Repository navigation

perf(qwen3_5 cuda): deliver the first token before drafting - #932

Closed

jrb-ai wants to merge 31 commits into

ashhart:mainfrom

jrb-ai:jon/pr836-first-token-handoff

Closed

perf(qwen3_5 cuda): deliver the first token before drafting#932
jrb-ai wants to merge 31 commits into
ashhart:mainfrom
jrb-ai:jon/pr836-first-token-handoff

Commits on Oct 14, 2026
Commits on Oct 15, 2026
- qwen3_5: attention chunks follow each row's own keys, not the forward's widest stream
  BobClawblaw
  committed
  8570e06
- qwen3_5: the DFlash2 config's divisors must be positive, and staging fits its selector rank
  BobClawblaw
  committed
  5e1c3fe
Commits on Oct 16, 2026
- perf(qwen3_5 cuda): deliver the first token before drafting
  jrb-ai
  committed
  d653cd5
Commits on Oct 17, 2026
Commits on Oct 18, 2026
- docs: clarify first-token replay clocks and startup receipt
  jrb-ai
  committed
  5248a33