fix(ali): stop injecting top_p into requests that omit it by ENCHIGO · Pull Request #6674 · QuantumNous/new-api

ENCHIGO · 2026-08-05T16:55:33Z

⚠️ 提交说明 / PR Notice

Important

请提供人工撰写的简洁摘要，避免直接粘贴未经整理的 AI 输出。

📝 变更描述 / Description

requestOpenAI2Ali 需要把 top_p 收进开区间，因为 DashScope 不接受 0 和 1 这两个边界值。但它是用 lo.FromPtrOr(request.TopP, 0) 读取指针的，这样**"没传 top_p"和"传了 0"就无法区分**——两者都会落进 topP <= 0 分支，被注入 0.001。

这带来两个后果：

部分模型直接 400。 0.001 / 0.999 都是三位小数，平台上部分模型只接受两位，会以 top_p参数非法：限制小数点[2]位 拒绝整个请求。实测 ZHIPU/GLM-5.2、xiaomi/mimo-v2.5-pro 在客户端不传 top_p 时 100% 失败。
不校验小数位的模型被静默改变采样行为。 请求不报错，但 top_p 被设成 0.001，等价于近乎贪婪解码，取代了模型自己的默认值。调用方无从察觉。

改法是把边界收敛限定在"调用方确实传了 top_p"的情况，并把 clamp 值改成两位小数：

request.TopP == nil 时原样返回，交由上游使用模型默认值 → 修复第 2 点
显式传值时 clamp 到 0.01 / 0.99 → 修复第 1 点

ThinkingBudget 的处理逻辑未改动。

🚀 变更类型 / Type of change

🐛 Bug 修复 (Bug fix) - 请关联对应 Issue，避免将设计取舍、理解偏差或预期不一致直接归类为 bug
✨ 新功能 (New feature) - 重大特性建议先通过 Issue 沟通
⚡ 性能优化 / 重构 (Refactor)
📝 文档更新 (Documentation)

🔗 关联任务 / Related Issue

Closes Ali 渠道对未传 top_p 的请求强制注入 0.001，导致百炼部分模型 400 且静默改变采样行为 #6671

✅ 提交前检查项 / Checklist

人工确认: 我已亲自整理并撰写此描述，没有直接粘贴未经处理的 AI 输出。
- 说明：本 PR 由 AI 协助完成，描述与代码均经我逐行复核后提交。为如实告知，此项不勾选。
非重复提交: 我已搜索现有的 Issues 与 PRs，确认不是重复提交。
Bug fix 说明: 已提交并关联对应 Issue Ali 渠道对未传 top_p 的请求强制注入 0.001，导致百炼部分模型 400 且静默改变采样行为 #6671；该问题为网关替调用方注入了未传的参数，属于实现缺陷而非设计取舍。
变更理解: 影响范围仅限 Ali 渠道（type=17）的 chat 请求转换。显式传入的 top_p 行为只在边界值上由三位小数变为两位小数；未传 top_p 的请求改为不再携带该字段，由上游使用模型默认值。
范围聚焦: 仅改动 relay/channel/ali/text.go 与新增对应单元测试，无其他无关改动。
本地验证: 已运行下方列出的测试与构建。
安全合规: 无任何凭据或敏感信息；沿用现有 testify 表驱动测试风格。

📸 运行证明 / Proof of Work

新增 relay/channel/ali/text_test.go，覆盖 6 种输入：

输入 `top_p`	期望输出	说明
未传（`nil`）	`nil`	不注入，回归本 PR 修复的核心问题
`0.8`	`0.8`	正常值原样通过
`1.0`	`0.99`	上边界，两位小数
`1.5`	`0.99`	越界，两位小数
`0.0`	`0.01`	下边界，两位小数
`-0.3`	`0.01`	越界，两位小数

$ go test ./relay/channel/ali/...
ok      github.com/QuantumNous/new-api/relay/channel/ali 0.833s

$ go vet ./relay/channel/ali/...
（无输出）

$ go build ./relay/... ./controller/... ./service/... ./model/... ./middleware/...
（无输出，退出码 0）

另在生产环境验证过等效行为：通过渠道 param_override 条件删除被注入的 top_p 后，ZHIPU/GLM-5.2 与 xiaomi/mimo-v2.5-pro 立即恢复正常，同渠道其余 11 个模型不受影响。

Summary by CodeRabbit

Bug Fixes
- Improved handling of the optional top_p setting.
- Preserved omitted values instead of applying an unintended default.
- Ensured out-of-range values are safely constrained between 0.01 and 0.99.
Tests
- Added coverage for omitted, valid, boundary, zero, negative, and excessive values.

DashScope rejects top_p at the 0 and 1 boundaries, so requestOpenAI2Ali clamped the value into the open interval. It read the pointer through lo.FromPtrOr(request.TopP, 0), so a request that omitted top_p was indistinguishable from one that sent 0 and got 0.001 injected. That had two effects. Models on the platform that validate two decimals rejected the request with "top_p参数非法：限制小数点[2]位". Models that do not validate it silently ran at near-greedy decoding instead of their own default. Only clamp when the caller actually sent top_p, and clamp to two decimals.

coderabbitai · 2026-08-05T16:56:09Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 4644556a-a9e8-4a3e-800c-2f4f1e207494

📥 Commits

Reviewing files that changed from the base of the PR and between 0ab0202 and b14363f.

