Implement temporal rolling VAE (Major VRAM reductions in Hunyuan and Kandinsky) (#10995)

* hunyuan upsampler: rework imports

Remove the transitive import of VideoConv3d and Resnet and takes these
from actual implementation source.

* model: remove unused give_pre_end

According to git grep, this is not used now, and was not used in the
initial commit that introduced it (see below).

This semantic is difficult to implement temporal roll VAE for (and would
defeat the purpose). Rather than implement the complex if, just delete
the unused feature.

(venv) rattus@rattus-box2:~/ComfyUI$ git log --oneline
220afe3 (HEAD) Initial commit.
(venv) rattus@rattus-box2:~/ComfyUI$ git grep give_pre
comfy/ldm/modules/diffusionmodules/model.py:                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
comfy/ldm/modules/diffusionmodules/model.py:        self.give_pre_end = give_pre_end
comfy/ldm/modules/diffusionmodules/model.py:        if self.give_pre_end:

(venv) rattus@rattus-box2:~/ComfyUI$ git co origin/master
Previous HEAD position was 220afe3 Initial commit.
HEAD is now at 9d8a817 Enable async offloading by default on Nvidia. (#10953)
(venv) rattus@rattus-box2:~/ComfyUI$ git grep give_pre
comfy/ldm/modules/diffusionmodules/model.py:                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
comfy/ldm/modules/diffusionmodules/model.py:        self.give_pre_end = give_pre_end
comfy/ldm/modules/diffusionmodules/model.py:        if self.give_pre_end:

* move refiner VAE temporal roller to core

Move the carrying conv op to the common VAE code and give it a better
name. Roll the carry implementation logic for Resnet into the base
class and scrap the Hunyuan specific subclass.

* model: Add temporal roll to main VAE decoder

If there are no attention layers, its a standard resnet and VideoConv3d
is asked for, substitute in the temporal rolloing VAE algorithm. This
reduces VAE usage by the temporal dimension (can be huge VRAM savings).

* model: Add temporal roll to main VAE encoder

If there are no attention layers, its a standard resnet and VideoConv3d
is asked for, substitute in the temporal rolling VAE algorithm. This
reduces VAE usage by the temporal dimension (can be huge VRAM savings).

Author: rattus128

comfy/ldm/hunyuan_video/upsampler.py

@@ -1,7 +1,8 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm, ResnetBlock, VideoConv3d
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d
+from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm
 import model_management, model_patcher
 
 class SRResidualCausalBlock3D(nn.Module):

comfy/ldm/hunyuan_video/vae_refiner.py

@@ -1,42 +1,12 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, CarriedConv3d, Normalize, conv_carry_causal_3d, torch_cat_if_needed
 import comfy.ops
 import comfy.ldm.models.autoencoder
 import comfy.model_management
 ops = comfy.ops.disable_weight_init
 
-class NoPadConv3d(nn.Module):
-    def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs):
-        super().__init__()
-        self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs)
-
-    def forward(self, x):
-        return self.conv(x)
-
-
-def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None):
-
-    x = xl[0]
-    xl.clear()
-
-    if conv_carry_out is not None:
-        to_push = x[:, :, -2:, :, :].clone()
-        conv_carry_out.append(to_push)
-
-    if isinstance(op, NoPadConv3d):
-        if conv_carry_in is None:
-            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate')
-        else:
-            carry_len = conv_carry_in[0].shape[2]
-            x = torch.cat([conv_carry_in.pop(0), x], dim=2)
-            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate')
-
-    out = op(x)
-
-    return out
-
 
 class RMS_norm(nn.Module):
     def __init__(self, dim):
@@ -49,7 +19,7 @@ def forward(self, x):
         return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device)
 
 class DnSmpl(nn.Module):
-    def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d):
+    def __init__(self, ic, oc, tds, refiner_vae, op):
         super().__init__()
         fct = 2 * 2 * 2 if tds else 1 * 2 * 2
         assert oc % fct == 0
@@ -109,7 +79,7 @@ def forward(self, x, conv_carry_in=None, conv_carry_out=None):
 
 
 class UpSmpl(nn.Module):
-    def __init__(self, ic, oc, tus=True, refiner_vae=True, op=VideoConv3d):
+    def __init__(self, ic, oc, tus, refiner_vae, op):
         super().__init__()
         fct = 2 * 2 * 2 if tus else 1 * 2 * 2
         self.conv = op(ic, oc * fct, kernel_size=3, stride=1, padding=1)
@@ -163,23 +133,6 @@ def forward(self, x, conv_carry_in=None, conv_carry_out=None):
 
