Skip to content

The input (6165048 tokens) is longer than the model's context length (1048580 tokens(todo fix)). #4068

New issue
New issue
Closed
Closed
The input (6165048 tokens) is longer than the model's context length (1048580 tokens(todo fix)).#4068
Labels
bugSomething isn't workingdshduplicateThis issue or pull request already existsp1scope/wiring

Description

@ranxianglei
ranxianglei
opened on Oct 21, 2026

失败原因：The input (6165048 tokens) is longer than the model's context length (1048580 tokens(todo fix)).
失败原因：413: {"code":3,"message":"Request size limit exceeded","details":[]}

image.png

billion-context-log (5).txt

billion-context-log (1) (1).txt

Mirrored from ework issue #4068

Activity

  1. ranxianglei commented on Oct 21, 2026

    @ranxianglei
    ranxianglei
    on Oct 21, 2026
    OwnerAuthor

    ework

  2. ranxianglei commented on Oct 21, 2026

    @ranxianglei
    ranxianglei
    on Oct 21, 2026
    OwnerAuthor

    🤖 Powered by ework · qwen3.8-27b

    (ai: vllm-qwen/qwen3.8-27b)

    查重结论(先答):重复 — 与 #3011 同一症状 + 同一根因,按「同症状同 issue」规则并入。#3011 先到(2026-10-13)保留;本 issue(#4068,2026-10-21)后到 → 关闭为重复。新证据已写入 #3011 最新楼层。

    标签:bug / dsh / scope/wiring / p1 / duplicate。(p1 理由:owner 2026-10-19 明确把「fork 继承被永久放弃」归为 p1 类,且该克隆会话实际不可用。)

    一句话人话:你在 DSH 里克隆了一个长会话,克隆出的子会话没继承到父会话的压缩状态,于是每轮都把约 1300 万 token 的原始全史原样重发给上游,直接超窗报错——不是操作问题,是 bili 的一个已知缺陷。

    结构化八问

    1. 是什么问题:DSH 分支克隆出的子会话「零折叠」起步(没继承父会话已压缩的状态),每轮重发未压缩全量历史,首轮即超窗失败(context_length_exceeded / 413 Request size limit exceeded),该克隆会话不可用。
    2. 什么原因导致(因果分层):
    3. 影响范围与严重程度:触发条件 = dsh(或任意 plugin 车道)克隆一个含 thinking/图片的长会话;命中即该克隆会话完全不可用(每轮超窗死循环)。我已在当前 master(bili v0.1.202)核实 C6 门仍在 ⇒ 最新版仍会触发。严重度 p1。
    4. 怎么修(方向,非本 issue 交付):C6 目标设计已在 [Bug] Fork of DSH branch clone: inheritance permanently abandoned by 409 SNAPSHOT_UNAVAILABLE — zero-fold clone triggers context_length_exceeded 400 on first turn | [Bug] DSH 分支克隆的 fork 继承被 409 SNAPSHOT_UNAVAILABLE 永久放弃：克隆零折叠，首轮即 context_length_exceeded … #3011 floor 13 给出推荐——把 opaque/thinking/图片字节纳入逐块 identity hash + 把 publicSnapshotTextComparable 从硬门降级为优化标记(fail-safe:对不齐就诚实截短,永不误归属)。此项触碰消息身份逻辑,属 §7.4 人审域,待你拍板方向后进实现(记在 [Bug] Fork adoption server-side completion window: After the child session's first degraded request is committed, POST fork is permanently latched by 409 CHILD_CONFLICT, preventing retroactive adoption | [Bug] fork adoption 服务端完成窗口：子会话首条降级请求落地后，POST fork… #3025-C6)。
    5. 是否有回归问题:不涉及(本 issue 仅关闭+归档证据,不改代码)。
    6. 是否破坏缓存:不涉及。
    7. 是否改了设计方向:不涉及(修复沿用既有 fork 继承管线,见 [Bug] Fork adoption server-side completion window: After the child session's first degraded request is committed, POST fork is permanently latched by 409 CHILD_CONFLICT, preventing retroactive adoption | [Bug] fork adoption 服务端完成窗口：子会话首条降级请求落地后，POST fork… #3025-C6)。
    8. 是否引入新配置:不涉及。

