ltx: vae: implement chunked encoder + CPU IO chunking (Big VRAM reductions) (#13062)

* ltx: vae: add cache state to downsample block

* ltx: vae: Add time stride awareness to causal_conv_3d

* ltx: vae: Automate truncation for encoder

Other VAEs just truncate without error. Do the same.

* sd/ltx: Make chunked_io a flag in its own right

Taking this bi-direcitonal, so make it a for-purpose named flag.

* ltx: vae: implement chunked encoder + CPU IO chunking

People are doing things with big frame counts in LTX including V2V
flows. Implement the time-chunked encoder to keep the VRAM down, with
the converse of the new CPU pre-allocation technique, where the chunks
are brought from the CPU JIT.

* ltx: vae-encode: round chunk sizes more strictly

Only powers of 2 and multiple of 8 are valid due to cache slicing.

comfy/ldm/lightricks/vae/causal_conv3d.py

@@ -23,6 +23,11 @@ class CausalConv3d(nn.Module):
         self.in_channels = in_channels
         self.out_channels = out_channels
 
+        if isinstance(stride, int):
+            self.time_stride = stride
+        else:
+            self.time_stride = stride[0]
+
         kernel_size = (kernel_size, kernel_size, kernel_size)
         self.time_kernel_size = kernel_size[0]
 
@@ -58,18 +63,23 @@ class CausalConv3d(nn.Module):
         pieces = [ cached, x ]
         if is_end and not causal:
             pieces.append(x[:, :, -1:, :, :].repeat((1, 1, (self.time_kernel_size - 1) // 2, 1, 1)))
+        input_length = sum([piece.shape[2] for piece in pieces])
+        cache_length = (self.time_kernel_size - self.time_stride) + ((input_length - self.time_kernel_size) % self.time_stride)
 
         needs_caching = not is_end
-        if needs_caching and x.shape[2] >= self.time_kernel_size - 1:
+        if needs_caching and cache_length == 0:
+            self.temporal_cache_state[tid] = (x[:, :, :0, :, :], False)
             needs_caching = False
-            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+        if needs_caching and x.shape[2] >= cache_length:
+            needs_caching = False
+            self.temporal_cache_state[tid] = (x[:, :, -cache_length:, :, :], False)
 
         x = torch.cat(pieces, dim=2)
         del pieces
         del cached
 
         if needs_caching:
-            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+            self.temporal_cache_state[tid] = (x[:, :, -cache_length:, :, :], False)
         elif is_end:
             self.temporal_cache_state[tid] = (None, True)
 

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -233,10 +233,7 @@ class Encoder(nn.Module):
 
         self.gradient_checkpointing = False
 
-    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)
+    def _forward_chunk(self, sample: torch.FloatTensor) -> Optional[torch.FloatTensor]:
         sample = self.conv_in(sample)
 
         checkpoint_fn = (
@@ -247,10 +244,14 @@ class Encoder(nn.Module):
 
         for down_block in self.down_blocks:
             sample = checkpoint_fn(down_block)(sample)
+            if sample is None or sample.shape[2] == 0:
+                return None
 
         sample = self.conv_norm_out(sample)
         sample = self.conv_act(sample)
         sample = self.conv_out(sample)
+        if sample is None or sample.shape[2] == 0:
+            return None
 
         if self.latent_log_var == "uniform":
             last_channel = sample[:, -1:, ...]
@@ -282,9 +283,35 @@ class Encoder(nn.Module):
 
         return sample
 
+    def forward_orig(self, sample: torch.FloatTensor, device=None) -> torch.FloatTensor:
+        r"""The forward method of the `Encoder` class."""
+
+        max_chunk_size = get_max_chunk_size(sample.device if device is None else device) * 2  # encoder is more memory-efficient than decoder
+        frame_size = sample[:, :, :1, :, :].numel() * sample.element_size()
+        frame_size = int(frame_size * (self.conv_in.out_channels / self.conv_in.in_channels))
+
+        outputs = []
+        samples = [sample[:, :, :1, :, :]]
+        if sample.shape[2] > 1:
+            chunk_t = max(2, max_chunk_size // frame_size)
+            if chunk_t < 4:
+                chunk_t = 2
+            elif chunk_t < 8:
+                chunk_t = 4
+            else:
+                chunk_t = (chunk_t // 8) * 8
+            samples += list(torch.split(sample[:, :, 1:, :, :], chunk_t, dim=2))
+        for chunk_idx, chunk in enumerate(samples):
+            if chunk_idx == len(samples) - 1:
+                mark_conv3d_ended(self)
+            chunk = patchify(chunk, patch_size_hw=self.patch_size, patch_size_t=1).to(device=device)
+            output = self._forward_chunk(chunk)
+            if output is not None:
+                outputs.append(output)
+
+        return torch_cat_if_needed(outputs, dim=2)
+
     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:
@@ -737,12 +764,25 @@ class SpaceToDepthDownsample(nn.Module):
             causal=True,
             spatial_padding_mode=spatial_padding_mode,
         )
+        self.temporal_cache_state = {}
 
