[DSA] Enable tilelang fp8_e4m3 KV cache on CUDA (raw layout) — verified on sm_121 - #36904

Closed

beastllama wants to merge 3 commits into

sgl-project:xinyuan/glm-5.3-flash-supportfrom

beastllama:fp8-kv-tilelang-cuda

Commits on Aug 28, 2026

[DSA] Enable tilelang fp8_e4m3 KV cache on CUDA (raw layout)
randomllama
committed
75a2241
test: port GB10-validated fp8 probe into pytest (one-hot exact, spread-within-budget, scrambled negative control)
randomllama
committed
06a1db7
Address CI + review: black formatting, CI registry for the fp8 test, reject tilelang-fp8 + DCP (raw writer lacks the DCP rank filter)
randomllama
committed
436a89b