glm5_next: support GLM-5.3-Flash (hybrid KDA + DSA, mHC, NVFP4 expert offload) - #270

Open

Cerynitius wants to merge 5 commits into

FlashML-org:mainfrom

Cerynitius:feat/glm5-flash

Commits on Aug 29, 2026

feat(glm5_next): support GLM-5.3-Flash (hybrid KDA + MLA/DSA, mHC, NVFP4 experts)
Cerynitius
committed
73fa396
feat(glm_moe_dsa): NoPE guards, DSA key-pooling, clamped-swiglu backend selection
Cerynitius
committed
9ccac9f
feat(cpu-moe): apply swiglu_limit in the generic activation epilogue
Cerynitius
committed
dfc69f4
docs(models): add GLM-5.3-Flash
Cerynitius
committed
04ad8df
feat(glm5_next): vision + video input (tower, dual-API intake, media-safe prefix caching)
Cerynitius
committed
857c167