Merge fdcc38b9ea into 379fbd1a82

comfy/ldm/lightricks/vae/causal_conv3d.py

@@ -65,13 +65,9 @@ class CausalConv3d(nn.Module):
             self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], 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)
-        elif is_end:
-            self.temporal_cache_state[tid] = (None, True)
 
         return self.conv(x) if x.shape[2] >= self.time_kernel_size else x[:, :, :0, :, :]
 

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -297,23 +297,7 @@ class Encoder(nn.Module):
                     module.temporal_cache_state.pop(tid, None)
 
 
-MIN_VRAM_FOR_CHUNK_SCALING = 6 * 1024 ** 3
-MAX_VRAM_FOR_CHUNK_SCALING = 24 * 1024 ** 3
-MIN_CHUNK_SIZE = 32 * 1024 ** 2
-MAX_CHUNK_SIZE = 128 * 1024 ** 2
-
-def get_max_chunk_size(device: torch.device) -> int:
-    total_memory = comfy.model_management.get_total_memory(dev=device)
-
-    if total_memory <= MIN_VRAM_FOR_CHUNK_SCALING:
-        return MIN_CHUNK_SIZE
-    if total_memory >= MAX_VRAM_FOR_CHUNK_SCALING:
-        return MAX_CHUNK_SIZE
-
-    interp = (total_memory - MIN_VRAM_FOR_CHUNK_SCALING) / (
-        MAX_VRAM_FOR_CHUNK_SCALING - MIN_VRAM_FOR_CHUNK_SCALING
-    )
-    return int(MIN_CHUNK_SIZE + interp * (MAX_CHUNK_SIZE - MIN_CHUNK_SIZE))
+MAX_CHUNK_SIZE=(128 * 1024 ** 2)
 
 class Decoder(nn.Module):
     r"""
@@ -541,11 +525,8 @@ class Decoder(nn.Module):
             timestep_shift_scale = ada_values.unbind(dim=1)
 
         output = []
-        max_chunk_size = get_max_chunk_size(sample.device)
 
-        def run_up(idx, sample_ref, ended):
-            sample = sample_ref[0]
-            sample_ref[0] = None
+        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:
@@ -573,21 +554,13 @@ class Decoder(nn.Module):
                 return
 
             total_bytes = sample.numel() * sample.element_size()
-            num_chunks = (total_bytes + max_chunk_size - 1) // max_chunk_size
+            num_chunks = (total_bytes + MAX_CHUNK_SIZE - 1) // MAX_CHUNK_SIZE
+            samples = torch.chunk(sample, chunks=num_chunks, dim=2)
 
-            if num_chunks == 1:
-                # when we are not chunking, detach our x so the callee can free it as soon as they are done
-                next_sample_ref = [sample]
-                del sample
-                run_up(idx + 1, next_sample_ref, ended)
-                return
-            else:
-                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)
 
-                for chunk_idx, sample1 in enumerate(samples):
-                    run_up(idx + 1, [sample1], ended and chunk_idx == len(samples) - 1)
-
-        run_up(0, [sample], True)
+        run_up(0, sample, True)
         sample = torch.cat(output, dim=2)
 
         sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)

nodes.py

@@ -952,7 +952,7 @@ class UNETLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "unet_name": (folder_paths.get_filename_list("diffusion_models"), ),
-                              "weight_dtype": (["default", "fp8_e4m3fn", "fp8_e4m3fn_fast", "fp8_e5m2"], {"advanced": True})
+                              "weight_dtype": (["default", "fp8_e4m3fn", "fp8_e4m3fn_fast", "fp8_e5m2"],)
                              }}
     RETURN_TYPES = ("MODEL",)
     FUNCTION = "load_unet"