diff --git a/.claude/scheduled_tasks.lock b/.claude/scheduled_tasks.lock new file mode 100644 index 0000000..83b14a8 --- /dev/null +++ b/.claude/scheduled_tasks.lock @@ -0,0 +1 @@ +{"sessionId":"45eccd49-4ce6-4a07-af1c-9bff461d40c9","pid":1030431,"procStart":"52754552","acquiredAt":1779633135793} \ No newline at end of file diff --git a/docs/01_fix_prefill_overflow.md b/docs/01_fix_prefill_overflow.md new file mode 100644 index 0000000..aa4396e --- /dev/null +++ b/docs/01_fix_prefill_overflow.md @@ -0,0 +1,10 @@ +# Chantier 1 — prefill multi-ubatch (cli hang pp512+) + +## Symptôme +cli prompt court OK; gros prefill pend. bench pp512 OK (1 ubatch). unaccounted=2^44 = cosmétique. +Hyp: prompt > n_ubatch → 2e ubatch prefill GDN → boucle handoff état S. + +## Étape 1: trouver le seuil + +## Conclusion: NON-BUG cote moteur +bench pp2048 (4 ubatches @b512) = 46.2 t/s OK. Multi-ubatch sain. cli court=8tok gen OK; cli ~160tok pend = bug cli interactif/reasoning, PAS lib. Kazeia = lib JNI, pas cli. unaccounted 2^44 = cosmetique breakdown. => chantier 1 CLOS, aucun fix moteur. Bench est le harness de ref. cli a eviter prod.