Repository navigation

Add StartLux Instant 4B native worker (#160) - #164

Draft

hsliuustc0106 wants to merge 11 commits into

mainfrom

cursor/startlux-instant-4b-native-ba7c

Draft

Add StartLux Instant 4B native worker (#160)#164
hsliuustc0106 wants to merge 11 commits into
mainfrom
cursor/startlux-instant-4b-native-ba7c

Commits on Oct 24, 2026
- Add the StartLux Instant 4B native worker
  
  cursoragent
  and
  hsliuustc0106
  committed
  30e87f5
- Share the Instant tokenizer across requests.
  
  cursoragent
  and
  hsliuustc0106
  committed
  5bb25be
Commits on Oct 25, 2026
- Reserve Instant prompt scratch before listen.
  
  cursoragent
  and
  hsliuustc0106
  committed
  2b4a586
- Load Instant KV tiles once per GQA group.
  
  cursoragent
  and
  hsliuustc0106
  committed
  3ae18f5
- Fit Instant grouped KV load inside sm_89 shared memory.
  
  cursoragent
  and
  hsliuustc0106
  committed
  136ddf8
- Pack Instant prompt rows inside the high-water scratch.
  
  cursoragent
  and
  hsliuustc0106
  committed
  1b5b1ee
- Queue Instant token ids from pinned memory without a stream sync.
  
  cursoragent
  and
  hsliuustc0106
  committed
  d139b22
- Skip the Instant forward's per-launch CUDA error query.
  
  cursoragent
  and
  hsliuustc0106
  committed
  d3ba9d5
- Time Instant GEMM candidates once at load.
  
  cursoragent
  and
  hsliuustc0106
  committed
  6c268cd
- Quantize Instant projections to FP8 E4M3.
  
  cursoragent
  and
  hsliuustc0106
  committed
  725c86a
- Quantize Instant projections to symmetric int8.
  
  cursoragent
  and
  hsliuustc0106
  committed
  867ed78