diff --git a/gpu_bdb/bdb_tools/cluster_startup.py b/gpu_bdb/bdb_tools/cluster_startup.py
index de078717..378c1036 100755
--- a/gpu_bdb/bdb_tools/cluster_startup.py
+++ b/gpu_bdb/bdb_tools/cluster_startup.py
@@ -103,12 +103,13 @@ def maybe_create_worker_directories(dask_worker):
     # Assumes all GPUs are the same size
     expected_workers = int(os.environ.get("NUM_WORKERS", 16))
     worker_counts = worker_count_info(client)
+    current_workers = None
     for gpu_size, count in worker_counts.items():
         if count != 0:
             current_workers = worker_counts.pop(gpu_size)
             break
-
-    if expected_workers is not None and expected_workers != current_workers:
+        
+    if expected_workers is not None and  current_workers is not None and expected_workers != current_workers:
         print(
             f"Expected {expected_workers} {gpu_size} workers in your cluster, but got {current_workers}. It can take a moment for all workers to join the cluster. You may also have misconfigred hosts."
         )
diff --git a/gpu_bdb/bdb_tools/q27_bert_utils.py b/gpu_bdb/bdb_tools/q27_bert_utils.py
index 0d28e6f3..ebf0a8a5 100755
--- a/gpu_bdb/bdb_tools/q27_bert_utils.py
+++ b/gpu_bdb/bdb_tools/q27_bert_utils.py
@@ -172,23 +172,30 @@ def tokenize_text_series(text_ser, seq_len, stride, vocab_hash_file):
          A dictionary with these keys {'token_ar':,'attention_ar':,'metadata':}
 
     """
+    from cudf._lib.nvtext.subword_tokenize import Hashed_Vocabulary as cpp_hashed_vocabulary
+    from cudf._lib.nvtext.subword_tokenize import subword_tokenize_inmem_hash as cpp_subword_tokenize
+
+    vocab_file = cpp_hashed_vocabulary(vocab_hash_file)
+
     if len(text_ser) == 0:
         return {"token_ar": None, "attention_ar": None, "metadata": None}
 
-    max_num_chars = text_ser.str.len().sum() + 1
-    max_rows_tensor = len(text_ser) * 2
+    # max_num_chars = text_ser.str.len().sum() + 1
+    # max_rows_tensor = len(text_ser) * 2
     max_length = seq_len - 2
 
-    tokens, attention_masks, metadata = text_ser.str.subword_tokenize(
-        vocab_hash_file,
-        do_lower=False,
-        max_num_strings=max_rows_tensor,
-        max_rows_tensor=max_rows_tensor,
-        max_num_chars=max_num_chars,
-        stride=stride,
-        max_length=max_length,
-        do_truncate=False,
-    )
+    tokens, attention_mask, metadata = cpp_subword_tokenize(
+            text_ser._column,
+            vocab_file,
+            do_lower=False,
+            stride=stride,
+            max_sequence_length=max_length,
+            do_truncate=False,
+        )
+    
+    tokens = cp.asarray(tokens)
+    attention_masks = cp.asarray(attention_mask)
+    metadata = cp.asarray(metadata)
     del text_ser
     ### reshape metadata into a matrix
     metadata = metadata.reshape(-1, 3)
@@ -301,7 +308,7 @@ def create_vocab_table(vocabpath):
     """
     id2vocab = []
     vocab2id = {}
-    with open(vocabpath) as f:
+    with open(vocabpath, encoding="utf-8") as f:
         for index, line in enumerate(f):
             token = line.split()[0]
             id2vocab.append(token)
diff --git a/gpu_bdb/bdb_tools/q27_get_review_sentence_utils.py b/gpu_bdb/bdb_tools/q27_get_review_sentence_utils.py
index b7ce3e21..1b7946d8 100755
--- a/gpu_bdb/bdb_tools/q27_get_review_sentence_utils.py
+++ b/gpu_bdb/bdb_tools/q27_get_review_sentence_utils.py
@@ -233,6 +233,7 @@ def get_org_sentences(sentence_boundary_df, org_df):
         inplace=True,
     )
     valid_left_loc = valid_left_loc.reset_index(drop=False)
+    valid_left_loc = valid_left_loc.rename(columns={"index":"flat_loc_org"})
 
     ### Better way to get the closeset row/col maybe
     valid_right_loc = (
@@ -249,8 +250,9 @@ def get_org_sentences(sentence_boundary_df, org_df):
     valid_right_loc = valid_right_loc[["r_fs_seq_row", "r_fs_seq_col"]].reset_index(
         drop=False
     )
+    valid_right_loc = valid_right_loc.rename(columns={"index":"flat_loc_org"})
 
-    valid_df = valid_left_loc.merge(valid_right_loc)
+    valid_df = valid_left_loc.merge(valid_right_loc, on="flat_loc_org")
     valid_df = valid_df.set_index(["flat_loc_org"])
 
     return valid_df
@@ -352,3 +354,4 @@ def get_org_df(pr_label_f, metadata_df, seq_len):
     org_df = org_df[flag]
 
     return org_df[["org_seq_row", "org_seq_col", "input_text_index", "flat_loc_org"]]
+
diff --git a/gpu_bdb/queries/q27/gpu_bdb_query_hf_27.py b/gpu_bdb/queries/q27/gpu_bdb_query_hf_27.py
index 312eb200..284c75fd 100755
--- a/gpu_bdb/queries/q27/gpu_bdb_query_hf_27.py
+++ b/gpu_bdb/queries/q27/gpu_bdb_query_hf_27.py
@@ -119,12 +119,11 @@ def main(client, config):
 
     import cudf
 
-    model_path = os.path.join(config["data_dir"], "../../q27_model_dir")
+    model_path = os.path.join(config["data_dir"], "/datasets/vjawa/distilbert-base-en-cased/")
     product_reviews_df = benchmark(
         read_tables,
         config=config,
         compute_result=config["get_read_time"],
-        dask_profile=config["dask_profile"],
     )
     product_reviews_df = product_reviews_df[
         product_reviews_df.pr_item_sk == q27_pr_item_sk
@@ -142,7 +141,8 @@ def main(client, config):
     )
     output_df = output_df.persist()
     wait(output_df)
-    client.run(del_model_attribute)
+    print(output_df.head())
+    #client.run(del_model_attribute)
     return output_df
 
 
@@ -153,5 +153,6 @@ def main(client, config):
 
     config = gpubdb_argparser()
     client, bc = attach_to_cluster(config)
-    client.run(rmm.reinitialize, pool_allocator=True, initial_pool_size=14e9)
+    
+    #client.run(rmm.reinitialize, pool_allocator=True, initial_pool_size=14e9)
     run_query(config=config, client=client, query_func=main)