docs: sync env template, migration notes and architecture for M2.5
This commit is contained in:
+7
-2
@@ -28,7 +28,7 @@ LLM__QWEN__1__TIMEOUT_S=120
|
||||
LLM_MAX_RETRIES=3 # 总尝试次数(含首次);或 LLM__RETRY__MAX_ATTEMPTS
|
||||
LLM_RETRY_BASE_DELAY=2.0 # 或 LLM__RETRY__BACKOFF_BASE_S
|
||||
LLM_RETRY_MAX_DELAY=30.0 # 或 LLM__RETRY__BACKOFF_MAX_S
|
||||
LLM_CIRCUIT_BREAKER_THRESHOLD=5 # 有效阈值自动取 max(此值, 并发×2);或 LLM__BREAKER__FAIL_THRESHOLD
|
||||
LLM_CIRCUIT_BREAKER_THRESHOLD=5 # 连续失败通道;有效阈值取 max(此值, 源级并发×2);或 LLM__BREAKER__FAIL_THRESHOLD
|
||||
LLM_CIRCUIT_BREAKER_COOLDOWN=60 # 或 LLM__BREAKER__COOLDOWN_S
|
||||
# LLM_TIMEOUT=120 # 源缺 TIMEOUT_S 时的缺省
|
||||
# LLM_TTFT_TIMEOUT=30 # 平铺看门狗缺省(成对生效)
|
||||
@@ -36,7 +36,12 @@ LLM_CIRCUIT_BREAKER_COOLDOWN=60 # 或 LLM__BREAKER__COOLDOWN_S
|
||||
# LLM__BREAKER__PROBE_TTL_S=240 # 缺省派生: max(2×最大源超时, cooldown, 最大源超时+5);显式值须 ≥ 最大源超时+5
|
||||
# LLM__BACKPRESSURE__STALL_WINDOW_S=300 # stall 双条件判死窗口;须 ≥ 最大源 TTFT
|
||||
# LLM__BACKPRESSURE__POLL_INTERVAL_S=0.05
|
||||
# LLM__SELECTOR=round_robin # round_robin(默认) | least_inflight
|
||||
# LLM__SELECTOR=health_aware # health_aware(默认,M2.5) | round_robin | least_inflight
|
||||
# ── M2.5 失败率熔断通道(可选,缺省即生产推荐值)──
|
||||
# LLM__BREAKER__MIN_CALLS=10 # 率通道最小样本(防低流量误判)
|
||||
# LLM__BREAKER__FAIL_RATE=0.6 # 窗口失败率阈值(429 不计入)
|
||||
# LLM__BREAKER__WINDOW_S=60 # 失败率窗口(双 30s 桶)
|
||||
# LLM__BREAKER__MAX_COOLDOWN_S=300 # 开路指数退避封顶(缺省 max(300, cooldown))
|
||||
# LLM__QUOTA_FULL=wait # wait(默认) | fail_fast
|
||||
|
||||
# ══ 装配选择(PGW_*)══
|
||||
|
||||
Reference in New Issue
Block a user