Improve offline latency by skipping checks for model updates

changelog.md

@@ -2,6 +2,10 @@
 
 ## 0.18.0
 
+### Added
+
+- Improve offline latency by skipping checks for model updates
+
 ### Fixed
 
 - Correct issue causing `choices` to be scored improperly

languagemodels/models.py

@@ -1,5 +1,7 @@
 import re
+from functools import partial
 from huggingface_hub import hf_hub_download, snapshot_download
+from huggingface_hub.utils import LocalEntryNotFoundError
 from tokenizers import Tokenizer
 import ctranslate2
 
@@ -34,7 +36,13 @@ def get_model_info(model_type="instruct"):
 def initialize_tokenizer(model_type, model_name):
     model_info = get_model_info(model_type)
 
-    tok_config = hf_hub_download(model_info["path"], "tokenizer.json")
+    get_model = partial(hf_hub_download, model_info["path"], "tokenizer.json")
+
+    try:
+        tok_config = get_model(local_files_only=True)
+    except LocalEntryNotFoundError:
+        tok_config = get_model()
+
     tokenizer = Tokenizer.from_file(tok_config)
 
     if model_type == "embedding":
@@ -47,8 +55,13 @@ def initialize_tokenizer(model_type, model_name):
 def initialize_model(model_type, model_name):
     model_info = get_model_info(model_type)
 
-    path = snapshot_download(model_info["path"], max_workers=1,
-                             allow_patterns=["*.bin", "*.txt", "*.json"])
+    get_path = partial(snapshot_download, model_info["path"], max_workers=1,
+                       allow_patterns=["*.bin", "*.txt", "*.json"])
+
+    try:
+        path = get_path(local_files_only=True)
+    except LocalEntryNotFoundError:
+        path = get_path()
 
     if model_info["architecture"] == "encoder-only-transformer":
         return ctranslate2.Encoder(path, "cpu", compute_type="int8", )