         return h + x
 
-class HunyuanRefinerResnetBlock(ResnetBlock):
-    def __init__(self, in_channels, out_channels, conv_op=NoPadConv3d, norm_op=RMS_norm):
-        super().__init__(in_channels=in_channels, out_channels=out_channels, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
-
-    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
-        h = x
-        h = [ self.swish(self.norm1(x)) ]
-        h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
-
-        h = [ self.dropout(self.swish(self.norm2(h))) ]
-        h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
-
-        if self.in_channels != self.out_channels:
-            x = self.nin_shortcut(x)
-
-        return x+h
-
 class Encoder(nn.Module):
     def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
                  ffactor_spatial, ffactor_temporal, downsample_match_channel=True, refiner_vae=True, **_):
@@ -191,7 +144,7 @@ def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
 
         self.refiner_vae = refiner_vae
         if self.refiner_vae:
-            conv_op = NoPadConv3d
+            conv_op = CarriedConv3d
             norm_op = RMS_norm
         else:
             conv_op = ops.Conv3d
@@ -206,9 +159,10 @@ def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
 
         for i, tgt in enumerate(block_out_channels):
             stage = nn.Module()
-            stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt,
-                                                                   out_channels=tgt,
-                                                                   conv_op=conv_op, norm_op=norm_op)
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=conv_op, norm_op=norm_op)
                                         for j in range(num_res_blocks)])
             ch = tgt
             if i < depth:
@@ -218,9 +172,9 @@ def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
             self.down.append(stage)
 
         self.mid = nn.Module()
-        self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
         self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
-        self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
 
         self.norm_out = norm_op(ch)
         self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1)
@@ -246,22 +200,20 @@ def forward(self, x):
             conv_carry_out = []
             if i == len(x) - 1:
                 conv_carry_out = None
+
             x1 = [ x1 ]
             x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
 
             for stage in self.down:
                 for blk in stage.block:
-                    x1 = blk(x1, conv_carry_in, conv_carry_out)
+                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
                 if hasattr(stage, 'downsample'):
                     x1 = stage.downsample(x1, conv_carry_in, conv_carry_out)
 
             out.append(x1)
             conv_carry_in = conv_carry_out
 
-        if len(out) > 1:
-            out = torch.cat(out, dim=2)
-        else:
-            out = out[0]
+        out = torch_cat_if_needed(out, dim=2)
 
         x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out)))
         del out
@@ -288,7 +240,7 @@ def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks,
 
         self.refiner_vae = refiner_vae
         if self.refiner_vae:
-            conv_op = NoPadConv3d
+            conv_op = CarriedConv3d
             norm_op = RMS_norm
         else:
             conv_op = ops.Conv3d
@@ -298,19 +250,20 @@ def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks,
         self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1)
 
         self.mid = nn.Module()
-        self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
         self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
-        self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
 
         self.up = nn.ModuleList()
         depth = (ffactor_spatial >> 1).bit_length()
         depth_temporal = (ffactor_temporal >> 1).bit_length()
 
         for i, tgt in enumerate(block_out_channels):
             stage = nn.Module()
-            stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt,
-                                                                   out_channels=tgt,
-                                                                   conv_op=conv_op, norm_op=norm_op)
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=conv_op, norm_op=norm_op)
                                         for j in range(num_res_blocks + 1)])
             ch = tgt
             if i < depth:
@@ -340,7 +293,7 @@ def forward(self, z):
                 conv_carry_out = None
             for stage in self.up:
                 for blk in stage.block:
-                    x1 = blk(x1, conv_carry_in, conv_carry_out)
+                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
                 if hasattr(stage, 'upsample'):
                     x1 = stage.upsample(x1, conv_carry_in, conv_carry_out)
 
@@ -350,10 +303,7 @@ def forward(self, z):
             conv_carry_in = conv_carry_out
         del x
 
-        if len(out) > 1:
-            out = torch.cat(out, dim=2)
-        else:
-            out = out[0]
+        out = torch_cat_if_needed(out, dim=2)
 
         if not self.refiner_vae:
             if z.shape[-3] == 1: