Repository navigation

fabric: tensor parallelism across CUDA processes (bootstrap, NCCL, mailbox and RoCE collectives, plan link, fail-fast) (DeepSeek-V4.1 stack 3/14) - #636

Closed

jayleaton wants to merge 33 commits into

ashhart:mainfrom

jayleaton:stack/03-tp-fabric

Closed

fabric: tensor parallelism across CUDA processes (bootstrap, NCCL, mailbox and RoCE collectives, plan link, fail-fast) (DeepSeek-V4.1 stack 3/14)#636
jayleaton wants to merge 33 commits into
ashhart:mainfrom
jayleaton:stack/03-tp-fabric

Commits on Oct 15, 2026
Commits on Oct 16, 2026
Commits on Oct 17, 2026
- CUDA arena: a reuse whose address-map insertion fails keeps its range free
  
  jayleaton
  and
  claude
  committed
  3aeebb9
- Merge stack/02-graph-cache-arena (the #635 arena fix) into stack/03-tp-fabric
  
  jayleaton
  and
  claude
  committed
  46161ac
Commits on Oct 18, 2026
- test(tp): reject malformed runner commands before configuration
  
  authored and
  jayleaton
  committed
  88418d8
- fix(tp): parse runner commands before opening sessions
  
  authored and
  jayleaton
  committed
  4c69b0d
- test(tp): retain verbs load failures in HCA diagnostics
  
  authored and
  jayleaton
  committed
  4df28ce
- fix(tp): distinguish verbs load errors from absent HCA ports
  
  authored and
  jayleaton
  committed
  0a80da0
- Merge ashhart/main (1.1.0, 76aafcf2) into stack/02-graph-cache-arena
  
  jayleaton
  and
  claude
  committed
  cdde1a4
- CUDA launches: a not-ready launch returns NotReady unlogged, as Driver.check does
  
  jayleaton
  and
  claude
  committed
  adbdb8a
- Merge stack/02-graph-cache-arena (adbdb8ab, main 76aafcf2) into stack/03-tp-fabric
  
  jayleaton
  and
  claude
  committed
  3bd6a6a