    在途相关工作(避免重复劳动)

    #3024(PR,fork 409 改可重试 + raw-view 兜底)、#3039(迟补全)、#3951(代理车道匿名 fork lineage)、#3025(家族 checklist,C6 待拍板)、#3978(preflight 耗时/断连)、#3872(发送前尺寸守卫)。注意:#3024/#3039/#3951 都不解开 C6——thinking 父会话的克隆继承率目前仍 ≈0,只有 C6 落地才根治。你本地可用的临时缓解仍是给该车道设 compress.modelContextLimit(如 900000)让 bili 转发前先折到位。

    收工判定

    #4068 关闭为 #3011 的重复。对照:症状同为「dsh 零折叠克隆 → 超窗不可用」,根因同为 C6(opaque-content snapshot 拒绝)。本 issue 的新价值 = ①确认 C6 在 v0.1.200/202 仍复现(floor 11 是 0.1.197);②记录「恢复路径卡死」次生机制——两者均已并入 #3011。后续修复走 #3025-C6,等你拍板方向。

  3. ranxianglei
    added
    bugSomething isn't working
    dsh
    duplicateThis issue or pull request already exists
    p1
    scope/wiring
    on Oct 21, 2026
  4. ranxianglei
    added a commit that references this issue on Oct 21, 2026

    fix(#4079): egress oversize bars never assume a window below 1M tokens

    e17e8df
  5. Zhishulo
    added 2 commits that reference this issue on Oct 21, 2026

    fix: egress window fuse — stop provably-doomed bodies at the last sen…

    18be796

    fix(ranxianglei#3900): overflow 400s teach the calibration ring — k̂ …

    3d456dc
  6. yanhuojiyusheng
    added a commit that references this issue on Oct 21, 2026

    fix(ranxianglei#4111): dsh unattributed gate — core-text byte math, s…

    ed1273c

  7. ranxianglei commented on Oct 21, 2026

    @ranxianglei
    ranxianglei
    on Oct 21, 2026
    OwnerAuthor

    🤖 Powered by ework · qwen3.8-27b

    [bot] 🏷
    (ai: vllm-qwen/qwen3.8-27b)

    处置完成 — #4068 判定为 #3011 的重复(同症状:dsh 克隆零折叠 → 超窗不可用;同根因:C6 opaque-content snapshot 拒绝)。

    已做:

    修复方向沿用既有 fork 继承管线:把 thinking/图片字节纳入逐块 identity hash + 把 comparable 标记降级为优化位,记在 #3025-C6,等你拍板方向后进实现(#3024/#3039/#3951 都不解开 C6)。临时缓解:给该车道设 compress.modelContextLimit(如 900000)让 bili 转发前先折到位。

    按先来后到:#3011 保留,#4068 作重复合并。

  8. ranxianglei commented on Oct 21, 2026

    @ranxianglei
    ranxianglei
    on Oct 21, 2026
    OwnerAuthor

    关闭:两条通路均已堵上,随 v0.1.203 发版(#4122)——(1) relay egress window fuse(#4090,"What this fixes" 首条即本条的 6,165,048-token 案例):必超窗请求在最后一处发送点本地 413 egress_window_fuse;(2) gate-none oversize arm(#4090,明确标注 "the #4068 channel specifically"):unattributed DIRECT 裸发通道的超窗体现在本地拒绝;(3) 校准闭环从拒绝学习(#4093)。如 0.1.203+ 仍复现请重开。

  9. ranxianglei
    closed this as completedon Oct 21, 2026
  10. ranxianglei
    added a commit that references this issue on Nov 7, 2026

    fix: #5613 dsh gate safety net — calibrated over-window estimate refu…

    e920fb0
  11. Louie0806
    added a commit that references this issue on Nov 8, 2026

    fix: ranxianglei#5898 review batch — 429-hold saturation warn + scope…

    81619f3
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingdshduplicateThis issue or pull request already existsp1scope/wiring

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions