From 962e0c3859b804f4bbc96304da8ec96dffcba040 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=95=85=E7=92=83?= Date: Fri, 20 Nov 2026 14:38:59 +0800 Subject: [PATCH] chore(release): complete 2.10.4 changelogs --- CHANGELOG.md | 20 ++++++++++++++++++++ CHANGELOG.zh.md | 20 ++++++++++++++++++++ 2 files changed, 40 insertions(+) diff --git a/CHANGELOG.md b/CHANGELOG.md index 9a934a7c..23bd086e 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -6,6 +6,26 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and [中文版](CHANGELOG.zh.md) · [README](README.md) · [Contributing](CONTRIBUTING.md) +## [2.10.4] - 2026-11-20 + +### Added + +- **RL training command group**: New `bl finetune rl` group for end-to-end reinforcement learning on Bailian. + - `bl finetune rl authorize` checks the admission whitelist, service-linked roles and service activation. Roles are never created automatically; a missing role stops with bilingual manual-authorization guidance. `--authorize` enables the RL training mode (high-risk, requires `--yes`). + - `bl finetune rl run` submits jobs from flags or a YAML spec and packages Rollout/Reward function code. Submission executes inside a managed, OS-sandboxed Python runtime (macOS/Linux only). Creating a job consumes billed MTU capacity and is gated by dry-run → explicit confirmation → `--yes`; `--watch` polls to a terminal state and a timeout does not cancel the cloud job. + - `bl finetune rl get` / `list` / `watch` / `logs` / `cancel` / `delete` manage jobs through the shared fine-tunes backend; `get` returns the full OpenAPI response including datasets and MTU spec/capacity. +- **OPD distillation command group**: New `bl finetune opd` group for pure-model online distillation (`pg_opd`) from a platform-registered teacher model. + - `bl finetune opd authorize` / `run` / `params` / `get` / `watch` / `logs` / `cancel` / `delete`; `run` previews the job spec with `--dry-run` before the billed submit, and `params` queries supported hyper-parameters for a student base model. +- **Managed runtime commands**: New `bl finetune runtime prepare` / `doctor` / `clean` for the RL/OPD sandbox runtime. + - `prepare` downloads a pinned CPython + dashscope runtime from reviewed multi-source manifests with SHA256 verification and mirror fallback, then caches it for reuse; `BAILIAN_RL_PYTHON_MIRROR` / `BAILIAN_RL_PIP_INDEX_URL` cover restricted regions and `BAILIAN_RL_RUNTIME` reuses a local development runtime. + - `doctor` verifies sandbox isolation (credential read denial, home redirection, SDK import, proxy egress control) and probes host outbound network reachability as an advisory check, so agent sandboxes that block network are reported with guidance instead of failing mid-submission. + - `clean` reclaims orphaned task directories and superseded runtime versions with a `--dry-run` preview; the active baseline runtime is never touched. +- **Sandbox network-denial hints**: Connect-level network denials (`EPERM` / `EACCES` / `ENETUNREACH` / `EHOSTUNREACH`) now produce bilingual hints explaining that the execution environment (for example an AI agent sandbox) may be blocking outbound network, with guidance to enable network access, run in a regular terminal, or set `HTTPS_PROXY`. Runtime download failures also carry the underlying errno. + +### Changed + +- **Proxy environment variables**: Lowercase `http_proxy` / `https_proxy` / `no_proxy` are now honored as a fallback when the uppercase variants are unset; uppercase still takes precedence. + ## [2.10.3] - 2026-11-18 ### Fixed diff --git a/CHANGELOG.zh.md b/CHANGELOG.zh.md index df6d8130..c39dee5a 100644 --- a/CHANGELOG.zh.md +++ b/CHANGELOG.zh.md @@ -6,6 +6,26 @@ [English](CHANGELOG.md) · [README](README.zh.md) · [参与贡献](CONTRIBUTING.zh.md) +## [2.10.4] - 2026-11-20 + +### 新增 + +- **RL 训练命令组**:新增 `bl finetune rl` 组,覆盖百炼强化学习训练的端到端链路。 + - `bl finetune rl authorize` 检查准入白名单、服务关联角色与业务开通状态;绝不自动创建角色,角色缺失时中止并给出双语手动授权指引。`--authorize` 开通 RL 训练模式(高风险,需 `--yes`)。 + - `bl finetune rl run` 通过 flag 或 YAML 规格提交训练任务,并打包 Rollout/Reward 函数代码;提交在托管的 OS 沙箱 Python 运行时内执行(仅 macOS/Linux)。创建任务消耗计费的 MTU 算力,遵循试运行 → 显式确认 → `--yes` 闸门;`--watch` 轮询至终态,超时不会取消云端任务。 + - `bl finetune rl get` / `list` / `watch` / `logs` / `cancel` / `delete` 经共享的 fine-tunes 后端管理任务;`get` 原样返回完整 OpenAPI 响应(含数据集与 MTU 规格/容量)。 +- **OPD 蒸馏命令组**:新增 `bl finetune opd` 组,使用平台登记的教师模型进行纯模型在线蒸馏(`pg_opd`)。 + - `bl finetune opd authorize` / `run` / `params` / `get` / `watch` / `logs` / `cancel` / `delete`;`run` 在计费提交前可用 `--dry-run` 预览任务规格,`params` 查询学生基座模型支持的超参数。 +- **托管运行时命令**:新增 `bl finetune runtime prepare` / `doctor` / `clean`,管理 RL/OPD 沙箱运行时。 + - `prepare` 按受审查的多源清单下载固定版本 CPython + dashscope 运行时,逐源 SHA256 校验并带镜像回退,缓存后长期复用;受限地区可设 `BAILIAN_RL_PYTHON_MIRROR` / `BAILIAN_RL_PIP_INDEX_URL`,`BAILIAN_RL_RUNTIME` 可复用本机开发态运行时。 + - `doctor` 验证沙箱隔离(凭据禁读、home 重定向、SDK 导入、代理出口管控),并以增强项探测宿主出站网络可达性:阻断网络的 agent 沙箱会收到降级提示与指引,而不是在提交中途失败。 + - `clean` 回收孤儿任务目录与旧版本运行时,支持 `--dry-run` 预览;绝不触碰当前基线运行时。 +- **沙箱网络拒绝提示**:连接级网络拒绝(`EPERM` / `EACCES` / `ENETUNREACH` / `EHOSTUNREACH`)现在给出双语提示,说明执行环境(如 AI agent 沙箱)可能阻断出站网络,并引导放行网络、改在普通终端执行或设置 `HTTPS_PROXY`;运行时下载失败也会携带底层 errno。 + +### 变更 + +- **代理环境变量**:大写变量未设置时,小写 `http_proxy` / `https_proxy` / `no_proxy` 现在作为兜底生效;大写仍优先。 + ## [2.10.3] - 2026-11-18 ### 修复