refactor: move tokenizer_data_utils with the rest of utils, add further unit testing.

scripts/run_inference.py

@@ -34,7 +34,7 @@
 import torch
 
 # Local
-from tuning.data import tokenizer_data_utils
+from tuning.utils.tokenizer_data_utils import tokenizer_and_embedding_resize
 
 
 ### Utilities
@@ -219,7 +219,7 @@ def load(
                     # where the model's layers are modified, in our case the embedding layer
                     # is modified, so we resize the backbone model's embedding layer with our own
                     # utility before passing it along to load the PEFT model.
-                    tokenizer_data_utils.tokenizer_and_embedding_resize(
+                    tokenizer_and_embedding_resize(
                         {}, tokenizer=tokenizer, model=base_model
                     )
                     model = PeftModel.from_pretrained(

tests/utils/test_embedding_resize.py

@@ -20,9 +20,10 @@
 import torch
 
 # Local
-from tuning.data import tokenizer_data_utils
+from tuning.utils.tokenizer_data_utils import tokenizer_and_embedding_resize
 
 MODEL_NAME = "Maykeye/TinyLLama-v0"
+INPUT_TEXT = "### Text: @NortonSupport Thanks much.\n\n### Label:"
 
 
 def _inference(
@@ -41,16 +42,16 @@ def _inference(
 
 
 def test_output_unaltered_across_embedding_resizes():
-    input_text = "### Text: @NortonSupport Thanks much.\n\n### Label:"
+    input_text = INPUT_TEXT
     tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
     model_not_resized = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
     model_resized = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
 
-    tokenizer_data_utils.tokenizer_and_embedding_resize(
+    tokenizer_and_embedding_resize(
         special_tokens_dict={}, tokenizer=tokenizer, model=model_resized, multiple_of=8
     )
 
-    tokenizer_data_utils.tokenizer_and_embedding_resize(
+    tokenizer_and_embedding_resize(
         special_tokens_dict={},
         tokenizer=tokenizer,
         model=model_not_resized,
@@ -74,3 +75,68 @@ def test_output_unaltered_across_embedding_resizes():
     )
 
     assert output_from_model_not_resized == output_from_model_resized
+
+
+def test_resize_with_special_tokens():
+    input_text = INPUT_TEXT
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+    model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
+
+    input_tokenizer_len = len(tokenizer.get_vocab())
+
+    special_tokens = {"sep_token": "<SEP>", "pad_token": "<PAD>"}
+    resize_result = tokenizer_and_embedding_resize(
+        special_tokens_dict=special_tokens,
+        tokenizer=tokenizer,
+        model=model,
+        multiple_of=1,
+    )
+
+    assert "<SEP>" in tokenizer.get_vocab()
+    assert "<PAD>" in tokenizer.get_vocab()
+
+    output_tokenizer_len = len(tokenizer.get_vocab())
+
+    assert output_tokenizer_len == input_tokenizer_len + 2
+    assert resize_result["num_new_tokens"] == output_tokenizer_len - input_tokenizer_len
+
+    output = _inference(
+        tokenizer=tokenizer, model=model, input_text=input_text, max_new_tokens=20
+    )
+    assert output is not None
+
+
+def test_no_resize_when_no_special_tokens():
+    input_text = INPUT_TEXT
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+    model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
+
+    input_tokenizer_len = len(tokenizer.get_vocab())
+
+    resize_result = tokenizer_and_embedding_resize(
+        special_tokens_dict={}, tokenizer=tokenizer, model=model, multiple_of=1
+    )
+
+    output_tokenizer_len = len(tokenizer.get_vocab())
+
+    assert input_tokenizer_len == output_tokenizer_len
+    assert resize_result["num_new_tokens"] == 0
+
+    output = _inference(
+        tokenizer=tokenizer, model=model, input_text=input_text, max_new_tokens=20
+    )
+
+    assert output is not None
+
+
+def test_resize_with_multiple_of():
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+    model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
+
+    resize_result = tokenizer_and_embedding_resize(
+        special_tokens_dict={}, tokenizer=tokenizer, model=model, multiple_of=8
+    )
+
+    assert model.get_input_embeddings().embedding_dim % 8 == 0
+    assert resize_result["new_embedding_size"] % 8 == 0
+    assert model.get_output_embeddings().out_features % 8 == 0

tests/utils/test_tokenizer_data_utils.py

@@ -7,7 +7,7 @@
 
 # Local
 # First party
-from tuning.data.tokenizer_data_utils import tokenizer_and_embedding_resize
+from tuning.utils.tokenizer_data_utils import tokenizer_and_embedding_resize
 
 
 def test_tokenizer_and_embedding_resize_return_values():

tuning/sft_trainer.py

@@ -53,7 +53,6 @@
     FileLoggingTrackerConfig,
     TrackerConfigFactory,
 )
-from tuning.data import tokenizer_data_utils
 from tuning.trackers.tracker_factory import FILE_LOGGING_TRACKER, get_tracker
 from tuning.trainercontroller import TrainerControllerCallback
 from tuning.utils.config_utils import get_hf_peft_config, get_json_config
@@ -70,6 +69,7 @@
     is_pretokenized_dataset,
     validate_data_args,
 )
+from tuning.utils.tokenizer_data_utils import tokenizer_and_embedding_resize
 
 
 def train(
@@ -294,7 +294,7 @@ def train(
 
     # TODO: lower priority but understand if resizing impacts inference quality and why its needed.
     # It makes sense if we manipulate tokenizer that we also save it and provide it to inference.
-    added_tokens_dict = tokenizer_data_utils.tokenizer_and_embedding_resize(
+    added_tokens_dict = tokenizer_and_embedding_resize(
         special_tokens_dict=special_tokens_dict,
         tokenizer=tokenizer,
         model=model,