allow passing trust remote code for tokenizer init in weight compress…

…ion (#830)

* allow passing trust remote code for tokenizer init in weight compression

* Update optimum/intel/openvino/modeling_decoder.py

* Apply suggestions from code review

Co-authored-by: Ilyas Moutawwakil <57442720+IlyasMoutawwakil@users.noreply.github.com>

---------

Co-authored-by: Ilyas Moutawwakil <57442720+IlyasMoutawwakil@users.noreply.github.com>

optimum/commands/export/openvino.py

@@ -262,6 +262,9 @@ def run(self):
                     "scale_estimation": self.args.scale_estimation,
                 }
 
+            if quantization_config.get("dataset", None) is not None:
+                quantization_config["trust_remote_code"] = self.args.trust_remote_code
+
             if self.args.weight_format in {"int4_sym_g128", "int4_asym_g128", "int4_sym_g64", "int4_asym_g64"}:
                 logger.warning(
                     f"--weight-format {self.args.weight_format} is deprecated, possible choices are fp32, fp16, int8, int4"

optimum/intel/openvino/configuration.py

@@ -263,6 +263,7 @@ def __init__(
         sym: bool = False,
         group_size: Optional[int] = None,
         tokenizer: Optional[str] = None,
+        trust_remote_code: bool = False,
         dataset: Optional[Union[str, List[str]]] = None,
         ratio: float = 1.0,
         all_layers: Optional[bool] = None,
@@ -275,6 +276,7 @@ def __init__(
     ):
         super().__init__(bits=bits, sym=sym, ignored_scope=ignored_scope, num_samples=num_samples)
         self.tokenizer = tokenizer
+        self.trust_remote_code = trust_remote_code
         self.dataset = dataset
         self.group_size = group_size or (-1 if bits == 8 else 128)
         self.ratio = ratio

optimum/intel/openvino/modeling_decoder.py

@@ -310,6 +310,7 @@ def _from_transformers(
             stateful=None,
             load_in_8bit=load_in_8bit,
             quantization_config=quantization_config,
+            trust_remote_code=trust_remote_code,
             **kwargs,
         )
 
@@ -781,6 +782,9 @@ def _from_pretrained(
 
         if isinstance(quantization_config, dict) and quantization_config == {"bits": 4}:
             quantization_config = get_default_int4_config(config.name_or_path)
+            if quantization_config.get("dataset", None) is not None:
+                quantization_config["trust_remote_code"] = kwargs.get("trust_remote_code", False)
+
         quantization_config = cls._prepare_weight_quantization_config(quantization_config, load_in_8bit)
 
         enable_compilation = kwargs.pop("compile", True) and not quantization_config

optimum/intel/openvino/quantization.py

@@ -672,7 +672,9 @@ def _remove_unused_columns(self, dataset: "Dataset"):
     def _prepare_builtin_dataset(self, quantization_config: OVWeightQuantizationConfig):
         from optimum.gptq.data import get_dataset, prepare_dataset
 
-        tokenizer = AutoTokenizer.from_pretrained(quantization_config.tokenizer)
+        tokenizer = AutoTokenizer.from_pretrained(
+            quantization_config.tokenizer, trust_remote_code=quantization_config.trust_remote_code
+        )
         nsamples = quantization_config.num_samples if quantization_config.num_samples else 128
         calibration_dataset = get_dataset(quantization_config.dataset, tokenizer, seqlen=32, nsamples=nsamples)
         calibration_dataset = prepare_dataset(calibration_dataset)