chore(deps): update torch/vision to cu128 for Blackwell

- Pin torch==2.10.0 and torchvision==0.25.0 to official CUDA 12.8 wheels
- Set --extra-index-url for cu128; remove old find-links/comments
- Ensures compatibility with Blackwell (sm_120) and RTX PRO 6000

pre-requirements.txt

@@ -1,15 +1,11 @@
-pip>=26.1.1
+pip>=26.2
 
-# Use the official PyTorch wheels matching CUDA 12.8 (cu128).
-# This ensures torchvision is available for the fast image processor.
-# The -f find-links page provides the compatible wheel builds.
--f https://download.pytorch.org/whl/cu128/torch_stable.html
---extra-index-url https://download.pytorch.org/whl/cu128/torch_stable.html
--f https://download.pytorch.org/whl/cu128/torch_stable.html
-# Pin normalized versions to avoid conflicts with external install-time constraints
-# The cu128 wheels will be resolved from the extra index when available for the platform.
-torch>=2.4.0
-torchvision>=0.19.0
+# Official PyTorch + torchvision wheels built for CUDA 12.8 (cu128).
+# These include sm_120 (Blackwell) kernels required by ZeroGPU’s RTX PRO 6000.
+--extra-index-url https://download.pytorch.org/whl/cu128
+
+torch==2.10.0
+torchvision==0.25.0
 
 # Previously used examples (kept for reference):
 # torch==2.4.0 --extra-index-url https://download.pytorch.org/whl/cu124