SourcesLMSYS
publisher · 1 source
LMSYS
1 source read from this publisher, newest first, with the methods each one describes.
Qwen3.8-Flash-Next: Day-0 Support in SGLang
technical report · vendor docs · 2026-08-26 · 19 methods
Gated ResidualQwen Sparse AttentionN-gram embedding lookup memoryHost-memory offload for N-gram embeddingsAsynchronous prefetching of offloaded N-gram embeddingsGated DeltaNet and Qwen Sparse Attention hybrid architectureReuse QSA index selection across speculative decoding stepsFour-stream Hyper-Connection Combine updateFusing activation, gating, and reduction into GEMM epiloguesHidden-dimension split Combine kernelIndexShare for multi-token prediction speculative decodingLow-rank gated Mix for Hyper-Connection residual streamsOffline weight reordering for local gate reductionOverlap pinned-host embedding gather with decoder computationQSA micro-block compression at ratio 4Shape-aware kernel dispatch for Hyper-ConnectionSparse pinned-host offloadSplit-K CuTe GEMMTop-512 block selection