P2: real engine-state feed replaces stale shadow counters for migration targeting

vLLM scheduler publishes real state (running/waiting, KV free, and the
max-in-progress-prefill signal /metrics lacks) to a tmpfs/redis store ~20Hz;
router reads it and avoids GIL-stall (mid-large-prefill) + KV-capacity-wall
targets, using real load over 30s-stale shadow counters. Components:
engine_state.py (canonical+reader), instrument_engine_state.py (scheduler
patch, file/redis writer), migration_target.py (scorer), proxy wiring
(--engine-state-uri, off=unchanged). All unit-tested without GPU; not yet
run live. See P2_ENGINE_STATE.md.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
2026-05-28 20:01:26 +08:00
parent 19191940e6
commit be948d32b8
5 changed files with 610 additions and 8 deletions

View File

@@ -111,6 +111,13 @@ class Settings:
v3_recent_mig_weight: float = 1.0 # how many "virtual requests" each
# recent migration counts as
# P2: real engine-state feed (replaces 30s-stale shadow counters for
# migration target selection). Empty = disabled (use shadow only).
engine_state_uri: str = "" # file:///dev/shm/... or redis://...
engine_state_period_ms: int = 50 # router poll period
es_big_prefill_threshold: int = 16000 # target mid-prefill >= this => avoid (GIL stall)
es_kv_wall_frac: float = 0.90 # target KV usage >= this => avoid (capacity wall)
# Direction B knob: LMetric fallback adds decode-token penalty to score.
# score = (pending_prefill + new + lmetric_decode_weight * ongoing_decode_tok) * num_req
# Empirical iter-time slope on H100 + Qwen3-30B-A3B: each decode token in
@@ -206,6 +213,9 @@ class InstanceState:
# recent-migration count over a sliding window, preventing back-to-back
# decisions from clustering on the same dst.
self.recent_mig_targeted_at: deque[float] = deque(maxlen=64)
# P2: latest real engine state (from the engine-state feed), or None
# when the feed is disabled/stale. Set by _engine_state_poll_loop.
self.real_state: dict | None = None
def estimate_cache_hit(self, token_ids: list[int] | None) -> int:
if not token_ids or len(token_ids) < BLOCK_SIZE:
@@ -672,20 +682,29 @@ def pick_instance_unified_v3(
now_mono = _time.monotonic()
cutoff = now_mono - SETTINGS.v3_recent_mig_window_s
def _real_load(inst):
# P2: prefer REAL engine state (running+waiting) over the proxy's
# 30s-stale shadow num_requests, when the engine-state feed is fresh.
rs = getattr(inst, "real_state", None)
if rs is not None:
return rs.get("num_running", 0) + rs.get("num_waiting", 0)
return inst.num_requests
def effective_load(inst):
# Drop expired entries lazily.
while inst.recent_mig_targeted_at and inst.recent_mig_targeted_at[0] < cutoff:
inst.recent_mig_targeted_at.popleft()
recent = len(inst.recent_mig_targeted_at)
return inst.num_requests + recent * SETTINGS.v3_recent_mig_weight
return _real_load(inst) + recent * SETTINGS.v3_recent_mig_weight
ph_load = _real_load(prefill_host)
threshold_loaded = max(1,
int(prefill_host.num_requests * SETTINGS.v3_target_load_ratio))
int(ph_load * SETTINGS.v3_target_load_ratio))
candidates = [
(i, inst) for i, inst in enumerate(instances)
if i != prefill_idx
and effective_load(inst) < threshold_loaded
and effective_load(inst) <= prefill_host.num_requests - SETTINGS.v3_min_load_gap
and effective_load(inst) <= ph_load - SETTINGS.v3_min_load_gap
]
if not candidates:
decision["v3_reason"] = (
@@ -700,11 +719,23 @@ def pick_instance_unified_v3(
# cache_hit DESC (more cache = less KV to transfer), then by effective_load
# (which includes recent-migration penalty), then by ongoing_tokens.
if SETTINGS.v3_prefer_cache_target:
decode_target_idx, decode_target = min(
candidates,
key=lambda x: (-x[1].estimate_cache_hit(token_ids),
effective_load(x[1]),
x[1].ongoing_tokens))
def _tgt_key(x):
