apply style, ignore visual on qwen

examples/multimodal_vision/mllama.py

@@ -1,10 +1,8 @@
 import os
 
-import torch
 from transformers import AutoProcessor
 
 from llmcompressor.modifiers.quantization import GPTQModifier
-
 from llmcompressor.transformers import oneshot
 from llmcompressor.transformers.tracing import TracableMllamaForConditionalGeneration
 from llmcompressor.transformers.utils.data_collator import mllama_data_collator

examples/multimodal_vision/pixtral.py

@@ -1,6 +1,5 @@
 import os
 
-import torch
 from transformers import AutoProcessor
 
 from llmcompressor.modifiers.quantization import GPTQModifier

examples/multimodal_vision/qwen2_vl.py

@@ -1,6 +1,5 @@
 import os
 
-import torch
 from compressed_tensors.quantization import (
     QuantizationArgs,
     QuantizationScheme,
@@ -43,7 +42,7 @@
             ),
         ),
     },
-    ignore=["re:.*lm_head"],
+    ignore=["re:visual.*", "re:.*lm_head"],
     dampening_frac=0.5,
 )
 

src/llmcompressor/modifiers/utils/pytorch_helpers.py

@@ -1,5 +1,5 @@
 from itertools import cycle
-from typing import Callable, Dict, List, Optional, Tuple, Any
+from typing import Any, Callable, Dict, List, Optional, Tuple
 
 import torch
 from torch.nn import Module

src/llmcompressor/pipelines/basic/pipeline.py

@@ -1,18 +1,18 @@
 import torch
 import torch.utils.data.dataloader
 import tqdm
+from compressed_tensors.utils import get_execution_device
 
 from llmcompressor.modifiers.utils.pytorch_helpers import apply_pad_mask_to_batch
 from llmcompressor.pytorch.utils.helpers import tensors_to_device
 from llmcompressor.utils.helpers import calibration_forward_context
-from compressed_tensors.utils import get_execution_device
 
 __all__ = ["run_pipeline"]
 
 
 def run_pipeline(model: torch.nn.Module, dataloader: torch.utils.data.DataLoader):
     model_device = get_execution_device(model)
-    
+
     with calibration_forward_context(model):
         for batch in tqdm.tqdm(dataloader, desc="Calibrating"):
             batch = apply_pad_mask_to_batch(batch)

src/llmcompressor/transformers/utils/data_collator.py

@@ -2,6 +2,7 @@
 
 __all__ = ["mllama_data_collator", "pixtral_data_collator"]
 
+
 def mllama_data_collator(batch):
     assert len(batch) == 1
     return {
@@ -13,6 +14,7 @@ def mllama_data_collator(batch):
         "cross_attention_mask": torch.tensor(batch[0]["cross_attention_mask"]),
     }
 
+
 def pixtral_data_collator(batch):
     assert len(batch) == 1
     return {
@@ -21,11 +23,12 @@ def pixtral_data_collator(batch):
         "pixel_values": torch.tensor(batch[0]["pixel_values"])[0],
     }
 
+
 def qwen2_vl_data_collator(batch):
     assert len(batch) == 1
     return {
         "input_ids": torch.LongTensor(batch[0]["input_ids"]),
         "attention_mask": torch.tensor(batch[0]["attention_mask"]),
         "pixel_values": torch.tensor(batch[0]["pixel_values"]),
         "image_grid_thw": torch.tensor(batch[0]["image_grid_thw"]),
-    }
+    }