Qwen3.8-27B serving line: full-precision adoption, captured block-draft speculation, small-M and state-stash tiers - #175

Open

LiangSu8899 wants to merge 44 commits into

feat/hf-kernels-structuresfrom

feat/qwen38-27b

Commits on Aug 14, 2026

adopt: dense projections quantize on adoption - the full-precision sibling of the expert-bank door
LiangSu8899
committed
56eeef5

Commits on Aug 15, 2026

kernels: sm_120 decode/prefill tiers with in-tree verification harnesses
LiangSu8899
committed
8b601d1
decode_loop: block-draft speculative member (DFlash/DSpark family), first assembly
LiangSu8899
committed
e2f4b16
decode_loop: the block-draft round rides the graph families
LiangSu8899
committed
96485cd
kernels: multi-row tier for the interleaved warp-split NVFP4 GEMM
LiangSu8899
committed
4b96408
adopt: small-M rows route to the native multi-row warp-split tier
LiangSu8899
committed
122b694
decode_loop: the draft forward rides the vendored fa2 and keeps bf16
LiangSu8899
committed
46467f6
gated_delta: per-row state stash - speculative rollback by selection
LiangSu8899
committed
6484484
structures: hub-first tiers, the wmma Gram swap, and the dspark member
LiangSu8899
committed
7791fa1
adopt: prefill slabs take the cooperative 256-tile tier
LiangSu8899
committed
917f202
structures: share activation quantization across sibling projections
LiangSu8899
committed
97eb499
decode_loop: sliced prompt pass for deep windows
LiangSu8899
committed
8fe5ec9
gated_delta_core: calibrated balance arm for the projection band
LiangSu8899
committed
8f23cec

Commits on Aug 16, 2026

Commits on Aug 17, 2026

Commits on Aug 18, 2026

Commits on Aug 19, 2026