations are ammended with temporal_cache_state{} which they can use to
hold any state then need for partial execution. Convolutions cache their
inputs behind the up to N-1 frames, and skip connections need to cache the
mismatch between convolution input and output that happens due to missing
future (non-causal) input.

Each call to run_up() processes a layer accross a range on input that
may or may not be complete. It goes depth first to process as much as
possible to try and digest frames to the final output ASAP. If layers run
out of input due to convolution losses, they simply return without action
effectively applying back-pressure to the earlier layers. As the earlier
layers do more work and caller deeper, the partial states are reconciled
and output continues to digest depth first as much as possible.

Chunking is done using a size quota rather than a fixed frame length and
any layer can initiate chunking, and multiple layers can chunk at different
granulatiries. This remove the old limitation of always having to process
1 latent frame to entirety and having to hold 8 full decoded frames as
the VRAM peak.

comfy/ldm/lightricks/vae/causal_conv3d.py

@@ -1,11 +1,11 @@
 from typing import Tuple, Union
 
+import threading
 import torch
 import torch.nn as nn
 import comfy.ops
 ops = comfy.ops.disable_weight_init
 
-
 class CausalConv3d(nn.Module):
     def __init__(
         self,
@@ -42,23 +42,34 @@ class CausalConv3d(nn.Module):
             padding_mode=spatial_padding_mode,
             groups=groups,
         )
+        self.temporal_cache_state={}
 
     def forward(self, x, causal: bool = True):
-        if causal:
+        tid = threading.get_ident()
-            first_frame_pad = x[:, :, :1, :, :].repeat(
+
-                (1, 1, self.time_kernel_size - 1, 1, 1)
+        cached, is_end = self.temporal_cache_state.get(tid, (None, False))
-            )
+        if cached is None:
-            x = torch.concatenate((first_frame_pad, x), dim=2)
+            padding_length = self.time_kernel_size - 1
-        else:
+            if not causal:
-            first_frame_pad = x[:, :, :1, :, :].repeat(
+                padding_length = padding_length // 2
-                (1, 1, (self.time_kernel_size - 1) // 2, 1, 1)
+            if x.shape[2] == 0:
-            )
+                return x
-            last_frame_pad = x[:, :, -1:, :, :].repeat(
+            cached = x[:, :, :1, :, :].repeat((1, 1, padding_length, 1, 1))
-                (1, 1, (self.time_kernel_size - 1) // 2, 1, 1)
+        pieces = [ cached, x ]
-            )
+        if is_end and not causal:
-            x = torch.concatenate((first_frame_pad, x, last_frame_pad), dim=2)
+            pieces.append(x[:, :, -1:, :, :].repeat((1, 1, (self.time_kernel_size - 1) // 2, 1, 1)))
-        x = self.conv(x)
+
-        return x
+        needs_caching = not is_end
+        if needs_caching and x.shape[2] >= self.time_kernel_size - 1:
+            needs_caching = False
+            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+
+        x = torch.cat(pieces, dim=2)
+
+        if needs_caching:
+            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+
+        return self.conv(x) if x.shape[2] >= self.time_kernel_size else x[:, :, :0, :, :]
 
     @property
     def weight(self):

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -1,4 +1,5 @@
 from __future__ import annotations
+import threading
 import torch
 from torch import nn
 from functools import partial
@@ -6,12 +7,35 @@ import math
 from einops import rearrange
 from typing import List, Optional, Tuple, Union
 from .conv_nd_factory import make_conv_nd, make_linear_nd
+from .causal_conv3d import CausalConv3d
 from .pixel_norm import PixelNorm
 from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
 import comfy.ops
+from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed
 
 ops = comfy.ops.disable_weight_init
 
+def mark_conv3d_ended(module):
+    tid = threading.get_ident()
+    for _, m in module.named_modules():
+        if isinstance(m, CausalConv3d):
+            current = m.temporal_cache_state.get(tid, (None, False))
+            m.temporal_cache_state[tid] = (current[0], True)
+
+def split2(tensor, split_point, dim=2):
+    return torch.split(tensor, [split_point, tensor.shape[dim] - split_point], dim=dim)
+
+def add_exchange_cache(dest, cache_in, new_input, dim=2):
+    if dest is not None:
+        if cache_in is not None:
+            cache_to_dest = min(dest.shape[dim], cache_in.shape[dim])
+            lead_in_dest, dest = split2(dest, cache_to_dest, dim=dim)
+            lead_in_source, cache_in = split2(cache_in, cache_to_dest, dim=dim)
+            lead_in_dest.add_(lead_in_source)
+        body, new_input = split2(new_input, dest.shape[dim], dim)
+        dest.add_(body)
+    return torch_cat_if_needed([cache_in, new_input], dim=dim)
+
 class Encoder(nn.Module):
     r"""
     The `Encoder` layer of a variational autoencoder that encodes its input into a latent representation.
@@ -205,7 +229,7 @@ class Encoder(nn.Module):
 
         self.gradient_checkpointing = False
 
-    def forward(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+    def forward_orig(self, sample: torch.FloatTensor) -> torch.FloatTensor:
         r"""The forward method of the `Encoder` class."""
 
