remove stuff, reduce diff

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

Author: kylesayrs

examples/quantization_w4a16/llama3_example.py

@@ -5,7 +5,7 @@
 from llmcompressor.transformers import oneshot
 
 # Select model and load it.
-#MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
+# MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
 MODEL_ID = "meta-llama/Llama-3.2-1B-Instruct"
 
 model = AutoModelForCausalLM.from_pretrained(
@@ -23,7 +23,7 @@
 # Increasing the number of samples can improve accuracy.
 NUM_CALIBRATION_SAMPLES = 512
 MAX_SEQUENCE_LENGTH = 2048
-BATCH_SIZE = 2
+BATCH_SIZE = 8
 
 # Load dataset and preprocess.
 ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)

src/llmcompressor/modifiers/quantization/gptq/base.py

@@ -254,8 +254,6 @@ def on_initialize(self, state: State, **kwargs) -> bool:
                 )
             if isinstance(exception, unfixable_errors):
                 raise exception
-            
-            raise exception
 
             warnings.warn("Falling back to layer_sequential pipeline")
             try:

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

@@ -63,7 +63,6 @@ def format_calibration_data(
     tokenized_calibration = tokenized_dataset.select(range(safe_calibration_samples))
 
     # collate data
-    breakpoint()
     if collate_fn is None:
         tokenizer = getattr(processor, "tokenizer", processor)
         if hasattr(tokenizer, "pad"):

src/llmcompressor/transformers/finetune/runner.py

@@ -49,15 +49,14 @@ def __init__(
         data_args: "DataTrainingArguments",
         model_args: "ModelArguments",
         training_args: "TrainingArguments",
-        processor: Processor,
     ):
         self._data_args = data_args
         self._model_args = model_args
         self._training_args = training_args
 
         self.datasets = {}
         self.trainer = None
-        self.processor = processor
+        self.processor = None
         self.parent_output_dir = self._training_args.output_dir
         self._output_dir = self._training_args.output_dir
 
@@ -69,7 +68,6 @@ def populate_datasets(self, processor: Processor, add_labels: bool = True):
         :param processor: processor or tokenizer to use for dataset tokenization
         :param add_labels: if True, add labels column to dataset splits
         """
-        # TODO: remove `processor` arg in favor of self.processor
         if self._data_args.dataset is None:
             logger.info(
                 "Running oneshot without calibration data. This is expected for "
@@ -111,7 +109,7 @@ def _get_split_name(inp_str):
                     registry_id,
                     data_args=self._data_args,
                     split=split_str,
-                    processor=self.processor,
+                    processor=processor,
                 )
                 tokenized_datasets[split_name] = dataset_manager(add_labels=add_labels)
 

src/llmcompressor/transformers/finetune/text_generation.py

@@ -394,12 +394,10 @@ def main(
 
     # Load datasets
     stage_runner = StageRunner(
-        model_args=model_args, data_args=data_args, training_args=training_args, processor=processor
+        model_args=model_args, data_args=data_args, training_args=training_args
     )
     add_labels = training_args.do_train or training_args.run_stages
-    stage_runner.populate_datasets(
-        processor=processor, add_labels=add_labels
-    )
+    stage_runner.populate_datasets(processor=processor, add_labels=add_labels)
     train_dataset = stage_runner.get_dataset_split("train")
     eval_dataset = stage_runner.get_dataset_split("validation")
     calib_dataset = stage_runner.get_dataset_split("calibration")