Merge branch 'master' into feat/point-cloud-gaussian-splat-nodes-v2

comfy/background_removal/birefnet.py

@@ -105,7 +105,7 @@ class WindowAttention(nn.Module):
 
         relative_position_bias = self.relative_position_bias_table[self.relative_position_index.long().view(-1)].view(
             self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
-        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        relative_position_bias = comfy.ops.cast_to_input(relative_position_bias.permute(2, 0, 1).contiguous(), attn)  # nH, Wh*Ww, Wh*Ww
         attn = attn + relative_position_bias.unsqueeze(0)
 
         if mask is not None:

comfy/clip_vision.py

@@ -2,7 +2,6 @@ from .utils import load_torch_file, transformers_convert, state_dict_prefix_repl
 import os
 import json
 import logging
-import torch
 
 import comfy.ops
 import comfy.model_patcher
@@ -50,10 +49,6 @@ class ClipVisionModel():
         self.load_device = comfy.model_management.text_encoder_device()
         offload_device = comfy.model_management.text_encoder_offload_device()
         self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
-        if self.model_type == "dinov3" and self.dtype == torch.float16:
-            # DINOv3's activations borderline fits fp16, preferring bf16 if available for better stability #TODO: further fp16 tests in practice
-            if comfy.model_management.should_use_bf16(self.load_device, prioritize_performance=True):
-                self.dtype = torch.bfloat16
         self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
         self.model.eval()
 

comfy/image_encoders/dino3.py

@@ -3,6 +3,7 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+import comfy.ops
 from comfy.ldm.modules.attention import optimized_attention_for_device
 from comfy.image_encoders.dino2 import LayerScale as DINOv3ViTLayerScale
 
@@ -166,17 +167,16 @@ class DINOv3ViTEmbeddings(nn.Module):
 
     def forward(self, pixel_values, bool_masked_pos=None):
         batch_size = pixel_values.shape[0]
-        target_dtype = self.patch_embeddings.weight.dtype
 
-        patch_embeddings = self.patch_embeddings(pixel_values.to(dtype=target_dtype))
+        patch_embeddings = self.patch_embeddings(pixel_values)
         patch_embeddings = patch_embeddings.flatten(2).transpose(1, 2)
 
         if bool_masked_pos is not None:
-            mask_token = self.mask_token.to(patch_embeddings.dtype)
+            mask_token = comfy.ops.cast_to_input(self.mask_token, patch_embeddings)
             patch_embeddings = torch.where(bool_masked_pos.unsqueeze(-1), mask_token, patch_embeddings)
 
-        cls_token = self.cls_token.expand(batch_size, -1, -1).to(patch_embeddings.device)
+        cls_token = comfy.ops.cast_to_input(self.cls_token.expand(batch_size, -1, -1), patch_embeddings)
-        register_tokens = self.register_tokens.expand(batch_size, -1, -1).to(patch_embeddings.device)
+        register_tokens = comfy.ops.cast_to_input(self.register_tokens.expand(batch_size, -1, -1), patch_embeddings)
         embeddings = torch.cat([cls_token, register_tokens, patch_embeddings], dim=1)
         return embeddings
 
@@ -244,7 +244,6 @@ class DINOv3ViTModel(nn.Module):
         return self.embeddings.patch_embeddings
 
     def forward(self, pixel_values, bool_masked_pos=None, **kwargs):
-        pixel_values = pixel_values.to(self.embeddings.patch_embeddings.weight.dtype)
         hidden_states = self.embeddings(pixel_values, bool_masked_pos=bool_masked_pos)
         position_embeddings = self.rope_embeddings(pixel_values)
 

comfy_extras/nodes_math.py

@@ -102,11 +102,18 @@ class MathExpressionNode(io.ComfyNode):
                 f"Math Expression '{expression}' must evaluate to a numeric result, "
                 f"got {type(result).__name__}: {result!r}"
             )
-        if not math.isfinite(result):
+        try:
+            float_result = float(result)
+        except OverflowError:
+            raise ValueError(
+                f"Math Expression '{expression}' produced a result too large to "
+                f"represent as a float: {result}"
+            ) from None
+        if not math.isfinite(float_result):
             raise ValueError(
                 f"Math Expression '{expression}' produced a non-finite result: {result}"
             )
-        return io.NodeOutput(float(result), int(result), bool(result))
+        return io.NodeOutput(float_result, int(result), bool(result))
 
 
 class MathExtension(ComfyExtension):

comfy_extras/nodes_triposplat.py

@@ -115,12 +115,11 @@ class TripoSplatConditioning(IO.ComfyNode):
         # feature1: DINOv3 token sequence (cls + registers + patches), ImageNet-normalized, with a final non-affine layer norm on top
         comfy.model_management.load_model_gpu(clip_vision.patcher)
         device = clip_vision.load_device
-        model_dtype = next(clip_vision.model.parameters()).dtype
         img = image.movedim(-1, 1).to(device)  # (B,3,H,W) in [0,1]
         mean = torch.tensor(_DINOV3_MEAN, device=device).view(1, 3, 1, 1)
         std = torch.tensor(_DINOV3_STD, device=device).view(1, 3, 1, 1)
         img = (img - mean) / std
-        seq = clip_vision.model(pixel_values=img.to(model_dtype))[0]
+        seq = clip_vision.model(pixel_values=img.float())[0]
         feature1 = F.layer_norm(seq.float(), seq.shape[-1:]).to(comfy.model_management.intermediate_device())
 
         # Second conditioning: the Flux2 VAE latent of the image, carried as a standard reference_latents entry
@@ -233,7 +232,9 @@ class TripoSplatSamplingPreview(IO.ComfyNode):
                     return
                 try:
                     if not state["loaded"]:
-                        comfy.model_management.load_models_gpu([vae.patcher], memory_required=memory_required)
+                        loaded_models = comfy.model_management.loaded_models(only_currently_used=True)
+                        loaded_models.append(vae.patcher)
+                        comfy.model_management.load_models_gpu(loaded_models, memory_required=memory_required)
                         state["loaded"] = True
                     img = decode_x0_to_image(vae, x0, cfg)
                     if state["pbar"] is None:

tests-unit/comfy_extras_test/nodes_math_test.py

@@ -197,3 +197,10 @@ class TestMathExpressionExecute:
     def test_pow_huge_exponent_raises(self):
         with pytest.raises(ValueError, match="Exponent .* exceeds maximum"):
             self._exec("pow(a, b)", a=10, b=10000000)
+
+    def test_huge_int_result_raises_value_error(self):
+        # Exponent is within the allowed MAX_EXPONENT range, so the result is a
+        # finite Python int that is nonetheless too large to convert to float.
+        # This must raise a clean ValueError, not an uncaught OverflowError.
+        with pytest.raises(ValueError, match="too large to represent as a float"):
+            self._exec("2 ** 3999")