added check to avoid div 0 errors in cache report; added in place wei…

…ght initial methods

Summary: as in title

Differential Revision: D44096435

fbshipit-source-id: bc71dbfd9382c17f1abd09540217b807a6887902

fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops.py

@@ -979,10 +979,11 @@ def print_uvm_cache_stats(self) -> None:
             f"N_conflict_unique_misses: {uvm_cache_stats[4]}\n"
             f"N_conflict_misses: {uvm_cache_stats[5]}\n"
         )
-        logging.info(
-            f"unique indices / requested indices: {uvm_cache_stats[2]/uvm_cache_stats[1]}\n"
-            f"unique misses / requested indices: {uvm_cache_stats[3]/uvm_cache_stats[1]}\n"
-        )
+        if uvm_cache_stats[1]:
+            logging.info(
+                f"unique indices / requested indices: {uvm_cache_stats[2]/uvm_cache_stats[1]}\n"
+                f"unique misses / requested indices: {uvm_cache_stats[3]/uvm_cache_stats[1]}\n"
+            )
 
     def prefetch(self, indices: Tensor, offsets: Tensor) -> None:
         self.timestep += 1
@@ -2347,10 +2348,11 @@ def print_uvm_cache_stats(self) -> None:
             f"N_conflict_unique_misses: {uvm_cache_stats[4]}\n"
             f"N_conflict_misses: {uvm_cache_stats[5]}\n"
         )
-        logging.info(
-            f"unique indices / requested indices: {uvm_cache_stats[2]/uvm_cache_stats[1]}\n"
-            f"unique misses / requested indices: {uvm_cache_stats[3]/uvm_cache_stats[1]}\n"
-        )
+        if uvm_cache_stats[1]:
+            logging.info(
+                f"unique indices / requested indices: {uvm_cache_stats[2]/uvm_cache_stats[1]}\n"
+                f"unique misses / requested indices: {uvm_cache_stats[3]/uvm_cache_stats[1]}\n"
+            )
 
     @torch.jit.export
     def prefetch(self, indices: Tensor, offsets: Tensor) -> None:
@@ -3042,6 +3044,23 @@ def fill_random_weights(self) -> None:
                 )
             )
 
+    def fill_fp16_random_weights(self) -> None:
+        """
+        Fill the buffer with random weights, table by table
+        """
+        self.initialize_weights()
+        weights = self.split_embedding_weights()
+        for dest_weight in weights:
+            dest_weight[0].view(torch.float16).normal_(0, 0.1).view(torch.uint8)
+
+    def copy_weights(self, emb_tensor: torch.Tensor) -> None:
+        """
+        Fill the buffer with random weights, table by table
+        """
+        self.initialize_weights()
+        weights = self.split_embedding_weights()
+        weights[0][0].copy_(emb_tensor.view(torch.uint8))
+
     def assign_embedding_weights(
         self, q_weight_list: List[Tuple[Tensor, Optional[Tensor]]]
     ) -> None: