From bc45753f771ffc509dfc24b58d0640e2eecdcd23 Mon Sep 17 00:00:00 2001 From: caoqianming Date: Mon, 3 Aug 2026 14:36:50 +0800 Subject: [PATCH] feat(models): unify thinking controls --- CHANGELOG.md | 4 ++ DESIGN.md | 2 + PROGRESS.md | 3 +- config/models/deepseek_v4.yaml | 17 +++-- config/models/doubao.yaml | 15 +++-- config/models/glm.yaml | 10 +-- config/models/local.yaml | 8 ++- config/models/unifyllm.yaml | 21 ++++-- core/__init__.py | 2 +- core/capabilities.py | 30 +++++++-- core/llm.py | 20 +++--- core/llm_params.py | 30 +++++++++ core/probe.py | 20 +++--- tests/test_llm_kwargs.py | 115 +++++++++++++++++++++++++++++++++ web/routers/models.py | 2 +- 15 files changed, 241 insertions(+), 58 deletions(-) create mode 100644 core/llm_params.py create mode 100644 tests/test_llm_kwargs.py diff --git a/CHANGELOG.md b/CHANGELOG.md index 22db6b2..5389de3 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -5,6 +5,10 @@ > 所以不是每个版本号都有条目。条目格式 `## <版本> — <日期>`,新条目加在最上面。 > 工程口径的完整记录见 `PROGRESS.md` / git log。 +## 0.60.26 — 2026-08-03 + +- 默认 DeepSeek Flash 已使用官方 0731 API 升级，继续展示思考过程并显式控制推理强度；不同模型的思考开关不再依赖服务端默认值，切换模型时行为更稳定。 + ## 0.60.25 — 2026-08-03 - HTML 产物现在可以在对话中直接显示，进入可视区域后自动加载交互内容，也可放大到弹窗查看；HTML 与 Markdown 预览均可在渲染结果和源文件之间切换，常见 HTTPS 图表、地图及网页资源可正常加载。 diff --git a/DESIGN.md b/DESIGN.md index fc461eb..8242f33 100644 --- a/DESIGN.md +++ b/DESIGN.md @@ -108,6 +108,8 @@ Session = 消息列表,ORM 直写 PG `messages`(append-only,jsonb 存 LiteLLM 默认 `deepseek_v4.flash`;复杂 bug / 终稿升 pro + reasoning_effort=max;fallback 手动切 Claude。成本量级:修 bug flash ~$0.01 / 完整申报书 flash ~$0.30(pro-max ~$1.5,Opus ~$10+)。99% 任务 flash 够用。 +模型思考参数由 profile 统一表达：`thinking_enabled` 只表示开关，`thinking_transport` 只表示已验证的传输协议，`reasoning_effort` 只表示开启后的推理强度；`core/llm_params.py` 是请求构造唯一入口。DeepSeek、GLM、方舟当前共享 `extra_body` 协议，未验证网关明确用 `none`、不猜参数协议，主循环不再按 family 分支。`/v1/models` 只返回语义明确的 `thinking_enabled`。 + --- ## 5. 设计哲学 diff --git a/PROGRESS.md b/PROGRESS.md index 54100cd..8309eb9 100644 --- a/PROGRESS.md +++ b/PROGRESS.md @@ -2,7 +2,7 @@ > 配合 `DESIGN.md`。本文件只记 phase 状态、决策偏差、文件量、下一步。每条 1-2 句:做了啥 + 关键判断;细节查 `git log` / `git diff` / `DESIGN §7.9`。 -最后更新:2026-08-03(交互式 HTML 预览 + 对话内嵌,bump 0.60.25) +最后更新:2026-08-03(DeepSeek Flash-0731 + thinking 参数统一,bump 0.60.26) --- @@ -23,6 +23,7 @@ ### 2026-08-03 +- **08-03 / 0.60.26 / DeepSeek Flash-0731 + thinking 参数统一**:默认 `deepseek-v4-flash` 无需换模型 ID 即接入官方 0731 后训练升级；Flash 改为显式开启 thinking 并透传 `reasoning_effort=high`，校准当前常规时段 token 成本但保留 8K 稳定输出预算。模型档案统一用 `thinking_enabled`(开关)+`thinking_transport`(协议)+`reasoning_effort`(强度)，DeepSeek/GLM/方舟共用纯函数请求构造，移除 family 分支与旧 `thinking_mode` 字段；方舟保持既有思考开启，GLM 保持生产验证过的显式关闭，未验证网关标记 `none`。`/v1/models` 同步只返回新字段；445 项 unittest 全绿(17 skip)，Ruff 与 diff 检查通过；未连生产 DB、未发真实模型请求，无 schema/migration/依赖变化。 - **08-03 / 0.60.25 / 交互式 HTML 预览 + 对话内嵌**:文件预览将 HTML 从普通源码提升为可切换“预览 / 源文件”的 sandbox iframe，允许脚本与 HTTPS CDN/接口但保持 opaque origin，禁止宿主权限、表单和顶层跳转；助手最终答复中的 HTML 产物改为进入可视区才加载的内嵌卡片，并可放大复用完整预览，Markdown 同步补源文件切换。Node 14 项、Python 27 项、JavaScript 语法及 diff 检查通过；当前环境无可用浏览器实例，真实页面点击/截图留部署后冒烟；无 schema、migration、HTTP API 或依赖变化。 - **08-03 / 0.60.24 / Web Mermaid 直出 + Markdown 围栏容错**:模型偶发用同长度围栏嵌套 Markdown/Mermaid 示例，CommonMark 会把后续正文吞进未闭合代码块；新增仅针对该明确形态的前后端确定性修复，提示词统一要求外层使用更长异类围栏，历史上下文加载时同样修正且不批量回写生产数据。聊天页本地 vendoring Mermaid 11.16.0，仅在助手文字段定稿后顺序渲染 `language-mermaid`，采用 strict 安全级别、文本/边数上限，语法错误或组件不可用时保留源码并提示；真实 Edge 冒烟确认中文流程图与 XYChart 柱线组合图可生成 SVG。Python 27 项、Node 9 项、Ruff、JS/Python 语法及 diff 检查通过；无 schema、migration、HTTP API 或 Python 依赖变化。 - **08-03 / 0.60.23 / Office→PDF 组件感知 + 展示路径兼容**:生产 task `92ac20cf` 暴露两层问题：上传消息给出 user-root 相对的 `测试pdf/x.doc`，host tool 又按 task_dir 拼接导致首次找不到；改用裸文件名后，host 仅安装 `libreoffice-impress` 却因只检测 `soffice` 而错误宣称支持 DOC，最终 Writer 导入返回 `source file could not be loaded`。现 `office_to_pdf` 同时解析 task 相对、user-root 展示路径与 `/workspace` 路径；Debian/Ubuntu 按 Writer/Calc/Impress 实际安装包缓存支持后缀，工具 schema 只声明可用格式，缺组件在启动转换前给出明确管理员提示。RUN bootstrap 与故障表同步要求 host 安装三组件；相关 30 项 unittest、Python 编译及 diff 格式检查通过，本机无 LibreOffice，真实 `.doc/.xlsx/.pptx` 冒烟留部署 host；无 schema、migration、HTTP API 或 Python 依赖变化，无需重建沙箱镜像。 diff --git a/config/models/deepseek_v4.yaml b/config/models/deepseek_v4.yaml index a7e2438..1e61ddb 100644 --- a/config/models/deepseek_v4.yaml +++ b/config/models/deepseek_v4.yaml @@ -9,21 +9,23 @@ variants: api_key_env: DEEPSEEK_API_KEY max_context: 1048576 reliable_context: 262144 - max_output: 8192 + max_output: 8192 # 官方上限 384K；zcbot 先保留稳定输出预算，压测后再放大 parallel_tools: false tool_calling_quality: good - thinking_mode: false - reasoning_effort_levels: [] - default_reasoning_effort: "" + thinking_enabled: true + thinking_transport: extra_body + reasoning_effort_levels: [low, high, max] + default_reasoning_effort: high code_quality: good enable_run_python: true max_iterations: 120 # backstop 兜底,非"轮"预算;真正的空转防护是 loop 的无进展熔断 + _RepeatGuard optimal_temperature: 0.3 prompt_caching: false extended_thinking: false + # Flash-0731 官方美元价按 1 USD ~= 7.2 CNY 折算；峰谷价正式生效后再同步。 input_cny_per_mtoken: 1.0 - output_cny_per_mtoken: 12.0 - cache_hit_cny_per_mtoken: 0.1 # DeepSeek 前缀缓存命中价(input 的 ~0.1x) + output_cny_per_mtoken: 2.0 + cache_hit_cny_per_mtoken: 0.02 # 官方 $0.0028 / M tokens pro: display_name: DeepSeek V4 Pro @@ -35,7 +37,8 @@ variants: max_output: 8192 parallel_tools: true tool_calling_quality: excellent - thinking_mode: true + thinking_enabled: true + thinking_transport: extra_body reasoning_effort_levels: [low, medium, high, max] default_reasoning_effort: medium code_quality: excellent diff --git a/config/models/doubao.yaml b/config/models/doubao.yaml index de74870..7d66093 100644 --- a/config/models/doubao.yaml +++ b/config/models/doubao.yaml @@ -3,10 +3,8 @@ # 与 config/models/local.yaml 同范式(避免 litellm volcengine provider 的版本/字段差异)。 # api_key 复用媒体侧的 ARK_API_KEY(同一火山账号),env 见 RUN.md。 # -# thinking_mode 暂设 false:Seed 2.1 是深度思考模型,但开关走 Ark body `thinking:{type:enabled}`, -# 与 OpenAI/DeepSeek 的 `reasoning_effort` 等级协议不同 —— 同 glm.yaml 的处理,要 core/llm.py -# 加 family 分支才能透传等级,留 TODO。设 false 只是不发 reasoning_effort 字段;模型默认仍会 -# 深度思考并返回 reasoning_content,不影响调用。 +# Seed 2.1 的 thinking 走统一 extra_body 协议；保持既有深度思考行为，但由服务端默认 +# 改为显式 enabled。