# P2: avoid a target that is mid-large-prefill (holds the GIL,
# stalls the mooncake receiver_loop = the ~45% control-plane
# residual layer-wise can't fix) or near the KV capacity wall,
# before ranking by cache-richness and real load.
inst = x[1]
ch = inst.estimate_cache_hit(token_ids)
rs = getattr(inst, "real_state", None)
stalls = near_wall = 0
if rs is not None:
if int(rs.get("max_prefill_remaining", 0)) >= SETTINGS.es_big_prefill_threshold:
stalls = 1
f = rs.get("gpu_kv_used_frac", 0.0) or 0.0
if float(f) >= SETTINGS.es_kv_wall_frac:
near_wall = 1
return (stalls, near_wall, -ch, effective_load(inst), inst.ongoing_tokens)
decode_target_idx, decode_target = min(candidates, key=_tgt_key)
else:
decode_target_idx, decode_target = min(
candidates, key=lambda x: (effective_load(x[1]), x[1].ongoing_tokens))
@@ -857,6 +888,57 @@ async def _fetch_vllm_inflight(inst: "InstanceState") -> tuple[int, int] | None:
return running, waiting
def _engine_state_read_all(uri: str, max_age_s: float = 2.0) -> dict:
"""P2 reader (inlined; mirrors engine_state.StateReader). Returns
{engine_id: state}, dropping records older than max_age_s."""
now = _time.time()
out: dict = {}
try:
if uri.startswith("file://"):
import glob
d = uri[len("file://"):]
for p in glob.glob(os.path.join(d, "*.json")):
try:
s = json.load(open(p))
except Exception:
continue
if now - s.get("ts", 0) <= max_age_s:
out[s.get("engine_id", os.path.basename(p)[:-5])] = s
elif uri.startswith("redis://"):
import redis
r = redis.Redis.from_url(uri)
for k in r.scan_iter("engine_state:*"):
v = r.get(k)
if not v:
continue
s = json.loads(v)
if now - s.get("ts", 0) <= max_age_s:
out[s.get("engine_id")] = s
except Exception:
pass
return out
async def _engine_state_poll_loop():
"""P2: poll the engine-state feed and attach real_state to each instance.
Instance i is keyed engine_{i} (matches AGENTIC_WORKER_ID in the launcher).
"""
uri = SETTINGS.engine_state_uri
if not uri:
return
period = max(0.01, SETTINGS.engine_state_period_ms / 1000.0)
insts = combined_instances or (prefill_instances + decode_instances)
print(f"[engine-state] polling {uri} every {period*1000:.0f}ms for {len(insts)} instances")
while True:
try:
await asyncio.sleep(period)
except asyncio.CancelledError:
return
states = await asyncio.to_thread(_engine_state_read_all, uri)
for i, inst in enumerate(insts):
inst.real_state = states.get(f"engine_{i}")
async def _reconcile_loop():
"""Periodic shadow-state reconciliation against vLLM /metrics truth.
@@ -960,6 +1042,7 @@ async def lifespan(app: FastAPI):
_verify_vllm_patch()
reconcile_task = asyncio.create_task(_reconcile_loop())
engine_state_task = asyncio.create_task(_engine_state_poll_loop())
if global_args.combined:
is_pd_sep = False
@@ -1720,6 +1803,10 @@ def parse_args():
" penalised. 0 = original behavior; 0.01 is a reasonable start.")
p.add_argument("--overload-factor", type=float, default=2.0,
help="Break session affinity when instance load > factor * avg")
p.add_argument("--engine-state-uri", type=str, default="",
help="P2: real engine-state feed for migration target "
"selection (file:///dev/shm/... or redis://...). "
"Empty=disabled (shadow counters only).")
# The four flags below are accepted for bench.sh backward compatibility but
# have no effect after the PD-sep offload path was retired (REPORT §3.9,
# commits 4c583f2 / cc6e562). Removing them would break scripts/bench.sh and
@@ -1752,6 +1839,7 @@ if __name__ == "__main__":
global_args = parse_args()
SETTINGS.heavy_threshold = global_args.heavy_threshold
SETTINGS.overload_factor = global_args.overload_factor
SETTINGS.engine_state_uri = getattr(global_args, 'engine_state_uri', '') or ''
SETTINGS.max_offload_inflight = global_args.max_offload_inflight
SETTINGS.cache_gate_ratio = global_args.cache_gate_ratio
SETTINGS.decode_iteration_s = getattr(global_args, 'decode_iteration_s', 0.05)