📒 Files selected for processing (2)

relay/channel/ali/text.go
relay/channel/ali/text_test.go

Walkthrough

The Ali request conversion now preserves omitted top_p values and clamps explicit out-of-range values to 0.01 or 0.99. Table-driven tests cover the supported input cases.

Changes

Ali top_p handling

Layer / File(s)	Summary
Preserve and clamp top_p values `relay/channel/ali/text.go`, `relay/channel/ali/text_test.go`	`requestOpenAI2Ali` no longer injects `top_p` when it is omitted. Explicit values at or below `0` are clamped to `0.01`, and values at or above `1` are clamped to `0.99`. Tests cover omission, valid values, boundaries, and out-of-range values.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested reviewers: calcium-ion

Poem

A rabbit checks the sampler’s gate,
Leaving missing values to their fate.
Zero hops to one,
One bounds to ninety-nine,
Two neat decimals make requests behave.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the main fix: Ali requests no longer inject top_p when callers omit it.
Linked Issues check	✅ Passed	The implementation satisfies `#6671` by preserving nil top_p and clamping explicit values to 0.01–0.99, with tests for the required cases.
Out of Scope Changes check	✅ Passed	All changes are related to `#6671`; the added table-driven tests directly verify the top_p conversion behavior.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

* fix(relay): set Request.GetBody so the HTTP/2 transport can transparently retry after an upstream stream reset (QuantumNous#6249) * fix(relay): set Request.GetBody so the HTTP/2 transport can transparently retry after an upstream stream reset The outbound request body is a type-erased io.Reader over BodyStorage, so net/http cannot derive Request.GetBody (it only does so for *bytes.Reader, *bytes.Buffer and *strings.Reader). With GetBody nil, the HTTP/2 transport cannot transparently retry a request once the body has been written and the upstream resets the stream with a retryable error (REFUSED_STREAM, or a connection-level GOAWAY); the relay request then fails with: http2: Transport: cannot retry err [...] after Request.Body was written; define Request.GetBody to avoid this error This affects every relay path that goes through DoApiRequest (chat, claude, gemini, responses, embedding, image, rerank). BodyStorage (memory and disk) already implements io.Seeker, so replay support only needed wiring: - NewOutboundJSONBody additionally returns a getBody that rewinds the storage and hands out a fresh non-closing reader. The transport only calls GetBody after the previous attempt's body has been abandoned, so the rewind cannot race an in-flight read. - RelayInfo carries it in the new UpstreamRequestGetBody field, set alongside UpstreamRequestBodySize by the handlers that build storage-backed bodies. - applyUpstreamGetBody (symmetric with applyUpstreamContentLength) wires it into DoApiRequest/DoFormRequest/DoTaskApiRequest, only when req.GetBody is still nil. Also remove the hand-rolled GetBody override in DoTaskApiRequest: it returned the same already-consumed reader, so any transport-level replay would have silently sent an empty body, and it clobbered the correct snapshot-based GetBody that net/http derives from the *bytes.Reader bodies the task adaptors pass in. For non-replayable bodies GetBody now stays nil, so a retry fails loudly instead of corrupting the request. Covered by unit tests plus an end-to-end raw-frame HTTP/2 test that resets the first stream with REFUSED_STREAM after the body is written and asserts the transport transparently retries with the complete body. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(relay): hand out independent readers from GetBody (address review) Per the http.Request.GetBody contract ("returns a new copy of Body"), each call must yield a reader with its own cursor. The previous implementation rewound and reused the shared BodyStorage, so two consecutive GetBody readers would interfere with each other, and a replay could disturb the primary body's offset under extreme transport timing (e.g. attempt N's body write not yet fully abandoned when the transport builds attempt N+1). Instead of snapshotting the payload (an extra copy), add BodyStorage.NewReader, which returns an independent zero-copy reader: - memory mode: a fresh bytes.Reader over the same immutable backing array; - disk mode: a separate file descriptor over the cache file, so the transport closing a replayed body only closes that descriptor. NewOutboundJSONBody's getBody now simply hands out storage.NewReader, and once the handler releases the storage, GetBody fails with ErrStorageClosed instead of replaying stale data. Tests: interleaved reads across two replay readers and the primary body each observe exactly their own byte stream, for both the memory and the disk-backed storage; the existing GetBody and HTTP/2 retry suites still pass (h2 e2e tests flake-free with -count=20). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(relay): bind replayable metadata on pass-through requests * fix(relay): reset upstream body metadata between channels * test(relay): cover replay across retries and channel attempts * fix(relay): stop following upstream redirects --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * refactor(relay): move replay metadata onto request bodies * Merge commit from fork * feat(channels): refine fetched model categorization (QuantumNous#6632) * feat(channels): refine fetched model categorization * fix: channel category * fix: hy3 category * fix: test Claude/Gemini endpoints with native request format (QuantumNous#6698) * feat(rate-limit): add user critical rate limit middleware for access token and aff transfer routes * fix: 修复兑换码额度精度损失 (QuantumNous#6685) * fix: 修复兑换码额度精度损失(QuantumNous#6680) * fix(redemption): guard update data integrity * CI: enhance release synchronization workflow with optional file syncing * fix(ali): stop injecting top_p into requests that omit it (QuantumNous#6674) * fix(channels): classify Qwen TTS models correctly (QuantumNous#6711) * feat(channels): add auto-disable-only channel test mode (QuantumNous#6728) * perf(web): debounce server and large-list searches (QuantumNous#6727) * fix: record reasoning effort consistently in usage logs (QuantumNous#6641) * feat(relay): expose user and group context to parameter overrides (QuantumNous#6534) * fix(ollama): preserve reasoning and tool-call context (QuantumNous#6605) * fix: backend length validation (QuantumNous#5548) * feat(billing): highlight matched conditional multipliers in logs (QuantumNous#6561) * feat(billing): highlight matched conditional multipliers in usage logs * fix(billing): make request rule tracing stable and type-safe * fix(web): require confirmation before rotating access token (QuantumNous#6749) --------- Co-authored-by: Lucas <hepo.lucas@gmail.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: CaIon <i@caion.me> Co-authored-by: RedwindA <128586631+RedwindA@users.noreply.github.com> Co-authored-by: Seefs <40468931+seefs001@users.noreply.github.com> Co-authored-by: lihu-001 <lihu9048@gmail.com> Co-authored-by: ENCHIGO <38551565+ENCHIGO@users.noreply.github.com>