     def forward(self, x, causal: bool = True):
-        if self.stride[0] == 2:
+        tid = threading.get_ident()
+        cached, pad_first, cached_x, cached_input = self.temporal_cache_state.get(tid, (None, True, None, None))
+        if cached_input is not None:
+            x = torch_cat_if_needed([cached_input, x], dim=2)
+            cached_input = None
+
+        if self.stride[0] == 2 and pad_first:
             x = torch.cat(
                 [x[:, :, :1, :, :], x], dim=2
             )  # duplicate first frames for padding
+            pad_first = False
+
+        if x.shape[2] < self.stride[0]:
+            cached_input = x
+            self.temporal_cache_state[tid] = (cached, pad_first, cached_x, cached_input)
+            return None
 
         # skip connection
         x_in = rearrange(
@@ -757,15 +797,26 @@ class SpaceToDepthDownsample(nn.Module):
 
         # conv
         x = self.conv(x, causal=causal)
-        x = rearrange(
-            x,
-            "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
-            p1=self.stride[0],
-            p2=self.stride[1],
-            p3=self.stride[2],
-        )
+        if self.stride[0] == 2 and x.shape[2] == 1:
+            if cached_x is not None:
+                x = torch_cat_if_needed([cached_x, x], dim=2)
+                cached_x = None
+            else:
+                cached_x = x
+                x = None
 
-        x = x + x_in
+        if x is not None:
+            x = rearrange(
+                x,
+                "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
+                p1=self.stride[0],
+                p2=self.stride[1],
+                p3=self.stride[2],
+            )
+
+        cached = add_exchange_cache(x, cached, x_in, dim=2)
+
+        self.temporal_cache_state[tid] = (cached, pad_first, cached_x, cached_input)
 
         return x
 
@@ -1098,6 +1149,8 @@ class processor(nn.Module):
         return (x - self.get_buffer("mean-of-means").view(1, -1, 1, 1, 1).to(x)) / self.get_buffer("std-of-means").view(1, -1, 1, 1, 1).to(x)
 
 class VideoVAE(nn.Module):
+    comfy_has_chunked_io = True
+
     def __init__(self, version=0, config=None):
         super().__init__()
 
@@ -1240,11 +1293,9 @@ class VideoVAE(nn.Module):
             }
         return config
 
-    def encode(self, x):
-        frames_count = x.shape[2]
-        if ((frames_count - 1) % 8) != 0:
-            raise ValueError("Invalid number of frames: Encode input must have 1 + 8 * x frames (e.g., 1, 9, 17, ...). Please check your input.")
-        means, logvar = torch.chunk(self.encoder(x), 2, dim=1)
+    def encode(self, x, device=None):
+        x = x[:, :, :max(1, 1 + ((x.shape[2] - 1) // 8) * 8), :, :]
+        means, logvar = torch.chunk(self.encoder(x, device=device), 2, dim=1)
         return self.per_channel_statistics.normalize(means)
 
     def decode_output_shape(self, input_shape):

comfy/sd.py

@@ -953,7 +953,7 @@ class VAE:
 
             # Pre-allocate output for VAEs that support direct buffer writes
             preallocated = False
-            if hasattr(self.first_stage_model, 'decode_output_shape'):
+            if getattr(self.first_stage_model, 'comfy_has_chunked_io', False):
                 pixel_samples = torch.empty(self.first_stage_model.decode_output_shape(samples_in.shape), device=self.output_device, dtype=self.vae_output_dtype())
                 preallocated = True
 
@@ -1038,8 +1038,13 @@ class VAE:
             batch_number = max(1, batch_number)
             samples = None
             for x in range(0, pixel_samples.shape[0], batch_number):
-                pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype).to(self.device)
-                out = self.first_stage_model.encode(pixels_in).to(self.output_device).to(dtype=self.vae_output_dtype())
+                pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype)
+                if getattr(self.first_stage_model, 'comfy_has_chunked_io', False):
+                    out = self.first_stage_model.encode(pixels_in, device=self.device)
+                else:
+                    pixels_in = pixels_in.to(self.device)
+                    out = self.first_stage_model.encode(pixels_in)
+                out = out.to(self.output_device).to(dtype=self.vae_output_dtype())
                 if samples is None:
                     samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 samples[x:x + batch_number] = out