replace tokenizer with processor

Signed-off-by: Kyle Sayers <[email protected]>

src/llmcompressor/pytorch/model_load/helpers.py

@@ -9,6 +9,7 @@
 
 from llmcompressor.core import active_session, create_session, pre_initialize_structure
 from llmcompressor.pytorch.utils import ModuleSparsificationInfo
+from llmcompressor.typing import Processor
 
 COMPLETED_STAGES_FILENAME = "completed_stages.json"
 
@@ -92,15 +93,16 @@ def initialize_recipe(model: Module, recipe_path: str):
 def save_model_and_recipe(
     model: Module,
     save_path: str,
-    tokenizer: Optional[Any] = None,
+    processor: Optional[Processor] = None,
     save_safetensors: bool = False,
     save_compressed: bool = False,
 ):
     """
-    Save a model, tokenizer and the currently loaded recipe to file
+    Save a model, processor and the currently loaded recipe to file
+
     :param model: pytorch model to save
     :param save_path: path to save output to
-    :param tokenizer: model tokenizer to save
+    :param processor: model processor or tokenizer to save
     :param save_safetensors: whether to save as safetensors or pickle (bin)
     :param save_compressed: whether to compress sparse weights on disk
     """
@@ -111,8 +113,8 @@ def save_model_and_recipe(
         save_path, save_compressed=save_compressed, safe_serialization=save_safetensors
     )
 
-    if tokenizer is not None:
-        tokenizer.save_pretrained(save_path)
+    if processor is not None:
+        processor.save_pretrained(save_path)
 
     logger.info("Saving output to {}".format(os.path.abspath(save_path)))
 

src/llmcompressor/transformers/finetune/data/base.py

@@ -3,14 +3,14 @@
 from compressed_tensors.registry import RegistryMixin
 from datasets import Dataset, IterableDataset
 from loguru import logger
-from transformers import AutoTokenizer
 
 from llmcompressor.transformers.finetune.data.data_args import DataTrainingArguments
 from llmcompressor.transformers.finetune.data.data_helpers import (
     LABELS_MASK_VALUE,
     get_custom_datasets_from_path,
     get_raw_dataset,
 )
+from llmcompressor.typing import Processor
 
 
 class TextGenerationDataset(RegistryMixin):
@@ -30,10 +30,10 @@ def __init__(
         text_column: str,
         data_args: DataTrainingArguments,
         split: str,
-        tokenizer: AutoTokenizer,
+        processor: Processor,
     ):
         self.text_column = text_column
-        self.tokenizer = tokenizer
+        self.processor = processor
         self.data_args = data_args
         self.raw_kwargs = data_args.raw_kwargs or {}
         self.split = split
@@ -50,20 +50,38 @@ def __init__(
         else:
             self.padding = False
 
-        if self.tokenizer:
+        # get tokenizer
+        self.tokenizer = getattr(self.processor, "tokenizer", self.processor)
+
+        if self.tokenizer is not None:
+            # fill in pad token
             if not self.tokenizer.pad_token:
                 self.tokenizer.pad_token = self.tokenizer.eos_token
 
-        # configure sequence length
-        max_seq_length = data_args.max_seq_length
-        model_max_length = tokenizer.model_max_length if tokenizer else max_seq_length
-        if self.tokenizer and max_seq_length > model_max_length:
-            logger.warning(
-                f"The max_seq_length passed ({max_seq_length}) is larger than "
-                f"the maximum length for the model ({tokenizer.model_max_length}). "
-                f"Using max_seq_length={tokenizer.model_max_length}."
+            # configure sequence length
+            max_seq_length = data_args.max_seq_length
+            if data_args.max_seq_length > self.tokenizer.model_max_length:
+                logger.warning(
+                    f"The max_seq_length passed ({max_seq_length}) is larger than "
+                    f"maximum length for model ({self.tokenizer.model_max_length}). "
+                    f"Using max_seq_length={self.tokenizer.model_max_length}."
+                )
+            self.max_seq_length = min(
+                data_args.max_seq_length, self.tokenizer.model_max_length
+            )
+
+            # configure padding
+            self.padding = (
+                False
+                if self.data_args.concatenate_data
+                else "max_length"
+                if self.data_args.pad_to_max_length
+                else False
             )
-        self.max_seq_length = min(data_args.max_seq_length, model_max_length)
+
+        else:
+            self.max_seq_length = None
+            self.padding = False
 
     def get_raw_dataset(self, cache_dir: Optional[str] = None) -> Dataset:
         """

src/llmcompressor/transformers/finetune/data/c4.py

@@ -10,12 +10,12 @@ class C4Dataset(TextGenerationDataset):
 
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
     """
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         data_args = deepcopy(data_args)
         data_args.dataset = "allenai/c4"
         super().__init__(
-            text_column="text", data_args=data_args, split=split, tokenizer=tokenizer
+            text_column="text", data_args=data_args, split=split, processor=processor
         )

src/llmcompressor/transformers/finetune/data/cnn_dailymail.py

@@ -24,18 +24,18 @@ class CNNDailyMailDataset(TextGenerationDataset):
 
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
     """
 
     SAMPLE_TEMPLATE = "Article:\n{article}\n\n### Summarization:\n{highlights}\n"
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         data_args = deepcopy(data_args)
         data_args.dataset = "cnn_dailymail"
         data_args.dataset_config_name = "3.0.0"
 
         super().__init__(
-            text_column="text", data_args=data_args, split=split, tokenizer=tokenizer
+            text_column="text", data_args=data_args, split=split, processor=processor
         )
 
     def get_raw_dataset(self, cache_dir: Optional[str] = None):

src/llmcompressor/transformers/finetune/data/custom.py

@@ -32,17 +32,17 @@ class CustomDataset(TextGenerationDataset):
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
         Can also be set to None to load all the splits
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
 
