[pull] master from comfyanonymous:master

comfy/ldm/chroma_radiance/model.py

@@ -189,15 +189,15 @@ def forward_nerf(
         nerf_pixels = nn.functional.unfold(img_orig, kernel_size=patch_size, stride=patch_size)
         nerf_pixels = nerf_pixels.transpose(1, 2) # -> [B, NumPatches, C * P * P]
 
+        # Reshape for per-patch processing
+        nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size)
+        nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2)
+
         if params.nerf_tile_size > 0 and num_patches > params.nerf_tile_size:
             # Enable tiling if nerf_tile_size isn't 0 and we actually have more patches than
             # the tile size.
-            img_dct = self.forward_tiled_nerf(img_out, nerf_pixels, B, C, num_patches, patch_size, params)
+            img_dct = self.forward_tiled_nerf(nerf_hidden, nerf_pixels, B, C, num_patches, patch_size, params)
         else:
-            # Reshape for per-patch processing
-            nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size)
-            nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2)
-
             # Get DCT-encoded pixel embeddings [pixel-dct]
             img_dct = self.nerf_image_embedder(nerf_pixels)
 
@@ -240,17 +240,8 @@ def forward_tiled_nerf(
             end = min(i + tile_size, num_patches)
 
             # Slice the current tile from the input tensors
-            nerf_hidden_tile = nerf_hidden[:, i:end, :]
-            nerf_pixels_tile = nerf_pixels[:, i:end, :]
-
-            # Get the actual number of patches in this tile (can be smaller for the last tile)
-            num_patches_tile = nerf_hidden_tile.shape[1]
-
-            # Reshape the tile for per-patch processing
-            # [B, NumPatches_tile, D] -> [B * NumPatches_tile, D]
-            nerf_hidden_tile = nerf_hidden_tile.reshape(batch * num_patches_tile, params.hidden_size)
-            # [B, NumPatches_tile, C*P*P] -> [B*NumPatches_tile, C, P*P] -> [B*NumPatches_tile, P*P, C]
-            nerf_pixels_tile = nerf_pixels_tile.reshape(batch * num_patches_tile, channels, patch_size**2).transpose(1, 2)
+            nerf_hidden_tile = nerf_hidden[i * batch:end * batch]
+            nerf_pixels_tile = nerf_pixels[i * batch:end * batch]
 
             # get DCT-encoded pixel embeddings [pixel-dct]
             img_dct_tile = self.nerf_image_embedder(nerf_pixels_tile)

comfy/model_detection.py

@@ -213,7 +213,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["nerf_mlp_ratio"] = 4
                 dit_config["nerf_depth"] = 4
                 dit_config["nerf_max_freqs"] = 8
-                dit_config["nerf_tile_size"] = 32
+                dit_config["nerf_tile_size"] = 512
                 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear"
                 dit_config["nerf_embedder_dtype"] = torch.float32
         else:

comfy/model_management.py

@@ -371,6 +371,9 @@ def amd_min_version(device=None, min_rdna_version=0):
 except:
     pass
 
+if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast:
+    torch.backends.cudnn.benchmark = True
+
 try:
     if torch_version_numeric >= (2, 5):
         torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)

comfy/ops.py

@@ -67,9 +67,6 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs):
 
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
-if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast:
-    torch.backends.cudnn.benchmark = True
-
 def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 

cuda_malloc.py

@@ -1,6 +1,6 @@
 import os
 import importlib.util
-from comfy.cli_args import args
+from comfy.cli_args import args, PerformanceFeature
 import subprocess
 
 #Can't use pytorch to get the GPU names because the cuda malloc has to be set before the first import.
@@ -75,8 +75,9 @@ def cuda_malloc_supported():
                 spec.loader.exec_module(module)
                 version = module.__version__
 
-        if int(version[0]) >= 2 and "+cu" in version: #enable by default for torch version 2.0 and up only on cuda torch
-            args.cuda_malloc = cuda_malloc_supported()
+        if int(version[0]) >= 2 and "+cu" in version:  # enable by default for torch version 2.0 and up only on cuda torch
+            if PerformanceFeature.AutoTune not in args.fast:  # Autotune has issues with cuda malloc
+                args.cuda_malloc = cuda_malloc_supported()
     except:
         pass