TokenProcessorConfig was added back to kvcache.Config (indexer level) at #444, but the scheduler never sets IndexerConfig.TokenProcessorConfig when building the indexer. It only sets the top-level TokenProcessorConfig (used to build the tokenProcessor), leaving indexer.blockSize() to fall back to the default of 16.
This causes a mismatch in GetPodScores: ComputeBlockExtraFeatures uses indexer.blockSize() (16) while tokenProcessor.chunkTokens uses the actual configured block size (e.g. 64), producing different counts for extraFeatures and token chunks.
The fix is either to propagate TokenProcessorConfig from the scheduler into IndexerConfig, or have indexer.blockSize() derive the block size from the tokenProcessor directly.
TokenProcessorConfigwas added back tokvcache.Config(indexer level) at #444, but the scheduler never setsIndexerConfig.TokenProcessorConfigwhen building the indexer. It only sets the top-levelTokenProcessorConfig(used to build the tokenProcessor), leavingindexer.blockSize()to fall back to the default of 16.This causes a mismatch in
GetPodScores:ComputeBlockExtraFeaturesusesindexer.blockSize()(16) whiletokenProcessor.chunkTokensuses the actual configured block size (e.g. 64), producing different counts for extraFeatures and token chunks.The fix is either to propagate
TokenProcessorConfigfrom the scheduler intoIndexerConfig, or haveindexer.blockSize()derive the block size from the tokenProcessor directly.