     """
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         data_args = deepcopy(data_args)
         super().__init__(
             text_column=data_args.text_column,
             data_args=data_args,
             split=split,
-            tokenizer=tokenizer,
+            processor=processor,
         )
         self.preprocessing_func = data_args.preprocessing_func
         self.remove_columns = data_args.remove_columns

src/llmcompressor/transformers/finetune/data/evolcodealpaca.py

@@ -24,7 +24,7 @@ class EvolCodeAlpacaDataset(TextGenerationDataset):
 
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
     """
 
     EVOL_ALPACA_TEMPLATE = (
@@ -34,11 +34,11 @@ class EvolCodeAlpacaDataset(TextGenerationDataset):
         "\n\n### Response:\n"
     )
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         data_args = deepcopy(data_args)
         data_args.dataset = "theblackcat102/evol-codealpaca-v1"
         super().__init__(
-            text_column="text", data_args=data_args, split=split, tokenizer=tokenizer
+            text_column="text", data_args=data_args, split=split, processor=processor
         )
 
     def get_raw_dataset(self, cache_dir: Optional[str] = None):

src/llmcompressor/transformers/finetune/data/gsm8k.py

@@ -11,16 +11,16 @@ class GSM8KDataset(TextGenerationDataset):
 
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
     """
 
     GSM_TEMPLATE = "Question: {question}\nAnswer:"
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         data_args = deepcopy(data_args)
         data_args.dataset = "gsm8k"
         super().__init__(
-            text_column="text", data_args=data_args, split=split, tokenizer=tokenizer
+            text_column="text", data_args=data_args, split=split, processor=processor
         )
 
     def get_raw_dataset(self, cache_dir: Optional[str] = None):

src/llmcompressor/transformers/finetune/data/open_platypus.py

@@ -24,7 +24,7 @@ class OpenPlatypusDataset(TextGenerationDataset):
 
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
     """
 
     ALPACA_TEMPLATE = {
@@ -37,11 +37,11 @@ class OpenPlatypusDataset(TextGenerationDataset):
         "instruction}\n\n### Response:\n",
     }
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         data_args = deepcopy(data_args)
         data_args.dataset = "garage-bAInd/Open-Platypus"
         super().__init__(
-            text_column="text", data_args=data_args, split=split, tokenizer=tokenizer
+            text_column="text", data_args=data_args, split=split, processor=processor
         )
 
     def get_raw_dataset(self, cache_dir: Optional[str] = None):

src/llmcompressor/transformers/finetune/data/ptb.py

@@ -10,15 +10,15 @@ class PtbDataset(TextGenerationDataset):
 
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
     """
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         data_args = deepcopy(data_args)
         data_args.dataset = "ptb_text_only"
         super().__init__(
             text_column="sentence",
             data_args=data_args,
             split=split,
-            tokenizer=tokenizer,
+            processor=processor,
         )

src/llmcompressor/transformers/finetune/data/ultrachat_200k.py

@@ -24,7 +24,7 @@ class UltraChatDataset(TextGenerationDataset):
 
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
     """
 
     DEFAULT_CHAT_TEMPLATE = (
@@ -40,7 +40,7 @@ class UltraChatDataset(TextGenerationDataset):
         "{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}"
     )
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         data_args = deepcopy(data_args)
         data_args.dataset = "HuggingFaceH4/ultrachat_200k"
 
@@ -51,13 +51,15 @@ def __init__(self, data_args, split, tokenizer):
             text_column="messages",
             data_args=data_args,
             split=split,
-            tokenizer=tokenizer,
+            processor=processor,
         )
 
         if (
             not hasattr(self.tokenizer, "chat_template")
             or self.tokenizer.chat_template is None
         ):
+            # note that since tokenizer is a member of processor,
+            # this change affects processor.apply_chat_template
             self.tokenizer.chat_template = self.DEFAULT_CHAT_TEMPLATE
 
     def get_raw_dataset(self, cache_dir: Optional[str] = None):
@@ -75,7 +77,7 @@ def restructure_fn(sample):
             if sample["messages"][0]["role"] != "system":
                 sample["messages"].insert(0, {"role": "system", "content": ""})
 
-            sample["messages"] = self.tokenizer.apply_chat_template(
+            sample["messages"] = self.processor.apply_chat_template(
                 sample["messages"], tokenize=False, add_generation_prompt=False
             )
             return sample

src/llmcompressor/transformers/finetune/data/wikitext.py

@@ -8,10 +8,10 @@ class WikiTextDataset(TextGenerationDataset):
 
     :param data_args: configuration settings for dataset loading
     :param split: split from dataset to load, for instance `test` or `train[:5%]`
-    :param tokenizer: tokenizer to use on dataset
+    :param processor: processor or tokenizer to use on dataset
     """
 
-    def __init__(self, data_args, split, tokenizer):
+    def __init__(self, data_args, split, processor):
         super().__init__(
-            text_column="text", data_args=data_args, split=split, tokenizer=tokenizer
+            text_column="text", data_args=data_args, split=split, processor=processor
         )

src/llmcompressor/transformers/finetune/model_args.py

@@ -34,6 +34,12 @@ class ModelArguments:
             "help": "Pretrained tokenizer name or path if not the same as model_name"
         },
     )
+    processor: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Pretrained processor name or path if not the same as model_name"
+        },
+    )
     cache_dir: Optional[str] = field(
         default=None,
         metadata={"help": "Where to store the pretrained data from huggingface.co"},