Fix gen_ai only build (#2603)

Summary:
with-proxy CUDA_HOME=/usr/local/cuda-12.1 python setup.py install -DTORCH_CUDA_ARCH_LIST="8.0;9.0;9.0a" -DFBGEMM_GENAI_ONLY=ON -j 60  --record files.txt 2>&1 | tee out.log

python FBGEMM/fbgemm_gpu/experimental/gen_ai/test/quantize/quantize_test.py


Reviewed By: jiawenliu64

Differential Revision: D57515269

Pulled By: jianyuh

fbgemm_gpu/experimental/gen_ai/test/quantize/quantize_test.py

@@ -259,3 +259,7 @@ def test_quantize_fp8_per_tensor_with_ub(
 
         zq_ref = (x @ w.T).to(torch.bfloat16)
         torch.testing.assert_close(zq, zq_ref, atol=1.0e-3, rtol=1.0e-3)
+
+
+if __name__ == "__main__":
+    unittest.main()

fbgemm_gpu/fbgemm_gpu/__init__.py

@@ -11,8 +11,18 @@
 
 try:
     torch.ops.load_library(os.path.join(os.path.dirname(__file__), "fbgemm_gpu_py.so"))
-except Exception as e:
-    print(e)
+except Exception as error_ranking:
+    try:
+        torch.ops.load_library(
+            os.path.join(
+                os.path.dirname(__file__),
+                "experimental/gen_ai/fbgemm_gpu_experimental_gen_ai_py.so",
+            )
+        )
+    except Exception as error_gen_ai:
+        # When both ranking/gen_ai so files are not available, print the error logs
+        print(error_ranking)
+        print(error_gen_ai)
 
 # Since __init__.py is only used in OSS context, we define `open_source` here
 # and use its existence to determine whether or not we are in OSS context
@@ -24,5 +34,8 @@
 # Export the version string from the version file auto-generated by setup.py
 from fbgemm_gpu.docs.version import __version__  # noqa: F401, E402
 
-# Trigger meta operator registrations
-from . import sparse_ops  # noqa: F401, E402
+try:
+    # Trigger meta operator registrations
+    from . import sparse_ops  # noqa: F401, E402
+except Exception:
+    pass

fbgemm_gpu/fbgemm_gpu/docs/__init__.py

@@ -6,4 +6,7 @@
 # LICENSE file in the root directory of this source tree.
 
 # Trigger the manual addition of docstrings to pybind11-generated operators
-from . import jagged_tensor_ops, table_batched_embedding_ops  # noqa: F401
+try:
+    from . import jagged_tensor_ops, table_batched_embedding_ops  # noqa: F401
+except Exception:
+    pass