feat(dflash): MoE 35B-A3B support + DDTree CUDA graph reuse#39

Open

dusterbloom wants to merge 11 commits intoLuce-Org:mainfrom

dusterbloom:feat/moe-35b-a3b

Commits on Apr 25, 2026

feat(dflash): accept qwen35moe arch in GGUF loader (MoE cycle 1)
dusterbloom
committed
feat(dflash): implement MoE FFN with expert routing + shared expert (cycle 2)
dusterbloom
committed
feat(dflash): parameterize target graph for 40-layer MoE + full forward test (cycle 3)
dusterbloom
committed
feat(dflash): parameterize draft model for 35B-A3B MoE + YaRN RoPE (cycle 4)
dusterbloom
committed
refactor(dflash): parameterize test code for multi-model support (cycle 5)
dusterbloom
committed
fix(dflash): reshape shared expert tensors for batched MoE FFN
dusterbloom
committed
perf(dflash): optimize MoE compute and graph allocator reuse
dusterbloom
committed
fix(dflash): correct YaRN attention scale for MoE draft model
dusterbloom
committed
perf(dflash): CUDA graph reuse + GPU argmax for MoE AR (64→144 tok/s)
dusterbloom
committed
perf(dflash): CUDA graph reuse for DDTree target verify
dusterbloom
committed

Commits on Apr 27, 2026

Merge remote-tracking branch 'origin/main' into feat/moe-35b-a3b
dusterbloom
committed