Merge pull request #91 from modelstudioai/feat/model-deploy-update

feat: add tts/i2i/t2i model deploy

Author: gujieye

AGENTS.md

@@ -104,6 +104,17 @@ CLI 只为「自己能权威解释的错误」发出语义化信号,服务端的
 
 如果命令调用 Console Gateway,`defineCommand` 必须设置 `auth: "console"`。runtime 会基于 `CONSOLE_AUTH_FLAGS` 自动在 help 中展示 `--console-region`、`--console-site`、`--console-switch-agent`、`--workspace-id`,并由 `authStage` 解析/注入 console credential。命令不要重复声明这些凭证域 flag,也不要手动从 env/config 解析 token。
 
+### 5. 禁止单字母变量命名
+
+所有变量、参数、回调形参必须使用有语义的命名,不允许单字母(如 `i`、`m`、`p`、`t`、`e`、`s`)。具体表现:
+
+- 回调参数: `.map((m) => ...)` → `.map((model) => ...)`, `.find((t) => ...)` → `.find((template) => ...)`
+- catch 变量: `catch (e)` → `catch (error)`
+- for-of 循环: `for (const i of items)` → `for (const item of items)`
+- 临时变量: `const s = ...` → `const strategy = ...`
+
+例外: 仅当作用域极小(≤3 行)且语义从上下文完全明确时,可使用 `k`/`v`(Object.entries 的 key/value)。
+
 ## 完成改动后的快速验证
 
 ```sh

README.md

@@ -38,7 +38,7 @@ Equip your AI Agent out-of-the-box with these capabilities, composable across co
 - **MCP integration** — Orchestrate Bailian MCP servers: list services, inspect tools, and invoke any tool directly from the terminal
 - **Web search** — Real-time internet retrieval for up-to-date, accurate answers
 - **Model recommendation** — Describe your scenario and get best-fit model suggestions; supports scoped search, model comparison, and alternative discovery
-- **Fine-tuning & deployment** — Upload datasets, create SFT/LoRA/DPO/CPT jobs (`finetune create`), probe job status non-blockingly (`finetune watch`), query per-model training capability (`finetune capability`), and deploy trained models as endpoints (`deploy create`)
+- **Fine-tuning & deployment** — Upload datasets, create text/audio/image fine-tune jobs (`finetune text|audio|image create`; text covers SFT/LoRA/DPO/CPT), probe job status non-blockingly (`finetune watch`), query per-model training capability (`finetune capability`), and deploy trained models as endpoints (`deploy text|audio|image create`)
 - **Console capabilities** — Browse Bailian apps (`app list`), check free-tier quota (`usage free`), view model usage statistics (`usage stats`), manage workspaces (`workspace list`), and manage rate limits (`quota list/request/check/history`)
 - **Local file auto-upload** — Every URL parameter accepts a local path; uploaded to free temp storage with 48-hour validity
 
@@ -114,10 +114,10 @@ bl auth login --console
 
 # Fine-tune & deploy — a one-shot train-to-serve workflow
 bl dataset upload --file ./train.jsonl                 # Upload a .jsonl dataset (validated first)
-bl finetune create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # Local paths auto-upload
+bl finetune text create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # Local paths auto-upload
 bl finetune watch --job-id ft-xxx --output json       # Non-blocking status probe (exit 0/1/3 = done/failed/running)
 bl finetune capability --model qwen3-8b               # Which training types a model supports
-bl deploy create --model qwen3-8b --name my-svc --plan mu  # Deploy the trained model as an endpoint
+bl deploy text create --model qwen3-8b --name my-svc --plan mu  # Deploy the trained model as an endpoint
 
 # Browse apps / free-tier quota / usage statistics / workspaces
 bl app list

README.zh.md

@@ -38,7 +38,7 @@ _专为 AI Agent 打造，每个命令均可作为结构化工具调用。_
 - **MCP 集成** — 统一调度百炼 MCP 服务：列出服务、查看工具、直接在终端调用任意工具
 - **联网搜索** — 实时互联网信息检索，提升回答准确性及时效性
 - **模型推荐** — 描述你的场景，智能推荐最适合的模型；支持限定范围搜索、模型对比和替代发现