方舟未声明 effort 等级，故只传 thinking、不猜 reasoning_effort。 # 单价见各 variant(元/百万 tokens,来源:火山方舟 2026-06 发布价)。 family: doubao @@ -21,7 +19,8 @@ variants: max_output: 16384 # 模型上限 128K(含思考),这里保守取值,需要长输出可调高 parallel_tools: true # Ark 兼容 parallel_tool_calls,默认 true tool_calling_quality: good - thinking_mode: false + thinking_enabled: true + thinking_transport: extra_body reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: good @@ -44,7 +43,8 @@ variants: max_output: 16384 # 模型上限 128K(含思考),这里保守取值,需要长输出可调高 parallel_tools: true tool_calling_quality: excellent - thinking_mode: false + thinking_enabled: true + thinking_transport: extra_body reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent @@ -69,7 +69,8 @@ variants: max_output: 16384 parallel_tools: true tool_calling_quality: excellent - thinking_mode: false + thinking_enabled: true + thinking_transport: extra_body reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent diff --git a/config/models/glm.yaml b/config/models/glm.yaml index b24e9fb..8064906 100644 --- a/config/models/glm.yaml +++ b/config/models/glm.yaml @@ -1,9 +1,7 @@ # 智谱 GLM 模型档案 # 走 litellm 原生 zai provider(1.83+ 内置)。litellm 默认 api_base 是国际站 api.z.ai; # 国内站 bigmodel.cn 通过 yaml 的 api_base 字段覆盖。两站 API key 不通用,env 也分开命名。 -# thinking 已接(core/llm.py _build_kwargs 的 family=="glm" 分支):GLM 协议是 body -# `{"thinking":{"type":"enabled|disabled"}}`(与 OpenAI/DeepSeek 的 `reasoning_effort` -# 等级不同族),走 extra_body 透传,由本档 `thinking_mode` 决定开关。当前均 false=禁用 —— +# thinking 走统一 extra_body 协议,由本档 `thinking_enabled` 决定开关。当前均 false=禁用 —— # 因网关侧默认开 thinking,重任务上会把输出预算烧在 reasoning_content 上撞满 65536 上限 # 被截断→空响应(task 35744bea 案);禁用后线上探针实测 reasoning_content 归零、正文照常。 family: glm @@ -19,7 +17,8 @@ variants: max_output: 8192 parallel_tools: false tool_calling_quality: good - thinking_mode: false + thinking_enabled: false + thinking_transport: extra_body reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: good @@ -41,7 +40,8 @@ variants: max_output: 8192 parallel_tools: false tool_calling_quality: good - thinking_mode: false + thinking_enabled: false + thinking_transport: extra_body reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent diff --git a/config/models/local.yaml b/config/models/local.yaml index d2c8677..64f6f3e 100644 --- a/config/models/local.yaml +++ b/config/models/local.yaml @@ -2,7 +2,7 @@ # 走 OpenAI 兼容协议(litellm provider 前缀 `openai/`,后段为实际 model 字段透传给 base_url)。 # 涉密任务时用户显式选 local.r1 / local.qwq 代替默认 deepseek_v4.flash;不走自动路由。 # 两个 variant 共用同一台推理服务器(api_base 同),api_key_env 也共用 LOCAL_LLM_API_KEY。 -# thinking_mode=false:R1 / QwQ 是天生推理模型,默认就思考,不通过 reasoning_effort 等级控制 +# thinking_enabled=false:R1 / QwQ 是天生推理模型,默认就思考,不通过 reasoning_effort 等级控制 # (那是 OpenAI / DeepSeek V4 风格);设 true 会发 reasoning_effort 字段,本地 vLLM / sglang # 多半不认,报 400。 