Further Reduce LTX VAE decode peak RAM usage (#13052)

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -473,6 +473,17 @@ class Decoder(nn.Module):
 
         self.gradient_checkpointing = False
 
+        # Precompute output scale factors: (channels, (t_scale, h_scale, w_scale), t_offset)
+        ts, hs, ws, to = 1, 1, 1, 0
+        for block in self.up_blocks:
+            if isinstance(block, DepthToSpaceUpsample):
+                ts *= block.stride[0]
+                hs *= block.stride[1]
+                ws *= block.stride[2]
+                if block.stride[0] > 1:
+                    to = to * block.stride[0] + 1
+        self._output_scale = (out_channels // (patch_size ** 2), (ts, hs * patch_size, ws * patch_size), to)
+
         self.timestep_conditioning = timestep_conditioning
 
         if timestep_conditioning:
@@ -494,11 +505,15 @@ class Decoder(nn.Module):
             )
 
 
-    # def forward(self, sample: torch.FloatTensor, target_shape) -> torch.FloatTensor:
+    def decode_output_shape(self, input_shape):
+        c, (ts, hs, ws), to = self._output_scale
+        return (input_shape[0], c, input_shape[2] * ts - to, input_shape[3] * hs, input_shape[4] * ws)
+
     def forward_orig(
         self,
         sample: torch.FloatTensor,
         timestep: Optional[torch.Tensor] = None,
+        output_buffer: Optional[torch.Tensor] = None,
     ) -> torch.FloatTensor:
         r"""The forward method of the `Decoder` class."""
         batch_size = sample.shape[0]
@@ -540,7 +555,13 @@ class Decoder(nn.Module):
             )
             timestep_shift_scale = ada_values.unbind(dim=1)
 
-        output = []
+        if output_buffer is None:
+            output_buffer = torch.empty(
+                self.decode_output_shape(sample.shape),
+                dtype=sample.dtype, device=comfy.model_management.intermediate_device(),
+            )
+        output_offset = [0]
+
         max_chunk_size = get_max_chunk_size(sample.device)
 
         def run_up(idx, sample_ref, ended):
@@ -556,7 +577,10 @@ class Decoder(nn.Module):
                     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.to(comfy.model_management.intermediate_device()))
+                    sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
+                    t = sample.shape[2]
+                    output_buffer[:, :, output_offset[0]:output_offset[0] + t].copy_(sample)
+                    output_offset[0] += t
                 return
 
             up_block = self.up_blocks[idx]
@@ -588,11 +612,8 @@ class Decoder(nn.Module):
                     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
+        return output_buffer
 
     def forward(self, *args, **kwargs):
         try:
@@ -1226,7 +1247,10 @@ class VideoVAE(nn.Module):
         means, logvar = torch.chunk(self.encoder(x), 2, dim=1)
         return self.per_channel_statistics.normalize(means)
 
-    def decode(self, x):
+    def decode_output_shape(self, input_shape):
+        return self.decoder.decode_output_shape(input_shape)
+
+    def decode(self, x, output_buffer=None):
         if self.timestep_conditioning: #TODO: seed
             x = torch.randn_like(x) * self.decode_noise_scale + (1.0 - self.decode_noise_scale) * x
-        return self.decoder(self.per_channel_statistics.un_normalize(x), timestep=self.decode_timestep)
+        return self.decoder(self.per_channel_statistics.un_normalize(x), timestep=self.decode_timestep, output_buffer=output_buffer)

comfy/sd.py

@@ -951,12 +951,23 @@ class VAE:
             batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
 
+            # Pre-allocate output for VAEs that support direct buffer writes
+            preallocated = False
+            if hasattr(self.first_stage_model, 'decode_output_shape'):
+                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
+
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x + batch_number].to(device=self.device, dtype=self.vae_dtype)
-                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True))
-                if pixel_samples is None:
-                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
-                pixel_samples[x:x+batch_number] = out
+                if preallocated:
+                    self.first_stage_model.decode(samples, output_buffer=pixel_samples[x:x+batch_number], **vae_options)
+                else:
+                    out = self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True)
+                    if pixel_samples is None:
+                        pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
+                    pixel_samples[x:x+batch_number].copy_(out)
+                    del out
+                self.process_output(pixel_samples[x:x+batch_number])
         except Exception as e:
             model_management.raise_non_oom(e)
             logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")