[AOTI][CPU] Consider bias=None case for fbgemm_linear_fp16_weight (#158535)

Test Plan:

Rollback Plan:

Differential Revision: D78458214

Pull Request resolved: https://github.com/pytorch/pytorch/pull/158535
Approved by: https://github.com/houseroad, https://github.com/henryoier, https://github.com/jingsh

test/quantization/core/test_quantized_op.py

@@ -3550,14 +3550,15 @@ class TestDynamicQuantizedOps(TestCase):
             (2, 4),         # batch_size
             (4, 5),     # input_channels
             (4, 7),      # output_channels
+            (True, False),         # bias None or not
         )
-        for batch_size, input_channels, output_channels in options:
+        for batch_size, input_channels, output_channels, bias_is_none in options:
             pack_op = torch.ops._quantized.wrapped_fbgemm_pack_gemm_matrix_fp16
             linear_op = torch.ops._quantized.wrapped_fbgemm_linear_fp16_weight
 
             x = torch.randn(batch_size, input_channels)
             w = torch.randn(output_channels, input_channels)
-            bias = torch.randn(output_channels)
+            bias = torch.randn(output_channels) if not bias_is_none else None
 
             w_packed = pack_op(w)
             out = linear_op(x, w_packed, bias, output_channels)
@@ -3591,6 +3592,18 @@ class TestDynamicQuantizedOps(TestCase):
 
         self.assertEqual(ref_out, compiled_out)
 
+        def func(X, W):
+            packed_W = torch.ops._quantized.wrapped_fbgemm_pack_gemm_matrix_fp16(W)
+            return torch.ops._quantized.wrapped_fbgemm_linear_fp16_weight(X, packed_W, None, W.size(0))
+
+        ref_out = func(x, w)
+
+        compiled = torch.compile(func)
+        compiled_out = compiled(x, w)
+
+        self.assertEqual(ref_out, compiled_out)
+
+
     """Tests the correctness of the dynamic quantized lstm/gru."""
 
     def _get_rnn_inputs(self, seq_len, num_batches, input_size, hidden_size, num_directions, reduce_range):