-- **微调与部署** — 上传数据集、创建 SFT/LoRA/DPO/CPT 调优任务（`finetune create`）、非阻塞探测任务状态（`finetune watch`）、按模型查训练能力（`finetune capability`），并把训练好的模型部署为推理服务（`deploy create`）
+- **微调与部署** — 上传数据集、创建文本/音频/图像调优任务（`finetune text|audio|image create`；文本涵盖 SFT/LoRA/DPO/CPT）、非阻塞探测任务状态（`finetune watch`）、按模型查训练能力（`finetune capability`），并把训练好的模型部署为推理服务（`deploy text|audio|image create`）
 - **控制台能力** — 浏览百炼应用（`app list`），查询模型免费额度（`usage free`），查看模型用量统计（`usage stats`），管理业务空间（`workspace list`），管理限流与提额（`quota list/request/check/history`）
 - **本地文件自动上传** — 所有 URL 参数同时支持本地路径，免费临时存储 48 小时
 
@@ -112,10 +112,10 @@ bl auth login --console
 
 # 微调与部署 — 从训练到服务的一站式流程
 bl dataset upload --file ./train.jsonl                 # 上传 .jsonl 数据集（先校验）
-bl finetune create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # 本地路径自动上传
+bl finetune text create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # 本地路径自动上传
 bl finetune watch --job-id ft-xxx --output json       # 非阻塞状态探测（退出码 0/1/3 = 成功/失败/进行中）
 bl finetune capability --model qwen3-8b               # 查询模型支持哪些训练方式
-bl deploy create --model qwen3-8b --name my-svc --plan mu  # 把训练好的模型部署为推理服务
+bl deploy text create --model qwen3-8b --name my-svc --plan mu  # 把训练好的模型部署为推理服务
 
 # 浏览应用 / 免费额度 / 用量统计 / 业务空间
 bl app list

packages/cli/README.md

@@ -38,7 +38,7 @@ Equip your AI Agent out-of-the-box with these capabilities, composable across co
 - **MCP integration** — Orchestrate Bailian MCP servers: list services, inspect tools, and invoke any tool directly from the terminal
 - **Web search** — Real-time internet retrieval for up-to-date, accurate answers
 - **Model recommendation** — Describe your scenario and get best-fit model suggestions; supports scoped search, model comparison, and alternative discovery
-- **Fine-tuning & deployment** — Upload datasets, create SFT/LoRA/DPO/CPT jobs (`finetune create`), probe job status non-blockingly (`finetune watch`), query per-model training capability (`finetune capability`), and deploy trained models as endpoints (`deploy create`)
+- **Fine-tuning & deployment** — Upload datasets, create text/audio/image fine-tune jobs (`finetune text|audio|image create`; text covers SFT/LoRA/DPO/CPT), probe job status non-blockingly (`finetune watch`), query per-model training capability (`finetune capability`), and deploy trained models as endpoints (`deploy text|audio|image create`)
 - **Console capabilities** — Browse Bailian apps (`app list`), check free-tier quota (`usage free`), view model usage statistics (`usage stats`), manage workspaces (`workspace list`), and manage rate limits (`quota list/request/check/history`)
 - **Local file auto-upload** — Every URL parameter accepts a local path; uploaded to free temp storage with 48-hour validity
 
@@ -114,10 +114,10 @@ bl auth login --console
 
 # Fine-tune & deploy — a one-shot train-to-serve workflow
 bl dataset upload --file ./train.jsonl                 # Upload a .jsonl dataset (validated first)
-bl finetune create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # Local paths auto-upload
+bl finetune text create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # Local paths auto-upload
 bl finetune watch --job-id ft-xxx --output json       # Non-blocking status probe (exit 0/1/3 = done/failed/running)
 bl finetune capability --model qwen3-8b               # Which training types a model supports
-bl deploy create --model qwen3-8b --name my-svc --plan mu  # Deploy the trained model as an endpoint