         sample = patchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
@@ -254,6 +278,22 @@ class Encoder(nn.Module):
 
         return sample
 
+    def forward(self, *args, **kwargs):
+        #No encoder support so just flag the end so it doesnt use the cache.
+        mark_conv3d_ended(self)
+        try:
+            return self.forward_orig(*args, **kwargs)
+        finally:
+            tid = threading.get_ident()
+            for _, module in self.named_modules():
+                # ComfyUI doesn't thread this kind of stuff today, but just in case
+                # we key on the thread to make it thread safe.
+                tid = threading.get_ident()
+                if hasattr(module, "temporal_cache_state"):
+                    module.temporal_cache_state.pop(tid, None)
+
+
+MAX_CHUNK_SIZE=(128 * 1024 ** 2)
 
 class Decoder(nn.Module):
     r"""
@@ -341,18 +381,6 @@ class Decoder(nn.Module):
                     timestep_conditioning=timestep_conditioning,
                     spatial_padding_mode=spatial_padding_mode,
                 )
-            elif block_name == "attn_res_x":
-                block = UNetMidBlock3D(
-                    dims=dims,
-                    in_channels=input_channel,
-                    num_layers=block_params["num_layers"],
-                    resnet_groups=norm_num_groups,
-                    norm_layer=norm_layer,
-                    inject_noise=block_params.get("inject_noise", False),
-                    timestep_conditioning=timestep_conditioning,
-                    attention_head_dim=block_params["attention_head_dim"],
-                    spatial_padding_mode=spatial_padding_mode,
-                )
             elif block_name == "res_x_y":
                 output_channel = output_channel // block_params.get("multiplier", 2)
                 block = ResnetBlock3D(
@@ -428,8 +456,9 @@ class Decoder(nn.Module):
             )
             self.last_scale_shift_table = nn.Parameter(torch.empty(2, output_channel))
 
+
     # def forward(self, sample: torch.FloatTensor, target_shape) -> torch.FloatTensor:
-    def forward(
+    def forward_orig(
         self,
         sample: torch.FloatTensor,
         timestep: Optional[torch.Tensor] = None,
@@ -437,6 +466,7 @@ class Decoder(nn.Module):
         r"""The forward method of the `Decoder` class."""
         batch_size = sample.shape[0]
 
+        mark_conv3d_ended(self.conv_in)
         sample = self.conv_in(sample, causal=self.causal)
 
         checkpoint_fn = (
@@ -445,24 +475,12 @@ class Decoder(nn.Module):
             else lambda x: x
         )
 
-        scaled_timestep = None
+        timestep_shift_scale = None
         if self.timestep_conditioning:
             assert (
                 timestep is not None
             ), "should pass timestep with timestep_conditioning=True"
             scaled_timestep = timestep * self.timestep_scale_multiplier.to(dtype=sample.dtype, device=sample.device)
-
-        for up_block in self.up_blocks:
-            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
-                sample = checkpoint_fn(up_block)(
-                    sample, causal=self.causal, timestep=scaled_timestep
-                )
-            else:
-                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
-
-        sample = self.conv_norm_out(sample)
-
-        if self.timestep_conditioning:
             embedded_timestep = self.last_time_embedder(
                 timestep=scaled_timestep.flatten(),
                 resolution=None,
@@ -483,16 +501,62 @@ class Decoder(nn.Module):
                 embedded_timestep.shape[-2],
                 embedded_timestep.shape[-1],
             )
-            shift, scale = ada_values.unbind(dim=1)
+            timestep_shift_scale = ada_values.unbind(dim=1)
-            sample = sample * (1 + scale) + shift
 
-        sample = self.conv_act(sample)
+        output = []
-        sample = self.conv_out(sample, causal=self.causal)
+
+        def run_up(idx, sample, ended):
+            if idx >= len(self.up_blocks):
+                sample = self.conv_norm_out(sample)
+                if timestep_shift_scale is not None:
+                    shift, scale = timestep_shift_scale
+                    sample = sample * (1 + scale) + shift
+                sample = self.conv_act(sample)
+                if ended:
+                    mark_conv3d_ended(self.conv_out)
+                sample = self.conv_out(sample, causal=self.causal)
+                if sample is not None and sample.shape[2] > 0:
+                    output.append(sample)
+                return
+
+            up_block = self.up_blocks[idx]
+            if (ended):
+                mark_conv3d_ended(up_block)
+            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
+                sample = checkpoint_fn(up_block)(
+                    sample, causal=self.causal, timestep=scaled_timestep
+                )
+            else:
+                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
+
+            if sample is None or sample.shape[2] == 0:
+                return
+
+            total_bytes = sample.numel() * sample.element_size()
+            num_chunks = (total_bytes + MAX_CHUNK_SIZE - 1) // MAX_CHUNK_SIZE
+            samples = torch.chunk(sample, chunks=num_chunks, dim=2)
+
+            for chunk_idx, sample1 in enumerate(samples):
+                run_up(idx + 1, sample1, ended and chunk_idx == len(samples) - 1)
+
+        run_up(0, sample, True)
+        sample = torch.cat(output, dim=2)
 
         sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
 
         return sample
 
+    def forward(self, *args, **kwargs):
+        try:
+            return self.forward_orig(*args, **kwargs)
+        finally:
+            for _, module in self.named_modules():
+                #ComfyUI doesn't thread this kind of stuff today, but just incase
+                #we key on the thread to make it thread safe.
+                tid = threading.get_ident()
+                if hasattr(module, "temporal_cache_state"):
+                    module.temporal_cache_state.pop(tid, None)
+
 
 class UNetMidBlock3D(nn.Module):
     """
@@ -663,8 +727,22 @@ class DepthToSpaceUpsample(nn.Module):
         )
         self.residual = residual
         self.out_channels_reduction_factor = out_channels_reduction_factor
+        self.temporal_cache_state = {}
 
     def forward(self, x, causal: bool = True, timestep: Optional[torch.Tensor] = None):
+        tid = threading.get_ident()
+        cached, drop_first_conv, drop_first_res = self.temporal_cache_state.get(tid, (None, True, True))
+        y = self.conv(x, causal=causal)
+        y = rearrange(
+            y,
+            "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
+        if self.stride[0] == 2 and y.shape[2] > 0 and drop_first_conv:
+            y = y[:, :, 1:, :, :]
+            drop_first_conv = False
         if self.residual:
             # Reshape and duplicate the input to match the output shape
             x_in = rearrange(
@@ -676,21 +754,20 @@ class DepthToSpaceUpsample(nn.Module):
             )
             num_repeat = math.prod(self.stride) // self.out_channels_reduction_factor
             x_in = x_in.repeat(1, num_repeat, 1, 1, 1)
-            if self.stride[0] == 2:
+            if self.stride[0] == 2 and x_in.shape[2] > 0 and drop_first_res:
                 x_in = x_in[:, :, 1:, :, :]
-        x = self.conv(x, causal=causal)
+                drop_first_res = False
-        x = rearrange(
+
-            x,
+            if y.shape[2] == 0:
-            "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+                y = None
-            p1=self.stride[0],
+
-            p2=self.stride[1],
+            cached = add_exchange_cache(y, cached, x_in, dim=2)
-            p3=self.stride[2],
+            self.temporal_cache_state[tid] = (cached, drop_first_conv, drop_first_res)
-        )
+
-        if self.stride[0] == 2:
+        else:
-            x = x[:, :, 1:, :, :]
+            self.temporal_cache_state[tid] = (None, drop_first_conv, False)
-        if self.residual:
+
-            x = x + x_in
+        return y
-        return x
 
 class LayerNorm(nn.Module):
     def __init__(self, dim, eps, elementwise_affine=True) -> None:
@@ -807,6 +884,8 @@ class ResnetBlock3D(nn.Module):
                 torch.randn(4, in_channels) / in_channels**0.5
             )
 
+        self.temporal_cache_state={}
+
     def _feed_spatial_noise(
         self, hidden_states: torch.FloatTensor, per_channel_scale: torch.FloatTensor
     ) -> torch.FloatTensor:
@@ -880,9 +959,12 @@ class ResnetBlock3D(nn.Module):
 
         input_tensor = self.conv_shortcut(input_tensor)
 
-        output_tensor = input_tensor + hidden_states
+        tid = threading.get_ident()
+        cached = self.temporal_cache_state.get(tid, None)
+        cached = add_exchange_cache(hidden_states, cached, input_tensor, dim=2)
+        self.temporal_cache_state[tid] = cached
 
-        return output_tensor
+        return hidden_states
 
 
 def patchify(x, patch_size_hw, patch_size_t=1):

comfy/ldm/modules/diffusionmodules/model.py

@@ -14,10 +14,13 @@ if model_management.xformers_enabled_vae():
     import xformers.ops
 
 def torch_cat_if_needed(xl, dim):
+    xl = [x for x in xl if x is not None and x.shape[dim] > 0]
     if len(xl) > 1:
         return torch.cat(xl, dim)
-    else:
+    elif len(xl) == 1:
         return xl[0]
+    else:
+        return None
 
 def get_timestep_embedding(timesteps, embedding_dim):
     """