Repository navigation
Qwen3.5/3.8 on CUDA: decode 37-73% faster on #836's engine, bits unchanged (on #836) - #978
Closed
BobClawblaw wants to merge 39 commits into
Closed
BobClawblaw wants to merge 39 commits into
BobClawblaw wants to merge 39 commits into
Commits
Commits on Oct 16, 2026
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
qwen3_5 cuda: the drafter's top-k over many blocks a row, the same ids and logits as tf_logprob_rows
committed- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed