Fix mask_embeddings_by_frequency edge case

Author: z52527

corelib/dynamicemb/src/dynamic_emb_op.cu

@@ -567,7 +567,7 @@ if (tables[0]->evict_strategy() == EvictStrategy::kLfu && frequency_threshold >
 #ifdef DEBUG
 
     if (frequency_threshold > 0 && mask_dims > 0) {
-      printf("Masking enabled\n");
+      // printf("Masking enabled\n");
       at::Tensor h_unique_embs = unique_embs.cpu();
       at::Tensor h_unique_embeddings_for_scatter = unique_embeddings_for_scatter.cpu();
       at::Tensor h_unique_output_scores = unique_output_scores.cpu();

corelib/dynamicemb/src/lookup_forward.cu

@@ -275,8 +275,8 @@ void mask_embeddings_by_frequency(void *embeddings_ptr, void *scores_ptr,
                                   int frequency_threshold, int mask_dims,
                                   DataType emb_type, DataType score_type,
                                   int device_num_sms, cudaStream_t stream) {
-  if (frequency_threshold <= 0 || mask_dims <= 0) {
-    return;  // No masking needed
+  if (frequency_threshold <= 0 || mask_dims <= 0 || num_embeddings <= 0) {
+    return;  // No masking needed or no embeddings to process
   }
 
   int num_warps_needed = num_embeddings; // each warp processes one embedding