The qwen4exp project has implemented recurrent state rollback support to enable effective Multi-Token Prediction (MTP) speculative decoding. This change allows the target state to move back by the number of rejected draft tokens, preventing unnecessary serialization of the entire recurrent state to host memory.
Previously, the `build_conv_state_at` function wrote only a single plane, causing rollbacks to restore an incomplete convolution history. The update now writes one snapshot per slot for both delta net QKV and PLE convolutions, ensuring accurate state restoration.
Benchmarks on Qwen3.8-Flash-Next UD-Q4_K_XL with MTP draft show decoding speeds of 183 tok/s on code and 144 tok/s on prose. This is a significant improvement over the previous fallback method, which achieved only 123 and 83 tok/s respectively.