Fix the int4 table batched embedding benchmark with mixed dim (#609)

Summary:
Pull Request resolved: #609

Fix the bench with "--mixed" dimension as reported by xcliang in https://www.internalfb.com/diff/D28248236 (0fe80ee014b936733278a77d0a24c9fe9a431c31)?dst_version_fbid=834500627143449&transaction_fbid=331571331658562

Differential Revision: D28466825

fbshipit-source-id: ac4725f37d89a3ecd2bcccd564b0424173f81230

fbgemm_gpu/bench/split_table_batched_embeddings_benchmark.py

@@ -791,7 +791,8 @@ def cpu(  # noqa C901
         feature_requires_grad = None
     if mixed:
         Ds = [
-            div_round_up(np.random.randint(low=int(0.5 * D), high=int(1.5 * D)), 4)
+            # int4 table batched emb op can only handle mixed D where D is multiple of 8
+            div_round_up(np.random.randint(low=int(0.5 * D), high=int(1.5 * D)), 8)
             for _ in range(T)
         ]
         D = np.average(Ds)
@@ -905,8 +906,9 @@ def int4_device(  # noqa C901
     else:
         feature_requires_grad = None
     if mixed:
+        # int4 table batched emb op can only handle mixed D where D is multiple of 8
         Ds = [
-            div_round_up(np.random.randint(low=int(0.5 * D), high=int(1.5 * D)), 4)
+            div_round_up(np.random.randint(low=int(0.5 * D), high=int(1.5 * D)), 8)
             for _ in range(T)
         ]
         D = np.average(Ds)