…s#6674)

Brings the fork up to upstream v1.0.0-rc.25 (47 commits, 233 files). Patches: - drop 6674-ali-top-p-passthrough (merged upstream as 2399de9) - drop 6573-custom-oauth-binding-status (superseded by upstream QuantumNous#6818, which aligned provider_id/provider_user_id in the frontend) - the remaining 12 still apply cleanly, in order, on rc.25 Conflict resolutions: - backend: took upstream's logic and re-applied our English log/error strings (atomic top-up settlement, Midjourney refund via service.RefundMidjourneyQuota, rune-based console length validation, common.QuotaFromFloat/QuotaFromDecimalStrict) - controller/channel-test.go: took upstream's worker-pool rewrite and re-applied ElevenLabs support plus our English strings - relaykit oai_chat request conversion: kept our web_search server-tool mapping alongside upstream's parameterless-tool handling; unioned both sides of the new test file - locales: 3-way union, 111 new upstream keys translated to pt-BR Frontend now runs on upstream's Vitest setup (jsdom): - converted three node:test files to Vitest - dropped the per-file happy-dom bootstrap that shadowed jsdom - test-setup installs an in-memory Storage because Bun's built-in localStorage throws and shadows jsdom's Also documents why a bare `patch --dry-run` reports success on macOS for an already-merged patch, which is what hid QuantumNous#6674. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CyvTzK8hTXFkGc7sWGg9ma

seefs001 approved these changes Aug 8, 2026

View reviewed changes

seefs001 merged commit 2399de9 into QuantumNous:main Aug 8, 2026
2 checks passed

0401lucky pushed a commit to 0401lucky/new-api that referenced this pull request Aug 16, 2026

fix(ali): stop injecting top_p into requests that omit it (QuantumNou…

a169feb

…s#6674)

DayFliggy pushed a commit to DayFliggy/Ren2Hub that referenced this pull request Aug 17, 2026

fix(ali): stop injecting top_p into requests that omit it (QuantumNou…

4b2ae01

…s#6674)

330079598 pushed a commit to 330079598/new-api that referenced this pull request Aug 19, 2026

fix(ali): stop injecting top_p into requests that omit it (QuantumNou…

e4eeb88

…s#6674)

Navigation Menu

Uh oh!

fix(ali): stop injecting top_p into requests that omit it - #6674

fix(ali): stop injecting top_p into requests that omit it#6674
seefs001 merged 1 commit into
QuantumNous:mainfrom
ENCHIGO:fix/ali-top-p-injection

ENCHIGO commented Aug 5, 2026 •

edited by coderabbitai Bot

Loading

Uh oh!

coderabbitai Bot commented Aug 5, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

ENCHIGO commented Aug 5, 2026 • edited by coderabbitai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

⚠️ 提交说明 / PR Notice

📝 变更描述 / Description

🚀 变更类型 / Type of change

🔗 关联任务 / Related Issue

✅ 提交前检查项 / Checklist

📸 运行证明 / Proof of Work

Summary by CodeRabbit

Uh oh!

coderabbitai Bot commented Aug 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ENCHIGO commented Aug 5, 2026 •

edited by coderabbitai Bot

Loading

coderabbitai Bot commented Aug 5, 2026 •

edited

Loading