+bl deploy text create --model qwen3-8b --name my-svc --plan mu  # Deploy the trained model as an endpoint
 
 # Browse apps / free-tier quota / usage statistics / workspaces
 bl app list

packages/cli/README.zh.md

@@ -38,7 +38,7 @@ _专为 AI Agent 打造，每个命令均可作为结构化工具调用。_
 - **MCP 集成** — 统一调度百炼 MCP 服务：列出服务、查看工具、直接在终端调用任意工具
 - **联网搜索** — 实时互联网信息检索，提升回答准确性及时效性
 - **模型推荐** — 描述你的场景，智能推荐最适合的模型；支持限定范围搜索、模型对比和替代发现
-- **微调与部署** — 上传数据集、创建 SFT/LoRA/DPO/CPT 调优任务（`finetune create`）、非阻塞探测任务状态（`finetune watch`）、按模型查训练能力（`finetune capability`），并把训练好的模型部署为推理服务（`deploy create`）
+- **微调与部署** — 上传数据集、创建文本/音频/图像调优任务（`finetune text|audio|image create`；文本涵盖 SFT/LoRA/DPO/CPT）、非阻塞探测任务状态（`finetune watch`）、按模型查训练能力（`finetune capability`），并把训练好的模型部署为推理服务（`deploy text|audio|image create`）
 - **控制台能力** — 浏览百炼应用（`app list`），查询模型免费额度（`usage free`），查看模型用量统计（`usage stats`），管理业务空间（`workspace list`），管理限流与提额（`quota list/request/check/history`）
 - **本地文件自动上传** — 所有 URL 参数同时支持本地路径，免费临时存储 48 小时
 
@@ -112,10 +112,10 @@ bl auth login --console
 
 # 微调与部署 — 从训练到服务的一站式流程
 bl dataset upload --file ./train.jsonl                 # 上传 .jsonl 数据集（先校验）
-bl finetune create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # 本地路径自动上传
+bl finetune text create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # 本地路径自动上传
 bl finetune watch --job-id ft-xxx --output json       # 非阻塞状态探测（退出码 0/1/3 = 成功/失败/进行中）
 bl finetune capability --model qwen3-8b               # 查询模型支持哪些训练方式
-bl deploy create --model qwen3-8b --name my-svc --plan mu  # 把训练好的模型部署为推理服务
+bl deploy text create --model qwen3-8b --name my-svc --plan mu  # 把训练好的模型部署为推理服务
 
 # 浏览应用 / 免费额度 / 用量统计 / 业务空间
 bl app list

packages/cli/src/commands.ts

@@ -52,7 +52,9 @@ import {
   datasetGet,
   datasetDelete,
   datasetValidate,
-  finetuneCreate,
+  finetuneTextCreate,
+  finetuneAudioCreate,
+  finetuneImageCreate,
   finetuneList,
   finetuneGet,
   finetuneCancel,
@@ -62,7 +64,9 @@ import {
   finetuneExport,
   finetuneWatch,
   finetuneCapability,
-  deployCreate,
+  deployTextCreate,
+  deployAudioCreate,
+  deployImageCreate,
   deployList,
   deployGet,
   deployModels,
@@ -133,7 +137,9 @@ export const commands: Record<string, AnyCommand> = {
   "dataset get": datasetGet,
   "dataset delete": datasetDelete,
   "dataset validate": datasetValidate,
-  "finetune create": finetuneCreate,
+  "finetune text create": finetuneTextCreate,
+  "finetune audio create": finetuneAudioCreate,
+  "finetune image create": finetuneImageCreate,
   "finetune list": finetuneList,
   "finetune get": finetuneGet,
   "finetune cancel": finetuneCancel,
@@ -143,7 +149,9 @@ export const commands: Record<string, AnyCommand> = {
   "finetune export": finetuneExport,
   "finetune watch": finetuneWatch,
   "finetune capability": finetuneCapability,
-  "deploy create": deployCreate,
+  "deploy text create": deployTextCreate,
+  "deploy audio create": deployAudioCreate,
+  "deploy image create": deployImageCreate,
   "deploy list": deployList,
   "deploy get": deployGet,
   "deploy models": deployModels,

packages/cli/tests/e2e/dataset.e2e.test.ts

@@ -214,6 +214,83 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: dataset (offline)", () => {
     expect(data.action).toBe("dataset.upload");
     expect(data.schema).toBe("dpo");
   });
+
+  test("dataset upload --schema image --no-validate --dry-run 采用 1GB 媒体上限", async () => {
+    // image schema raises the upload cap to 1 GiB (vs 300 MB for text).
+    // --no-validate keeps this offline (the jsonl fixture is not a real zip).
