woof

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

Author: kylesayrs

examples/multimodal_vision/phi3_vision_example.py

@@ -3,7 +3,6 @@
 from transformers import AutoModelForCausalLM, AutoProcessor
 
 from llmcompressor.modifiers.quantization import GPTQModifier
-from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
 from llmcompressor.transformers import oneshot
 
 # Load model.
@@ -16,7 +15,6 @@
     _attn_implementation="eager",
 )
 processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
-processor.chat_template = processor.tokenizer.chat_template
 
 # Oneshot arguments
 DATASET_ID = "lmms-lab/flickr30k"
@@ -68,7 +66,6 @@ def data_collator(batch):
 
 # Recipe
 recipe = [
-    SmoothQuantModifier(smoothing_strength=0.8),
     GPTQModifier(
         targets="Linear",
         scheme="W4A16",

src/llmcompressor/pipelines/sequential/helpers.py

@@ -71,7 +71,6 @@ def trace_subgraphs(
     concrete_args = populate_concrete_args(model, sample_input)
 
     # trace
-    breakpoint()
     with (
         calibration_forward_context(model),
         HooksMixin.disable_hooks(),

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

@@ -81,6 +81,7 @@ def format_calibration_data(
         else SequentialSampler(tokenized_calibration),
         "collate_fn": collate_fn,
         "pin_memory": True,
+        "drop_last": False,
     }
 
     calib_dataloader = DataLoader(tokenized_calibration, **dataloader_params)

src/llmcompressor/transformers/finetune/runner.py

@@ -68,6 +68,7 @@ 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
         """
+        self.processor = processor  # TODO: pass processor into init instead of this fn
         if self._data_args.dataset is None:
             logger.info(
                 "Running oneshot without calibration data. This is expected for "