2026-05 monthly summary focused on delivering Ascend-specific speculative decoding enhancements for vLLM on Ascend NPUs, plus robust test coverage and performance-oriented improvements. Delivered the AscendExtractHiddenStatesProposer to support the extract_hidden_states speculative decoding method on Ascend hardware by adapting the base proposer to ACL graphs and implementing Ascend-specific token ID preparation. Updated the model runner to support KV cache allocation and reshaping for cache_only_layers. Added end-to-end and unit tests to validate correctness and performance in the spec_decode workflow. This work expands hardware support, improves throughput and latency for large language model inference on Ascend, and demonstrates strong systems-level engineering across inference, data paths, and test automation.