ggml adapter: SM120/Qwen3.6-35B target — binding-driven LLM decode windows, MoE catalog family, packless artifacts - #186

Open

LiangSu8899 wants to merge 10 commits into

feat/structures-ggml-adapterfrom

feat/ggml-sm120-qwen36

Commits on Aug 25, 2026

ggml adapter: add SM120/Qwen3.6-35B target (LLM decode window set)
LiangSu8899
committed
3be46d8
ggml adapter sm120: in-process weight repack for the FP4 regions
LiangSu8899
committed
f4ab9e1
catalog: map the sm120/Qwen3.6 target onto real structure families
LiangSu8899
committed
c8e1a8b
catalog: charter — structures express boundaries, never adjudicate
LiangSu8899
committed
87e6a30
ggml adapter sm120: window code consumes the binding (model facts unsoldered)
LiangSu8899
committed
83e828f
csrc: promote the M-rows f32out warpsplit GEMV and activation quantizer
LiangSu8899
committed
5303064
ggml adapter sm120: windows default on (zero-configuration safe tier)
LiangSu8899
committed
2954324
qualification: one-command gate runner for the sm120/Qwen3.6 target
LiangSu8899
committed
65f9fb2
docs: sm120 target invariants (PDL capability, template-M, speculative state discipline)
LiangSu8899
committed
3b77186
ggml adapter sm120: NVFP4-head artifact path (packless full tier)
LiangSu8899
committed
ba8a5b7