leavelet 17cbfa31c9 E1 lane-stats instrumentation + 2a reserve-at-admission (default-off) ...

Two CP HiCache observability/admission additions, both behind default-off env flags (behavior-neutral until enabled).

E1 instrumentation (SGLANG_CP_HICACHE_LANE_STATS_S=<sec>, default 0):
- environ.py: flag. hiradix_cache.py: _cp_maybe_log_lane_stats() (rank-0 per-owner-lane L2 occupancy + imbalance + hot-prefix histogram) wired into check_hicache_events; per-cause drop counters in _warn_cp_hicache_fallback.
- Used for E1: verdict = lane imbalance is a non-issue (1.02-1.03 across full fill->evict->refill, zero host_reservation_failed).

2a reserve-at-admission (SGLANG_CP_HICACHE_RESERVE_ADMISSION, default off):
- hiradix_cache.py: cp_host_backup_admission_budget() (min-over-ranks available+evictable host tokens from the replicated radix snapshot) + _cp_host_evictable_tokens_by_rank.
- schedule_policy.py: PrefillAdder rem_l2_backup_tokens field + worst-lane footprint debit + budget_state gate (OTHER).
- scheduler.py: pass cp_size to PrefillAdder.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

2026-06-20 12:03:14 +00:00

..

sglang

E1 lane-stats instrumentation + 2a reserve-at-admission (default-off)

2026-06-20 12:03:14 +00:00

pyproject_cpu.toml

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

pyproject_npu.toml

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

pyproject_other.toml

chore: extract diffusion_common in python/pyproject_other.toml (#20803)

2026-03-19 10:39:16 +08:00

pyproject_xpu.toml

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

pyproject.toml

new pyprojectoml

2026-06-10 05:54:43 +08:00