OPT-175B

CHANGELOG.md

@@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 
+### Added
+
+- Support for OPT-175B (AI2 only)
+
 ### Fixed
 
 - Fixed the way we compute SQuAD metrics.

catwalk/models/__init__.py

@@ -112,6 +112,10 @@ def _shorten_hf_name(hf_name: str) -> str:
     MODELS[f"metaicl::{name}"] = MetaICLModel(hf_name)
     MODELS[f"promptsource::{name}"] = PromptsourceDecoderOnlyRCModel(hf_name)
 
+MODELS["rc::opt-175b"] = DecoderOnlyRCModel(
+    "/net/nfs.cirrascale/allennlp/opt/opt-175b-huggingface",
+    pretrained_tokenizer_name_or_path="facebook/opt-66b")
+
 
 def short_name_for_model_object(model: Model) -> Optional[str]:
     for model_name, model_object in MODELS.items():

catwalk/models/rank_classification.py

@@ -25,6 +25,7 @@ def __init__(
         self,
         pretrained_model_name_or_path: str,
         *,
+        pretrained_tokenizer_name_or_path: Optional[str] = None,
         likelihood_averaging: str = 'token',
         **model_kwargs
     ):
@@ -41,13 +42,20 @@ def __init__(
         """
         assert likelihood_averaging in {'char', 'token'}
         self.pretrained_model_name_or_path = pretrained_model_name_or_path
+        if pretrained_tokenizer_name_or_path is None:
+            pretrained_tokenizer_name_or_path = pretrained_model_name_or_path
+        self.pretrained_tokenizer_name_or_path = pretrained_tokenizer_name_or_path
+
         self.likelihood_averaging = likelihood_averaging
         self.model_kwargs = model_kwargs
 
     @classmethod
     def _make_model(cls, pretrained_model_name_or_path: str, *, make_copy: bool = False, **kwargs) -> _Model:
         raise NotImplementedError
 
+    def _make_tokenizer(self) -> AutoTokenizer:
+        return cached_transformers.get_tokenizer(AutoTokenizer, self.pretrained_tokenizer_name_or_path)
+
     def predict(  # type: ignore
         self,
         task: Task,
@@ -62,7 +70,7 @@ def predict(  # type: ignore
             self.pretrained_model_name_or_path,
             device_map="auto" if torch.cuda.device_count() > 0 else None,
             **self.model_kwargs).eval()
-        tokenizer = cached_transformers.get_tokenizer(AutoTokenizer, self.pretrained_model_name_or_path)
+        tokenizer = self._make_tokenizer()
 
         for instance_chunk in more_itertools.chunked(instances, max_instances_in_memory):
             yield from self.predict_chunk(