Update asset caching logic for toxicity model

langkit/asset_downloader.py

@@ -65,7 +65,7 @@ def _is_zip_file(file_path: str) -> bool:
         return False
 
 
-@retry(stop=stop_after_attempt(3), wait=wait_exponential_jitter(max=5))
+# @retry(stop=stop_after_attempt(3), wait=wait_exponential_jitter(max=5))
 def _download_asset(asset_id: str, tag: str = "0"):
     asset_path = _get_asset_path(asset_id, tag)
     response: GetAssetResponse = cast(GetAssetResponse, assets_api.get_asset(asset_id))

langkit/metrics/toxicity.py

@@ -22,23 +22,34 @@ def __toxicity(pipeline: TextClassificationPipeline, max_length: int, text: List
     return [result["score"] if result["label"] == "toxic" else 1.0 - result["score"] for result in results]  # type: ignore
 
 
-__model_path = "martin-ha/toxic-comment-model"
+_model_path = "martin-ha/toxic-comment-model"
+_revision = "9842c08b35a4687e7b211187d676986c8c96256d"
+
+
+def _cache_assets():
+    AutoModelForSequenceClassification.from_pretrained(_model_path, revision=_revision)
+    AutoTokenizer.from_pretrained(_model_path, revision=_revision)
 
 
 @lru_cache
 def _get_tokenizer() -> PreTrainedTokenizerBase:
-    return AutoTokenizer.from_pretrained(__model_path)
+    return AutoTokenizer.from_pretrained(_model_path, local_files_only=True, revision=_revision)
 
 
 @lru_cache
 def _get_pipeline() -> TextClassificationPipeline:
     use_cuda = torch.cuda.is_available() and not bool(os.environ.get("LANGKIT_NO_CUDA", False))
-    model: PreTrainedTokenizerBase = AutoModelForSequenceClassification.from_pretrained(__model_path)
+    model: PreTrainedTokenizerBase = AutoModelForSequenceClassification.from_pretrained(
+        _model_path, local_files_only=True, revision=_revision
+    )
     tokenizer = _get_tokenizer()
     return TextClassificationPipeline(model=model, tokenizer=tokenizer, device=0 if use_cuda else -1)
 
 
 def toxicity_metric(column_name: str) -> Metric:
+    def cache_assets():
+        _cache_assets()
+
     def init():
         _get_pipeline()
 
@@ -51,7 +62,9 @@ def udf(text: pd.DataFrame) -> SingleMetricResult:
         metrics = __toxicity(_pipeline, max_length, col)
         return SingleMetricResult(metrics=metrics)
 
-    return SingleMetric(name=f"{column_name}.toxicity.toxicity_score", input_names=[column_name], evaluate=udf, init=init)
+    return SingleMetric(
+        name=f"{column_name}.toxicity.toxicity_score", input_names=[column_name], evaluate=udf, init=init, cache_assets=cache_assets
+    )
 
 
 prompt_toxicity_metric = partial(toxicity_metric, "prompt")

langkit/metrics/whylogs_compat.py

@@ -105,6 +105,8 @@ def to_udf_schema_args(metric: Metric) -> List[UdfSchemaArgs]:
 
 def create_whylogs_udf_schema(eval_conf: WorkflowMetricConfig) -> UdfSchema:
     for metric in eval_conf.metrics:
+        if metric.cache_assets:
+            metric.cache_assets()
         if metric.init:
             metric.init()