diff --git a/gemma/common.h b/gemma/common.h index 0ca7e5c..663b2ca 100644 --- a/gemma/common.h +++ b/gemma/common.h @@ -36,7 +36,7 @@ ByteStorageT AllocateSizeof() { return hwy::AllocateAligned(sizeof(T)); } -constexpr size_t kPrefillBatchSize = 64; +constexpr size_t kPrefillBatchSize = 512; constexpr size_t kDecodeBatchSize = 1; constexpr size_t kBatchedQueryBatchSize = 16; constexpr size_t kMinAdjustedPrefillBatchSize =