perf(deepseek_v4): faster exact DeepSeek-V4 prefill (wsdpa fusion + scheduler/cache amortization) - #2568

Merged

jundot merged 6 commits into

jundot:mainfrom

jonathan308:pr/dsv4-wsdpa-prefill

Aug 10, 2026

Commits on Aug 10, 2026

perf(deepseek_v4): fuse windowed+pooled prefill attention for DeepSeek-V4
jonathan308
committed
c2f5583
fix(deepseek_v4): build native prefill kernels reliably
jonathan308
committed
d372749
perf(deepseek_v4): skip discarded lm_head projection in chunked prefill
jonathan308
committed
19827b9
perf(deepseek_v4): append pooled cache in place during prefill
jonathan308
committed
a77ab09
perf(deepseek_v4): fold pooled-ratio causal mask into indexer score kernel
jonathan308
committed
760fe94
test: restore prefill fixture compatibility

PhilipJohnBasile
authored and
jonathan308
committed
9423196