family: local @@ -18,7 +18,8 @@ variants: max_output: 8192 parallel_tools: false tool_calling_quality: fair - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: good @@ -41,7 +42,8 @@ variants: max_output: 8192 parallel_tools: false tool_calling_quality: fair - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: good diff --git a/config/models/unifyllm.yaml b/config/models/unifyllm.yaml index f35f93e..62c0ce0 100644 --- a/config/models/unifyllm.yaml +++ b/config/models/unifyllm.yaml @@ -23,7 +23,8 @@ variants: max_output: 8192 parallel_tools: true tool_calling_quality: excellent - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent @@ -43,7 +44,8 @@ variants: max_output: 8192 parallel_tools: true tool_calling_quality: excellent - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent @@ -63,7 +65,8 @@ variants: max_output: 8192 parallel_tools: true tool_calling_quality: excellent - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent @@ -83,7 +86,8 @@ variants: max_output: 8192 parallel_tools: true tool_calling_quality: excellent - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent @@ -103,7 +107,8 @@ variants: max_output: 8192 parallel_tools: true tool_calling_quality: excellent - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent @@ -123,7 +128,8 @@ variants: max_output: 8192 parallel_tools: true tool_calling_quality: good - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: good @@ -143,7 +149,8 @@ variants: max_output: 8192 parallel_tools: false # gemini 走网关未实测该参数,保守关闭 tool_calling_quality: good - thinking_mode: false + thinking_enabled: false + thinking_transport: none reasoning_effort_levels: [] default_reasoning_effort: "" code_quality: excellent diff --git a/core/__init__.py b/core/__init__.py index 4cb7cbb..f9862a4 100644 --- a/core/__init__.py +++ b/core/__init__.py @@ -1,3 +1,3 @@ # zcbot 版本号单一事实源:web/app.py 的 FastAPI version、/healthz 返回、前端展示都引这里。 # 改版本只动这一行。 -__version__ = "0.60.25" +__version__ = "0.60.26" diff --git a/core/capabilities.py b/core/capabilities.py index cd578dc..e355655 100644 --- a/core/capabilities.py +++ b/core/capabilities.py @@ -7,6 +7,8 @@ from typing import List import yaml +from .llm_params import THINKING_TRANSPORTS + @dataclass class ModelCapabilities: @@ -24,8 +26,10 @@ class ModelCapabilities: parallel_tools: bool = False tool_calling_quality: str = "good" - # 思考模式 - thinking_mode: bool = False + # 思考开关 + thinking_enabled: bool = False + # none=不猜 provider 默认值；extra_body=显式发送 thinking.type，effort 同体透传。 + thinking_transport: str = "none" reasoning_effort_levels: List[str] = field(default_factory=list) default_reasoning_effort: str = "" @@ -71,9 +75,27 @@ class ModelCapabilities: f"档案 {path} 没有 variant={variant};可选: {list(variants)}" ) - var = variants[variant] + var = dict(variants[variant]) valid_keys = {f.name for f in fields(cls)} kwargs = {k: v for k, v in var.items() if k in valid_keys} kwargs["family"] = data.get("family", family) kwargs["variant"] = variant - return cls(**kwargs) + caps = cls(**kwargs) + if caps.thinking_transport not in THINKING_TRANSPORTS: + raise ValueError( + f"档案 {path} 的 thinking_transport={caps.thinking_transport!r} 无效;" + f"可选: {sorted(THINKING_TRANSPORTS)}" + ) + if caps.thinking_enabled and caps.thinking_transport == "none": + raise ValueError( + f"档案 {path} 开启 thinking 时必须声明可验证的 thinking_transport" + ) + if ( + caps.default_reasoning_effort + and caps.default_reasoning_effort not in caps.reasoning_effort_levels + ): + raise ValueError( + f"档案 {path} 的 default_reasoning_effort=" + f"{caps.default_reasoning_effort!r} 不在 reasoning_effort_levels 中" + ) + return caps diff --git a/core/llm.py b/core/llm.py index dfbda66..fa213a3 100644 --- a/core/llm.py +++ b/core/llm.py @@ -25,6 +25,7 @@ from litellm.exceptions import ( ) from .capabilities import ModelCapabilities +from .llm_params import build_thinking_kwargs # 单次 LLM 请求超时(秒),默认与 litellm 一致(600s)但显式化 + env 可调 ── # 长思考模型真被掐("600s 无字节 → run 标 error")时调大 ZCBOT_LLM_TIMEOUT_S 即可, @@ -64,18 +65,13 @@ class LLM: kwargs["tools"] = tools if self.caps.parallel_tools and parallel_tool_calls is not False: kwargs["parallel_tool_calls"] = True - if self.caps.thinking_mode and reasoning_effort: - kwargs["reasoning_effort"] = reasoning_effort - # GLM(zai)的 thinking 网关侧默认开着 —— 重任务上把整个输出预算烧在 reasoning_content - # 上、撞满模型自带输出上限(65536)被截断,回来 content 空 + 无 tool_call,被 loop 判空 - # 响应整轮丢弃 + 无效重试(task 35744bea 案:5 次 empty 全 tokens_out=65536,重试同上下文 - # 再撞)。GLM 的 thinking 协议是 body {"thinking":{"type":...}}(与 OpenAI 的 - # reasoning_effort 等级协议不同族),走 extra_body 透传 —— 线上探针实测 disabled 后 - # reasoning_content 归零、正文/工具照常。config thinking_mode 决定开关(当前 glm 档均 false)。 - if self.caps.family == "glm": - kwargs["extra_body"] = { - "thinking": {"type": "enabled" if self.caps.thinking_mode else "disabled"} - } + kwargs.update( + build_thinking_kwargs( + enabled=self.caps.thinking_enabled, + transport=self.caps.thinking_transport, + reasoning_effort=reasoning_effort, + ) + ) if self.caps.prompt_caching: kwargs["extra_headers"] = {"anthropic-beta": "prompt-caching-2024-07-31"} return kwargs diff --git a/core/llm_params.py b/core/llm_params.py new file mode 100644 index 0000000..6ec1b87 --- /dev/null +++ b/core/llm_params.py @@ -0,0 +1,30 @@ +"""跨 provider 的可选 LLM 请求参数构造。""" +from __future__ import annotations + +from typing import Any + +THINKING_TRANSPORTS = {"none", "extra_body"} + + +def build_thinking_kwargs( + *, enabled: bool, transport: str, reasoning_effort: str | None +) -> dict[str, Any]: + """把统一的 thinking 配置转换为 LiteLLM 调用参数。 + + ``none`` 表示该模型没有经过验证的显式控制协议，不猜 provider 默认值。 + ``extra_body`` 对应当前 DeepSeek、GLM 与方舟 ChatCompletions 的共同协议； + effort 仅在开启且档案提供非空值时发送。 + """ + if transport == "none": + return {} + if transport != "extra_body": + raise ValueError( + f"不支持的 thinking_transport={transport!r};可选: {sorted(THINKING_TRANSPORTS)}" + ) + + body: dict[str, Any] = { + "thinking": {"type": "enabled" if enabled else "disabled"} + } + if enabled and reasoning_effort: + body["reasoning_effort"] = reasoning_effort + return {"extra_body": body} diff --git a/core/probe.py b/core/probe.py index 554b274..9a02250 100644 --- a/core/probe.py +++ b/core/probe.py @@ -6,7 +6,7 @@ 四项探测: - basic_chat:连通性。失败则跳过其余。 - parallel_tools:给两个独立工具,看 single response 里 tool_calls 数量。 -- thinking_mode:对 declared=True 的模型传 reasoning_effort,看 API 是否接受 + 是否产出 thinking。 +- thinking:对 enabled=True 的模型传统一参数,看 API 是否接受 + 是否产出 thinking。 - long_context(opt-in):needle-in-haystack 简化版,默认探 reliable_context 的 1/8。 """ from __future__ import annotations @@ -133,11 +133,11 @@ def probe_parallel_tools(llm: LLM, caps: ModelCapabilities) -> ProbeResult: ) -def probe_thinking_mode(llm: LLM, caps: ModelCapabilities) -> ProbeResult: - declared = caps.thinking_mode +def probe_thinking(llm: LLM, caps: ModelCapabilities) -> ProbeResult: + declared = caps.thinking_enabled if not declared: return ProbeResult( - name="thinking_mode", + name="thinking", declared=False, observed=None, status="skip", @@ -145,7 +145,7 @@ def probe_thinking_mode(llm: LLM, caps: ModelCapabilities) -> ProbeResult: ) effort = ( caps.default_reasoning_effort - or (caps.reasoning_effort_levels[0] if caps.reasoning_effort_levels else "medium") + or (caps.reasoning_effort_levels[0] if caps.reasoning_effort_levels else None) ) try: resp = llm.chat( @@ -162,22 +162,22 @@ def probe_thinking_mode(llm: LLM, caps: ModelCapabilities) -> ProbeResult: ) observed = bool(rc) return ProbeResult( - name="thinking_mode", + name="thinking", declared=True, observed=observed, status="ok" if observed else "mismatch", detail=( - f"reasoning_effort={effort} accepted; " + (f"reasoning_effort={effort} accepted; " if effort else "thinking enabled; ") + ("thinking content returned" if observed else "no thinking content in response") ), ) except Exception as e: return ProbeResult( - name="thinking_mode", + name="thinking", declared=True, observed=False, status="mismatch", - detail=f"reasoning_effort rejected: {type(e).__name__}: {e}", + detail=f"thinking parameters rejected: {type(e).__name__}: {e}", ) @@ -237,7 +237,7 @@ def probe_capabilities( if report.results[0].status == "error": return report report.add(probe_parallel_tools(llm, caps)) - report.add(probe_thinking_mode(llm, caps)) + report.add(probe_thinking(llm, caps)) if include_long_context: report.add(probe_long_context(llm, caps)) return report diff --git a/tests/test_llm_kwargs.py b/tests/test_llm_kwargs.py new file mode 100644 index 0000000..a0f9fd5 --- /dev/null +++ b/tests/test_llm_kwargs.py @@ -0,0 +1,115 @@ +import os +import unittest +from pathlib import Path +from unittest.mock import patch + +from core.capabilities import ModelCapabilities +from core.llm import LLM + + +class LLMKwargsTests(unittest.TestCase): + def _llm( + self, *, family: str, thinking_enabled: bool, thinking_transport: str + ) -> LLM: + caps = ModelCapabilities( + family=family, + model_id=f"{family}/model", + api_key_env="TEST_LLM_API_KEY", + thinking_enabled=thinking_enabled, + thinking_transport=thinking_transport, + optimal_temperature=0.3, + ) + with patch.dict(os.environ, {"TEST_LLM_API_KEY": "test-key"}): + return LLM(caps) + + def test_deepseek_explicitly_enables_thinking_and_sets_effort(self) -> None: + llm = self._llm( + family="deepseek_v4", thinking_enabled=True, thinking_transport="extra_body" + ) + + kwargs = llm._build_kwargs( + [{"role": "user", "content": "hello"}], None, None, "high" + ) + + self.assertNotIn("reasoning_effort", kwargs) + self.assertEqual( + kwargs["extra_body"], + {"thinking": {"type": "enabled"}, "reasoning_effort": "high"}, + ) + + def test_deepseek_explicitly_disables_thinking_without_effort(self) -> None: + llm = self._llm( + family="deepseek_v4", thinking_enabled=False, thinking_transport="extra_body" + ) + + kwargs = llm._build_kwargs( + [{"role": "user", "content": "hello"}], None, None, "high" + ) + + self.assertNotIn("reasoning_effort", kwargs) + self.assertEqual( + kwargs["extra_body"], {"thinking": {"type": "disabled"}} + ) + + def test_other_openai_compatible_provider_gets_no_thinking_body(self) -> None: + llm = self._llm( + family="unifyllm", thinking_enabled=False, thinking_transport="none" + ) + + kwargs = llm._build_kwargs( + [{"role": "user", "content": "hello"}], None, None, None + ) + + self.assertNotIn("extra_body", kwargs) + + def test_extra_body_transport_enables_thinking_without_effort(self) -> None: + llm = self._llm( + family="doubao", thinking_enabled=True, thinking_transport="extra_body" + ) + + kwargs = llm._build_kwargs( + [{"role": "user", "content": "hello"}], None, None, None + ) + + self.assertEqual( + kwargs["extra_body"], {"thinking": {"type": "enabled"}} + ) + + def test_invalid_transport_is_rejected_before_request(self) -> None: + llm = self._llm( + family="test", thinking_enabled=True, thinking_transport="unknown" + ) + + with self.assertRaisesRegex(ValueError, "thinking_transport"): + llm._build_kwargs( + [{"role": "user", "content": "hello"}], None, None, "high" + ) + + def test_flash_profile_matches_0731_capabilities(self) -> None: + caps = ModelCapabilities.load( + "deepseek_v4.flash", Path(__file__).resolve().parents[1] / "config" / "models" + ) + + self.assertTrue(caps.thinking_enabled) + self.assertEqual(caps.reasoning_effort_levels, ["low", "high", "max"]) + self.assertEqual(caps.default_reasoning_effort, "high") + self.assertEqual(caps.max_output, 8192) + self.assertEqual(caps.output_cny_per_mtoken, 2.0) + self.assertEqual(caps.cache_hit_cny_per_mtoken, 0.02) + self.assertEqual(caps.thinking_transport, "extra_body") + + def test_other_controllable_profiles_declare_transport(self) -> None: + models_dir = Path(__file__).resolve().parents[1] / "config" / "models" + + glm = ModelCapabilities.load("glm.pro52", models_dir) + doubao = ModelCapabilities.load("doubao.turbo", models_dir) + + self.assertFalse(glm.thinking_enabled) + self.assertEqual(glm.thinking_transport, "extra_body") + self.assertTrue(doubao.thinking_enabled) + self.assertEqual(doubao.thinking_transport, "extra_body") + self.assertEqual(doubao.default_reasoning_effort, "") + + +if __name__ == "__main__": + unittest.main() diff --git a/web/routers/models.py b/web/routers/models.py index 3c132ad..ddf5559 100644 --- a/web/routers/models.py +++ b/web/routers/models.py @@ -54,7 +54,7 @@ def register_model_routes(app, *, require_user) -> None: "display_name": caps.display_name or profile, "family": caps.family, "variant": caps.variant, - "thinking_mode": caps.thinking_mode, + "thinking_enabled": caps.thinking_enabled, "is_default": profile == default, }) return {"models": out}