Merge branch 'main' into 32bit_optimizer

Author: YangKai0616

.github/FUNDING.yml

@@ -0,0 +1 @@
+open_collective: bitsandbytes

.github/scripts/build-cuda.sh

@@ -15,10 +15,10 @@ elif [ "${build_arch}" = "aarch64" ]; then
     [[ "${cuda_version}" == 12.8.* || "${cuda_version}" == 12.9.* ]] && build_capability="75;80;90;100;120"
 else
     # By default, target Maxwell through Hopper.
-    build_capability="50;52;60;61;70;75;80;86;89;90"
+    build_capability="50;60;70;75;80;86;89;90"
 
-    # CUDA 12.8+: Add sm100 and sm120; remove < sm75 to align with PyTorch 2.7+cu128 minimum
-    [[ "${cuda_version}" == 12.8.* || "${cuda_version}" == 12.9.* ]] && build_capability="75;80;86;89;90;100;120"
+    # CUDA 12.8+: Add sm100 and sm120; remove < sm70 to align with PyTorch 2.8+cu128 minimum
+    [[ "${cuda_version}" == 12.8.* || "${cuda_version}" == 12.9.* ]] && build_capability="70;75;80;86;89;90;100;120"
 fi
 
 [[ "${build_os}" = windows-* ]] && python3 -m pip install ninja