linting and formatting

Signed-off-by: Yash Mehan <yashmehan@gmail.com>

Author: yash4242

build/accelerate_launch.py

@@ -120,7 +120,7 @@ def main():
         logging.error(traceback.format_exc())
         # v5: torch.distributed raises ChildFailedError with per-rank exit codes
         # Check if the root cause was a user error
-        if hasattr(e, 'failures'):
+        if hasattr(e, "failures"):
             root_codes = [f.exitcode for f in e.failures.values()]
             if any(c == USER_ERROR_EXIT_CODE for c in root_codes):
                 sys.exit(USER_ERROR_EXIT_CODE)

pyproject.toml

@@ -28,9 +28,9 @@ classifiers=[
 dependencies = [
 "numpy>=1.26.4,<2.2.0",
 "accelerate>=1.9.0,<2.0.0",
-"transformers>=5.2.0,<=5.3.0",
-"torch>=2.10.0,<2.11.0",
-"torchvision<=0.25.1",
+"transformers>=5.2.0,<5.3.0",
+"torch>2.7.0,<=2.9.0",
+"torchvision<=0.24.0",
 "sentencepiece>=0.1.99,<0.3",
 "tokenizers<=0.23.0",
 "tqdm>=4.66.2,<5.0",
@@ -40,7 +40,7 @@ dependencies = [
 "simpleeval>=0.9.13,<2.0",
 "pillow>=12.1.1",
 "kernels>=0.12.1,<0.13.0",
-"huggingface_hub>=1.3.0,<1.4.0"
+"huggingface_hub>=1.3.0,<1.4.0",
 ]
 
 [project.optional-dependencies]

tests/utils/test_embedding_resize.py

@@ -20,11 +20,9 @@
 
 # Third Party
 from transformers import (
-    AutoModelForCausalLM,
-    AutoModelForImageTextToText, #AutoModelForVision2Seq was renamed to this in transformers v5
-    AutoProcessor,
-    AutoTokenizer,
+    AutoModelForImageTextToText,  # AutoModelForVision2Seq was renamed to this in transformers v5
 )
+from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer
 import torch
 
 # First Party
@@ -136,8 +134,9 @@ def test_special_tokens_before_and_after():
     special_tokens_dict = {"sep_token": "<SEP>", "pad_token": "<PAD>"}
     addn_spl_tokens_added = ["<NotSeenTokenA>", "<NotSeenTokenB>", "<NotSeenTokenC>"]
     # for transformers v5: merge existing extra_special_tokens with new ones to prevent replacement
-    special_tokens_dict["additional_special_tokens"] = list(tokenizer.extra_special_tokens) + addn_spl_tokens_added
-
+    special_tokens_dict["additional_special_tokens"] = (
+        list(tokenizer.extra_special_tokens) + addn_spl_tokens_added
+    )
 
     resize_result = tokenizer_and_embedding_resize(
         special_tokens_dict=special_tokens_dict,
@@ -210,7 +209,9 @@ def test_resize_with_multiple_of():
 
 
 def test_resize_llama_vision_model():
-    model = AutoModelForImageTextToText.from_pretrained(TINY_LLAMA_VISION_MODEL_NAME) # AutoModelForVision2Seq was renamed to AutoModelForImageTextToText in transformers v5
+    model = AutoModelForImageTextToText.from_pretrained(
+        TINY_LLAMA_VISION_MODEL_NAME
+    )  # AutoModelForVision2Seq was renamed to AutoModelForImageTextToText in transformers v5
     processor = AutoProcessor.from_pretrained(TINY_LLAMA_VISION_MODEL_NAME)
     tokenizer = processor.tokenizer
 

tuning/data/tokenizer_utils.py

@@ -45,31 +45,32 @@ def get_special_tokens_dict(
     special_tokens_dict = {}
     if not tokenizer_name_or_path:
         llama_classes = tuple(
-            cls for cls in [
+            cls
+            for cls in [
                 getattr(transformers, "LlamaTokenizer", None),
                 getattr(transformers, "LlamaTokenizerFast", None),
-            ] if cls is not None
+            ]
+            if cls is not None
         )
         is_llama_tokenizer = (
-            (bool(llama_classes) and isinstance(tokenizer, llama_classes))
-            or "llama" in (getattr(tokenizer, "name_or_path", "") or "").lower()
-        )
+            bool(llama_classes) and isinstance(tokenizer, llama_classes)
+        ) or "llama" in (getattr(tokenizer, "name_or_path", "") or "").lower()
 
         gpt_neox_classes = tuple(
-            cls for cls in [
+            cls
+            for cls in [
                 getattr(transformers, "GPTNeoXTokenizerFast", None),
                 getattr(transformers, "GPTNeoXTokenizer", None),
-            ] if cls is not None
+            ]
+            if cls is not None
         )
 
         if is_llama_tokenizer:
             special_tokens_dict["bos_token"] = "<s>"
             special_tokens_dict["eos_token"] = "</s>"
             special_tokens_dict["unk_token"] = "<unk>"
             special_tokens_dict["pad_token"] = "<pad>"
-        elif isinstance(
-            tokenizer, (transformers.GPT2Tokenizer, *gpt_neox_classes)
-        ):
+        elif isinstance(tokenizer, (transformers.GPT2Tokenizer, *gpt_neox_classes)):
             special_tokens_dict["pad_token"] = "<pad>"
 
         # Add special tokens only when a custom tokenizer is not passed
@@ -117,7 +118,7 @@ def tokenizer_and_embedding_resize(
         dict: Metadata on number of added tokens.
     """
     num_new_tokens = tokenizer.add_special_tokens(
-        special_tokens_dict=special_tokens_dict, 
+        special_tokens_dict=special_tokens_dict,
         # replace_additional_special_tokens=False
     )
     embedding_size = int(multiple_of * math.ceil(len(tokenizer) / multiple_of))

tuning/sft_trainer.py

@@ -28,9 +28,11 @@
 from peft import LoraConfig
 from peft.utils.other import fsdp_auto_wrap_policy
 from torch.cuda import OutOfMemoryError
+from transformers import (
+    AutoModelForImageTextToText,  # AutoModelForVision2Seq was renamed in transformers v5
+)
 from transformers import (
     AutoModelForCausalLM,
-    AutoModelForImageTextToText, # AutoModelForVision2Seq was renamed in transformers v5
     AutoProcessor,
     AutoTokenizer,
     TrainerCallback,