Update for autoloading

Author: matthewdouglas

setup.py

@@ -6,5 +6,5 @@
     package_dir={"": "src"},
     packages=find_packages(where="src"),
     install_requires=["bitsandbytes"],
-    entry_points={"torch.backends": ["bitsandbytes_intel = bitsandbytes_intel:_autoload"]},
+    entry_points={"bitsandbytes.backends": ["bitsandbytes_intel = bitsandbytes_intel:_autoload"]},
 )

src/bitsandbytes_intel/__main__.py

@@ -8,7 +8,7 @@ def main():
 
     try:
         A = torch.randint(-128, 127, (32, 64), dtype=torch.int8).to("xpu")
-        B = torch.randint(-128, 127, (64, 128), dtype=torch.int8).to("xpu")
+        B = torch.randint(-128, 127, (128, 64), dtype=torch.int8).to("xpu")
 
         result = torch.ops.bitsandbytes.int8_linear_matmul(A, B)
         # Simple output verification

src/bitsandbytes_intel/cpu_xpu_common.py

@@ -203,8 +203,9 @@ def int8_linear_matmul_impl(
     A_reshaped = A.reshape(m, k)
 
     # torch._int_mm is available on CPU since torch 2.4, XPU since torch 2.6
-    if (A.device.type == "cpu" and _torch_version_prereq(2, 4)) or (
-        A.device.type == "xpu" and _torch_version_prereq(2, 6)
+    if (
+        A.device.type == "cpu" and _torch_version_prereq(2, 4)
+        # or (A.device.type == "xpu" and _torch_version_prereq(2, 6)
     ):
         C = torch._int_mm(A_reshaped, B.T).to(dtype)
     else: