Fix Linux release build exceeding 4GB PyInstaller limit

Install CPU-only PyTorch on Linux and exclude nvidia modules from
non-CUDA PyInstaller builds.

.github/workflows/release.yml

@@ -61,17 +61,16 @@ jobs:
           pip install pyinstaller
           pip install -r backend/requirements.txt
 
+      - name: Install CPU-only PyTorch (Linux)
+        if: matrix.platform == 'ubuntu-22.04'
+        run: |
+          pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu --force-reinstall --no-deps
+
       - name: Install MLX dependencies (Apple Silicon only)
         if: matrix.backend == 'mlx'
         run: |
           pip install -r backend/requirements-mlx.txt
 
-      # - name: Install PyTorch with CUDA (Windows only)
-      #   if: matrix.platform == 'windows-latest'
-      #   run: |
-      #     pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall --no-deps
-      #     pip install torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
-
       - name: Build Python server (Linux/macOS)
         if: matrix.platform != 'windows-latest'
         run: |

backend/build_binary.py

@@ -90,6 +90,18 @@ def build_server(cuda=False):
             '--hidden-import', 'torch.cuda',
             '--hidden-import', 'torch.backends.cudnn',
         ])
+    else:
+        # Exclude NVIDIA CUDA packages from CPU-only builds to keep binary under 4GB.
+        # On Linux, pip may pull CUDA-enabled PyTorch by default which includes ~3GB
+        # of NVIDIA shared libraries that PyInstaller would bundle.
+        nvidia_packages = [
+            'nvidia', 'nvidia.cublas', 'nvidia.cuda_cupti', 'nvidia.cuda_nvrtc',
+            'nvidia.cuda_runtime', 'nvidia.cudnn', 'nvidia.cufft', 'nvidia.curand',
+            'nvidia.cusolver', 'nvidia.cusparse', 'nvidia.nccl', 'nvidia.nvjitlink',
+            'nvidia.nvtx',
+        ]
+        for pkg in nvidia_packages:
+            args.extend(['--exclude-module', pkg])
 
     # Add MLX-specific imports if building on Apple Silicon (never for CUDA builds)
     if is_apple_silicon() and not cuda: