Enable export paths for LLMs [CodeGen, OPT, Bloom] (#1562)

* initital commit

* missing config arg

Author: dbogunowicz

src/sparseml/transformers/export.py

@@ -136,6 +136,13 @@ def load_task_model(task: str, model_path: str, config: Any) -> Module:
             model_type="model",
         )
 
+    if task == "text-generation":
+        return SparseAutoModel.text_generation_from_pretrained(
+            model_name_or_path=model_path,
+            config=config,
+            model_type="model",
+        )
+
     raise ValueError(f"unrecognized task given of {task}")
 
 
@@ -263,6 +270,9 @@ def export_transformer_to_onnx(
     tokenizer = AutoTokenizer.from_pretrained(
         model_path, model_max_length=sequence_length
     )
+    if task == "text-generation":
+        tokenizer.pad_token = tokenizer.eos_token
+
     model = load_task_model(task, model_path, config)
     _LOGGER.info(f"loaded model, config, and tokenizer from {model_path}")
 
@@ -353,12 +363,14 @@ def export_transformer_to_onnx(
     # run export
     model = model.eval()
     onnx_file_path = os.path.join(model_path, onnx_file_name)
+    kwargs = {"input_names": list(inputs.keys())} if task == "text-generation" else {}
 
     export_onnx(
         model,
         inputs,
         onnx_file_path,
         convert_qat=convert_qat,
+        **kwargs,
     )
     _LOGGER.info(f"ONNX exported to {onnx_file_path}")
 

src/sparseml/transformers/utils/model.py

@@ -20,6 +20,7 @@
 import torch
 from torch.nn import Module
 from transformers import (
+    AutoModelForCausalLM,
     AutoModelForMaskedLM,
     AutoModelForQuestionAnswering,
     AutoModelForSequenceClassification,
@@ -235,6 +236,41 @@ def text_classification_from_pretrained_distil(
 
         return model, teacher
 
+    @staticmethod
+    def text_generation_from_pretrained(
+        model_name_or_path: str,
+        model_type: str,
+        **kwargs,
+    ) -> Module:
+        """
+        :param model_name_or_path: the name of or path to the model to load
+        :param model_type: specify the type of model loaded for logging;
+            ex one of [model, student, teacher]
+        :param kwargs: keyword arguments to pass through to the AutoModel call
+        :return: the created model for text generation
+        """
+        SparseAutoModel._check_tf(model_name_or_path)
+        if not kwargs:
+            kwargs = {}
+        kwargs["from_tf"] = False
+        delayed = False
+        if "state_dict" not in kwargs:
+            kwargs["state_dict"], delayed = SparseAutoModel._loadable_state_dict(
+                model_name_or_path
+            )
+        # Export decoder model without kv cache support
+        kwargs["config"].is_decoder = True
+        kwargs["config"].use_cache = False
+        kwargs["config"].use_past = False
+
+        model = AutoModelForCausalLM.from_pretrained(
+            model_name_or_path,
+            **kwargs,
+        )
+        SparseAutoModel.log_model_load(model, model_name_or_path, model_type, delayed)
+
+        return model
+
     @staticmethod
     def token_classification_from_pretrained(
         model_name_or_path: str,