Key Feature Changes

Vision SDK

  • N/A

Index SDK

  • Enhanced standard-state ILFlat algorithm
  • IVFFlat retrieval algorithm supported
  • IVFSP algorithm's Add interface performance optimized for small-batch scenarios

RAG SDK

  • Smart markdown document parsing supported
  • Reference design for knowledge QA applications provided
  • Embedding/Reranker model accelerated

Rec SDK

  • Performance optimized for dense_to_jagged, jagged_to_padded_dense, permute_2d_sparse_data, and asynchronous_complete_cumsum operators
  • Gradient accumulation supported in the Torch scenario
  • PyTorch 2.7 adaptation supported in the Torch scenario
  • DP parallel training supported in the Torch scenario
  • pagedHSTU inference enabled for HSTU, supporting different QKV lengths, head_num up to 16, and custom mask