add OUTPUT_NODE = True to ImageCropV2

Allow non QuantizedTensor layer to set want_requant to get the post lora
calculation stochastic cast down to the original input dtype.

This is then used by the legacy fp8 Linear implementation to set the
compute_dtype to the preferred lora dtype but then want_requant it back
down to fp8.

This fixes the issue with --fast fp8_matrix_mult is combined with
--fast dynamic_vram which doing a lora on an fp8_ non QT model.

comfy/ldm/wan/vae.py

@@ -485,7 +485,7 @@ class WanVAE(nn.Module):
         iter_ = 1 + (t - 1) // 4
         feat_map = None
         if iter_ > 1:
-            feat_map = [None] * count_conv3d(self.decoder)
+            feat_map = [None] * count_conv3d(self.encoder)
         ## 对encode输入的x，按时间拆分为1、4、4、4....
         for i in range(iter_):
             conv_idx = [0]

comfy/ops.py

@@ -167,17 +167,15 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             x = to_dequant(x, dtype)
         if not resident and lowvram_fn is not None:
             x = to_dequant(x, dtype if compute_dtype is None else compute_dtype)
-            #FIXME: this is not accurate, we need to be sensitive to the compute dtype
             x = lowvram_fn(x)
-            if (isinstance(orig, QuantizedTensor) and
-                (want_requant and len(fns) == 0 or update_weight)):
+            if (want_requant and len(fns) == 0 or update_weight):
                 seed = comfy.utils.string_to_seed(s.seed_key)
-                y = QuantizedTensor.from_float(x, s.layout_type, scale="recalculate", stochastic_rounding=seed)
-                if want_requant and len(fns) == 0:
-                    #The layer actually wants our freshly saved QT
-                    x = y
-            elif update_weight:
-                y = comfy.float.stochastic_rounding(x, orig.dtype, seed = comfy.utils.string_to_seed(s.seed_key))
+                if isinstance(orig, QuantizedTensor):
+                    y = QuantizedTensor.from_float(x, s.layout_type, scale="recalculate", stochastic_rounding=seed)
+                else:
+                    y = comfy.float.stochastic_rounding(x, orig.dtype, seed=seed)
+            if want_requant and len(fns) == 0:
+                x = y
             if update_weight:
                 orig.copy_(y)
         for f in fns:
@@ -617,7 +615,8 @@ def fp8_linear(self, input):
 
     if input.ndim != 2:
         return None
-    w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
+    lora_compute_dtype=comfy.model_management.lora_compute_dtype(input.device)
+    w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True, compute_dtype=lora_compute_dtype, want_requant=True)
     scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
 
     scale_input = torch.ones((), device=input.device, dtype=torch.float32)

comfy_extras/nodes_images.py

@@ -65,6 +65,8 @@ class ImageCropV2(IO.ComfyNode):
             outputs=[IO.Image.Output()],
         )
 
+    OUTPUT_NODE = True
+
     @classmethod
     def execute(cls, image, crop_region) -> IO.NodeOutput:
         x = crop_region.get("x", 0)
@@ -706,8 +708,8 @@ class SplitImageToTileList(IO.ComfyNode):
     @staticmethod
     def get_grid_coords(width, height, tile_width, tile_height, overlap):
         coords = []
-        stride_x = max(1, tile_width - overlap)
-        stride_y = max(1, tile_height - overlap)
+        stride_x = round(max(tile_width * 0.25, tile_width - overlap))
+        stride_y = round(max(tile_width * 0.25, tile_height - overlap))
 
         y = 0
         while y < height:
@@ -764,34 +766,6 @@ class ImageMergeTileList(IO.ComfyNode):
             ],
         )
 
-    @staticmethod
-    def get_grid_coords(width, height, tile_width, tile_height, overlap):
-        coords = []
-        stride_x = max(1, tile_width - overlap)
-        stride_y = max(1, tile_height - overlap)
-
-        y = 0
-        while y < height:
-            x = 0
-            y_end = min(y + tile_height, height)
-            y_start = max(0, y_end - tile_height)
-
-            while x < width:
-                x_end = min(x + tile_width, width)
-                x_start = max(0, x_end - tile_width)
-
-                coords.append((x_start, y_start, x_end, y_end))
-
-                if x_end >= width:
-                    break
-                x += stride_x
-
-            if y_end >= height:
-                break
-            y += stride_y
-
-        return coords
-
     @classmethod
     def execute(cls, image_list, final_width, final_height, overlap):
         w = final_width[0]
@@ -804,7 +778,7 @@ class ImageMergeTileList(IO.ComfyNode):
         device = first_tile.device
         dtype = first_tile.dtype
 
-        coords = cls.get_grid_coords(w, h, t_w, t_h, ovlp)
+        coords = SplitImageToTileList.get_grid_coords(w, h, t_w, t_h, ovlp)
 
         canvas = torch.zeros((b, h, w, c), device=device, dtype=dtype)
         weights = torch.zeros((b, h, w, 1), device=device, dtype=dtype)