+    const file = join(__dirname, ".dataset-valid.jsonl");
+    const { stdout, stderr, exitCode } = await runCli([
+      "dataset",
+      "upload",
+      "--file",
+      file,
+      "--schema",
+      "image",
+      "--no-validate",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<{ action: string; schema: string; max_bytes: number }>(stdout);
+    expect(data.action).toBe("dataset.upload");
+    expect(data.schema).toBe("image");
+    expect(data.max_bytes).toBe(1024 * 1024 * 1024);
+  });
+
+  test.each(["tts", "image"])("dataset upload --dry-run 接受媒体 schema %s", async (schema) => {
+    const file = join(__dirname, ".dataset-valid.jsonl");
+    const { stdout, stderr, exitCode } = await runCli([
+      "dataset",
+      "upload",
+      "--file",
+      file,
+      "--schema",
+      schema,
+      "--no-validate",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<{ action: string; schema: string }>(stdout);
+    expect(data.action).toBe("dataset.upload");
+    expect(data.schema).toBe(schema);
+  });
+
+  test("dataset validate --schema video 拒绝（视频生成入口已隐藏）", async () => {
+    const file = join(__dirname, ".dataset-valid.jsonl");
+    const { stdout, stderr, exitCode } = await runCli([
+      "dataset",
+      "validate",
+      "--file",
+      file,
+      "--schema",
+      "video",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stdout + stderr).not.toBe(0);
+    expect(`${stdout}\n${stderr}`).toMatch(/--schema video is not supported/);
+  });
+
+  test("dataset upload --schema video 拒绝（视频生成入口已隐藏）", async () => {
+    const file = join(__dirname, ".dataset-valid.jsonl");
+    const { stdout, stderr, exitCode } = await runCli([
+      "dataset",
+      "upload",
+      "--file",
+      file,
+      "--schema",
+      "video",
+      "--no-validate",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stdout + stderr).not.toBe(0);
+    expect(`${stdout}\n${stderr}`).toMatch(/--schema video is not supported/);
+  });
 });
 
 describe.skipIf(!isDashScopeE2EReady())("e2e: dataset (DashScope)", () => {

packages/cli/tests/e2e/deploy.e2e.test.ts

@@ -22,14 +22,15 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: deploy (offline)", () => {
   });
 
   test("deploy create --help 正常退出并展示必填项", async () => {
-    const { stderr, exitCode } = await runCli(["deploy", "create", "--help"]);
+    const { stderr, exitCode } = await runCli(["deploy", "text", "create", "--help"]);
     expect(exitCode, stderr).toBe(0);
     expect(stderr).toMatch(/--model|--name/i);
   });
 
   test("deploy create --dry-run 构造 lora 部署请求体", async () => {
     const { stdout, stderr, exitCode } = await runCli([
       "deploy",
+      "text",
       "create",
       "--model",
       "qwen-plus-2025-12-01",
@@ -56,6 +57,65 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: deploy (offline)", () => {
     expect(data.body.capacity).toBe(1);
   });
 
+  test("deploy create --plan mu --deploy-spec --dry-run 透传 deploy_spec", async () => {
+    const { stdout, stderr, exitCode } = await runCli([
+      "deploy",
+      "text",
+      "create",
+      "--model",
+      "qwen3-8b",
+      "--name",
+      "my-qwen3-mu",
+      "--plan",
+      "mu",
+      "--deploy-spec",
+      "MU1",
+      "--capacity",
+      "2",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<{
+      action: string;
+      body: { plan: string; deploy_spec?: string; capacity?: number };
+    }>(stdout);
+    expect(data.action).toBe("deploy.create");
+    expect(data.body.plan).toBe("mu");
+    expect(data.body.deploy_spec).toBe("MU1");
+    expect(data.body.capacity).toBe(2);
+  });
+
+  test("deploy audio create --dry-run 默认 plan=mu（CosyVoice 部署契约）", async () => {
+    // Audio (CosyVoice TTS) outputs deploy model-unit-billed: the modality fixes
+    // the default plan to `mu` (text/image stay `lora`). In dry-run the mu
+    // strategy skips the catalog lookup, so deploy_spec is omitted and capacity
+    // falls back to 1 with billing_method POST_PAY.
+    const { stdout, stderr, exitCode } = await runCli([
+      "deploy",
+      "audio",
+      "create",
+      "--model",
+      "my-cosyvoice-ft",
+      "--name",
+      "my-tts",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<{
+      action: string;
+      body: { plan: string; name: string; billing_method?: string; capacity?: number };
+    }>(stdout);
+    expect(data.action).toBe("deploy.create");
+    expect(data.body.plan).toBe("mu");
+    expect(data.body.name).toBe("my-tts");
+    expect(data.body.billing_method).toBe("POST_PAY");
+    expect(data.body.capacity).toBe(1);
+  });
+
   test("deploy scale --dry-run 转发 capacity", async () => {
     const { stdout, stderr, exitCode } = await runCli([
       "deploy",