Skip calculate feature score threshold for ttl in feature score eviction policy (#3552)

EddyLXJ · meta-codesync[bot] · commit 5f98e016260a · 2025-11-18T01:11:50.000-08:00
Summary: Pull Request resolved: #3552 AS title, no need to calculate feature score threshold for ttl in feature score eviction policy Reviewed By: steven1327 Differential Revision: D87278882 Privacy Context Container: L1158701 fbshipit-source-id: 4e1874b50de9c1f2bc28375d33a04770136e7cb4
diff --git a/torchrec/modules/embedding_configs.py b/torchrec/modules/embedding_configs.py
@@ -237,7 +237,7 @@ class FeatureScoreBasedEvictionPolicy(VirtualTableEvictionPolicy):
         0  # max number of inference ids per rank, default is training_id_keep_count
     )
     inference_eviction_feature_score_threshold: Optional[float] = (
-        None  # 0 means no eviction
+        None  # [DO NOT SET THIS] feature score threshold for inference eviction policy. This is used for inference only, and will be set during publish time
     )
     inference_eviction_ttl_mins: Optional[int] = None  # 0 means no eviction
     feature_score_mapping: Optional[Dict[str, float]] = None  # feature score mapping

Original file line number	Diff line number	Diff line change
`@@ -237,7 +237,7 @@ class FeatureScoreBasedEvictionPolicy(VirtualTableEvictionPolicy):`
`237`	`237`	`0 # max number of inference ids per rank, default is training_id_keep_count`
`238`	`238`	`)`
`239`	`239`	`inference_eviction_feature_score_threshold: Optional[float] = (`
`240`		`- None # 0 means no eviction`
	`240`	`+ None # [DO NOT SET THIS] feature score threshold for inference eviction policy. This is used for inference only, and will be set during publish time`
`241`	`241`	`)`
`242`	`242`	`inference_eviction_ttl_mins: Optional[int] = None # 0 means no eviction`
`243`	`243`	`feature_score_mapping: Optional[Dict[str, float]] = None # feature score mapping`