Fix Bfloat16 support in iree-turbine

Author: DavidGinten

.github/workflows/ci.yaml

@@ -62,7 +62,7 @@ jobs:
       - name: Run unit tests
         if: ${{ !cancelled() }}
         run: |
-          pytest -n 4 --capture=tee-sys -vv .
+          pytest --capture=tee-sys -vv .
 
       - name: Run LIT tests
         if: ${{ !cancelled() }}

iree/turbine/aot/support/ir_utils.py

@@ -324,10 +324,7 @@ def create_tensor_global(
                 )
             else:
                 # Emit inline initialized.
-                detached_tensor = t.detach().contiguous().cpu()
-                array = np.array(detached_tensor)
-                # We know that a Numpy array is a ReadableBuffer so ignore type error.
-                contents = memoryview(array)  # type: ignore
+                contents = torch.to_dlpack(t)
                 blob_name = symbol_name
                 elements_attr = DenseResourceElementsAttr.get_from_buffer(
                     contents, blob_name, tensor_type