• Joined on 2026-07-08
iomgaa pushed to main at iomgaa/PolyGateway 2026-09-14 15:05:05 +08:00
b0232b7466 merge: release 1.3.8 net generation time and mid-stream stall hedging
cf6e3fa008 chore: release 1.3.8
a2a00c0ef2 test: sync PG telemetry integration helpers with 1.3.8 stall columns
8489b8ea50 docs: resolve branch review note (final-state comment for content_event wiring)
ca8743bf33 docs: document net generation time and mid-stream stall hedging
Compare 13 commits »
iomgaa created repository iomgaa/RivalBits 2026-09-14 12:15:28 +08:00
iomgaa opened issue iomgaa/PolyGateway#28 2026-09-14 10:42:33 +08:00
P2:首 token 到后中段沉默,对冲永不触发
iomgaa opened issue iomgaa/PolyGateway#27 2026-09-14 10:42:32 +08:00
P1:generation_ms 把流式中段沉默计入墙钟,裸生成时间名不副实
iomgaa closed issue iomgaa/PolyGateway#24 2026-09-11 03:21:48 +08:00
长尾延迟没有对冲机制,TIMEOUT_S 一个旋钮调不出好取值
iomgaa commented on issue iomgaa/PolyGateway#24 2026-09-11 03:21:47 +08:00
长尾延迟没有对冲机制,TIMEOUT_S 一个旋钮调不出好取值

已在 1.3.7 解决(opt-in,默认关闭):

  • {SCOPE}__HEDGE__AFTER_S 开启 chat 单路并发对冲:阈值到点向另一个等价源并发第二请求,谁先回用谁,输家取消收口。触发判据分流式(首 token 未至,误杀不了慢生成)与…
iomgaa pushed to main at iomgaa/PolyGateway 2026-09-11 03:19:46 +08:00
a28d94451e docs: record 1.3.7 release gate results
a04d87e8d7 merge: release 1.3.7 hedged requests and bare generation time
c60011b034 chore: release 1.3.7
9f7d407120 fix: skip hedge dispatch when primary finishes during admission
0572611af7 docs: document hedged requests and bare generation time
Compare 9 commits »
iomgaa pushed tag v1.3.7 to iomgaa/PolyGateway 2026-09-11 03:19:46 +08:00
iomgaa closed issue iomgaa/PolyGateway#22 2026-09-10 23:36:52 +08:00
单源 + 上游 429 无 Retry-After 时,一次调用挂满整个 stall window(生产配置下 >20 分钟)
iomgaa commented on issue iomgaa/PolyGateway#22 2026-09-10 23:36:51 +08:00
单源 + 上游 429 无 Retry-After 时,一次调用挂满整个 stall window(生产配置下 >20 分钟)

已在 1.3.6 解决( opt-in ):

  • 新增可选整体调用期限 {SCOPE}__CALL_DEADLINE_S 或 per-call call_deadline_s=(缺省 None 不启用,不配置即逐字保持 1.3.5 语义)。期限覆盖缓存读、准入排队、退避…
iomgaa commented on issue iomgaa/PolyGateway#24 2026-09-10 23:34:33 +08:00
长尾延迟没有对冲机制,TIMEOUT_S 一个旋钮调不出好取值

还需要注意前端需要返回真实的生成时间,所以可能涉及到添加多一个额外的参数,这个参数是排除了重试以及等待时间的裸生成时间,并且应该取对冲请求中快的那个。这样子就可以获取真实的请求时间,而不包含因为llm…

iomgaa pushed to main at iomgaa/PolyGateway 2026-09-10 23:31:43 +08:00
166b2865d0 docs: record 1.3.6 release gate results and matrix exemption
b0ab39edbf docs: scope the slow release gate by diff intersection
a07a6b096b merge: release 1.3.6 call deadline and cancellation settlement
53d2f089e1 chore: release 1.3.6
5e2c35a812 test: pin the expiry window and close settlement review gaps
Compare 11 commits »
iomgaa pushed tag v1.3.6 to iomgaa/PolyGateway 2026-09-10 23:31:43 +08:00
iomgaa created branch docs/1.3.5-release-evidence in iomgaa/PolyGateway 2026-09-10 02:02:27 +08:00
iomgaa pushed to docs/1.3.5-release-evidence at iomgaa/PolyGateway 2026-09-10 02:02:27 +08:00
d2455e8fd4 docs: record 1.3.5 release completion and external verification
iomgaa closed issue iomgaa/PolyGateway#23 2026-09-10 01:59:22 +08:00
LLMResponse 答不出「这次调用总共等了多久、试了几次」
iomgaa commented on issue iomgaa/PolyGateway#23 2026-09-10 01:59:22 +08:00
LLMResponse 答不出「这次调用总共等了多久、试了几次」

本 issue 由 1.3.5 关闭(main ab00aa44576bd4cd47d0680d107d5191679dc1ee,tag v1.3.5,已上传 registry)。

治理单位从「一次尝试」补齐到「一次逻辑调用」:新导出 frozen dataclass `polygateway.CallSta…

iomgaa commented on issue iomgaa/PolyGateway#19 2026-09-10 01:59:21 +08:00
遥测:上游状态码被压平、三分之二失败行没有归因列、21 条错误详情为空

本 issue 由 1.3.5 关闭(main ab00aa44576bd4cd47d0680d107d5191679dc1ee,tag v1.3.5,已上传 registry)。

归因缺失的三条本体已修:上游状态码不再被压平(中转把 529 改写成 503 就记 503,**不猜回…