Merge pull request #276 from AInVFX/main

fix: enhance Conv3d workaround compatibility for PyTorch dev builds and AMD ROCm
Fixes #178 and #265

Author: adrientoupet

README.md

@@ -36,6 +36,10 @@ We're actively working on improvements and new features. To stay informed:
 
 ## 🚀 Updates
 
+**2025.11.10 - Version 2.5.7**
+
+- **🔧 Fix: Conv3d workaround compatibility** - Enhanced platform detection and added graceful fallback to prevent errors on PyTorch dev builds and AMD ROCm systems
+
 **2025.11.09 - Version 2.5.6**
 
 - 🎨 **Fix: Restored natural look for 7b model** - Corrected torch.compile optimization that was causing overly plastic/ high-specular appearance in upscaled videos with 7b model.

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "seedvr2_videoupscaler"
 description = "SeedVR2 official ComfyUI integration: ByteDance-Seed's one-step diffusion-based video/image upscaling with memory-efficient inference"
-version = "2.5.6"
+version = "2.5.7"
 authors = [
     {name = "numz"},
     {name = "adrientoupet"}

src/models/video_vae_v3/modules/causal_inflation_lib.py

@@ -96,17 +96,21 @@ def _conv_forward(self, input, weight, bias, *args, **kwargs):
             hasattr(torch.backends.cudnn, 'is_available') and
             torch.backends.cudnn.is_available() and
             getattr(torch.backends.cudnn, 'enabled', True)):
-            # Direct cuDNN call bypasses buggy PyTorch dispatch layer (NVIDIA only)
-            out = torch.cudnn_convolution(
-                input, weight, self.padding, self.stride, self.dilation, self.groups,
-                benchmark=False, deterministic=False, allow_tf32=True
-            )
-            if bias is not None:
-                out += bias.reshape((1, -1) + (1,) * (out.ndim - 2))
-            return out
-        else:
-            # Use standard path for unaffected configurations
-            return super()._conv_forward(input, weight, bias, *args, **kwargs)
+            try:
+                # Direct cuDNN call bypasses buggy PyTorch dispatch layer (NVIDIA only)
+                out = torch.cudnn_convolution(
+                    input, weight, self.padding, self.stride, self.dilation, self.groups,
+                    benchmark=False, deterministic=False, allow_tf32=True
+                )
+                if bias is not None:
+                    out += bias.reshape((1, -1) + (1,) * (out.ndim - 2))
+                return out
+            except RuntimeError:
+                # Fallback if direct cuDNN call fails (dev builds, edge cases)
+                pass
+        
+        # Use standard path for unaffected configurations or if workaround failed
+        return super()._conv_forward(input, weight, bias, *args, **kwargs)
 
     def memory_limit_conv(
         self,

src/optimization/compatibility.py

@@ -105,7 +105,21 @@ def _check_conv3d_memory_bug():
     with fp16/bfloat16 due to buggy dispatch layer.
     """
     try:
-        if not (torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 3):
+        # Exclude AMD ROCm/HIP builds (they use MIOpen, not cuDNN)
+        if hasattr(torch.version, 'hip') and torch.version.hip is not None:
+            return False
+        
+        # Must have CUDA available
+        if not (hasattr(torch, 'cuda') and torch.cuda.is_available()):
+            return False
+        
+        # Must have cuDNN actually available (not just the attribute)
+        if not (hasattr(torch.backends.cudnn, 'is_available') and 
+                torch.backends.cudnn.is_available()):
+            return False
+        
+        # Check device capability (NVIDIA GPUs)
+        if torch.cuda.get_device_capability()[0] < 3:
             return False
 
         # Parse torch version