fix(oom): multimodal OOM fix with auto-profile#1273

Open

sufubao wants to merge 9 commits intomainfrom

Commits on Apr 16, 2026

fix(oom): layer 0 + layer 1 — visual/audio stress and semaphore tightening
sufubao
committed
feat(oom): phase 1 probe + log-only phase 2 target (auto-profile prep)
sufubao
committed
feat(oom): phase 3 + 4 auto-profile rebuild and validate loop
sufubao
committed
feat(oom): flip mem_fraction default to 1.0 and remove DISABLE_CHECK_MAX_LEN_INFER
sufubao
committed
fix(oom): default mem_fraction to 0.95 and fix startup assertion
sufubao
committed
fix(oom): shrink auto-profile probe size from gmbs*bmt to bmt+gmbs
sufubao
committed
fix(oom): include max_req_total_len in probe size floor
sufubao
committed
fix(oom): relax _check_mem_size assertion during probe phase
sufubao
committed
fix(oom): remove duplicate mem_fraction subtraction in mamba profile_size
sufubao
committed