speed up and reduce VRAM of QWEN VAE and WAN (less so) (Comfy-Org#12036)

* ops: introduce autopad for conv3d

This works around pytorch missing ability to causal pad as part of the
kernel and avoids massive weight duplications for padding.

* wan-vae: rework causal padding

This currently uses F.pad which takes a full deep copy and is liable to
be the VRAM peak. Instead, kick spatial padding back to the op and
consolidate the temporal padding with the cat for the cache.

* wan-vae: implement zero pad fast path

The WAN VAE is also QWEN where it is used single-image. These
convolutions are however zero padded 3d convolutions, which means the
VAE is actually just 2D down the last element of the conv weight in
the temporal dimension. Fast path this, to avoid adding zeros that
then just evaporate in convoluton math but cost computation.

Author: rattus128

comfy/ldm/wan/vae.py

@@ -5,7 +5,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 from einops import rearrange
-from comfy.ldm.modules.diffusionmodules.model import vae_attention
+from comfy.ldm.modules.diffusionmodules.model import vae_attention, torch_cat_if_needed
 
 import comfy.ops
 ops = comfy.ops.disable_weight_init
@@ -20,22 +20,29 @@ class CausalConv3d(ops.Conv3d):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self._padding = (self.padding[2], self.padding[2], self.padding[1],
-                         self.padding[1], 2 * self.padding[0], 0)
-        self.padding = (0, 0, 0)
+        self._padding = 2 * self.padding[0]
+        self.padding = (0, self.padding[1], self.padding[2])
 
     def forward(self, x, cache_x=None, cache_list=None, cache_idx=None):
         if cache_list is not None:
             cache_x = cache_list[cache_idx]
             cache_list[cache_idx] = None
 
-        padding = list(self._padding)
-        if cache_x is not None and self._padding[4] > 0:
-            cache_x = cache_x.to(x.device)
-            x = torch.cat([cache_x, x], dim=2)
-            padding[4] -= cache_x.shape[2]
+        if cache_x is None and x.shape[2] == 1:
+            #Fast path - the op will pad for use by truncating the weight
+            #and save math on a pile of zeros.
+            return super().forward(x, autopad="causal_zero")
+
+        if self._padding > 0:
+            padding_needed = self._padding
+            if cache_x is not None:
+                cache_x = cache_x.to(x.device)
+                padding_needed = max(0, padding_needed - cache_x.shape[2])
+            padding_shape = list(x.shape)
+            padding_shape[2] = padding_needed
+            padding = torch.zeros(padding_shape, device=x.device, dtype=x.dtype)
+            x = torch_cat_if_needed([padding, cache_x, x], dim=2)
             del cache_x
-        x = F.pad(x, padding)
 
         return super().forward(x)
 

comfy/ops.py

@@ -203,7 +203,9 @@ class Conv3d(torch.nn.Conv3d, CastWeightBiasOp):
         def reset_parameters(self):
             return None
 
-        def _conv_forward(self, input, weight, bias, *args, **kwargs):
+        def _conv_forward(self, input, weight, bias, autopad=None, *args, **kwargs):
+            if autopad == "causal_zero":
+                weight = weight[:, :, -input.shape[2]:, :, :]
             if NVIDIA_MEMORY_CONV_BUG_WORKAROUND and weight.dtype in (torch.float16, torch.bfloat16):
                 out = torch.cudnn_convolution(input, weight, self.padding, self.stride, self.dilation, self.groups, benchmark=False, deterministic=False, allow_tf32=True)
                 if bias is not None:
@@ -212,15 +214,15 @@ def _conv_forward(self, input, weight, bias, *args, **kwargs):
             else:
                 return super()._conv_forward(input, weight, bias, *args, **kwargs)
 
-        def forward_comfy_cast_weights(self, input):
+        def forward_comfy_cast_weights(self, input, autopad=None):
             weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = self._conv_forward(input, weight, bias)
+            x = self._conv_forward(input, weight, bias, autopad=autopad)
             uncast_bias_weight(self, weight, bias, offload_stream)
             return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
-            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0 or "autopad" in kwargs:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)