feat: merge self-built-framework

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

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

@@ -214,6 +214,52 @@ 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/video schemas raise 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", "video"])(
+    "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);
+    },
+  );
 });
 
 describe.skipIf(!isDashScopeE2EReady())("e2e: dataset (DashScope)", () => {

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

@@ -56,6 +56,98 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: deploy (offline)", () => {
     expect(data.body.capacity).toBe(1);
   });
 
+  test("deploy create --dry-run 组装视频 LoRA 的 aigc_config", async () => {
+    const { stdout, stderr, exitCode } = await runCli([
+      "deploy",
+      "create",
+      "--model",
+      "wan2.5-i2v-preview-ft-xxx",
+      "--name",
+      "my-video-lora",
+      "--aigc-prompt",
+      "a cat surfing",
+      "--aigc-lora-prompt-default",
+      "trigger-word",
+      "--aigc-use-input-prompt",
+      "true",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<{
+      action: string;
+      body: {
+        plan: string;
+        aigc_config?: {
+          use_input_prompt?: boolean;
+          prompt?: string;
+          lora_prompt_default?: string;
+        };
+      };
+    }>(stdout);
+    expect(data.action).toBe("deploy.create");
+    expect(data.body.plan).toBe("lora");
+    expect(data.body.aigc_config).toEqual({
+      use_input_prompt: true,
+      prompt: "a cat surfing",
+      lora_prompt_default: "trigger-word",
+    });
+  });
+
+  test("deploy create --aigc-* 仅对 plan=lora 有效（plan=ptu 时报错）", async () => {
+    const { stdout, stderr, exitCode } = await runCli([
+      "deploy",
+      "create",
+      "--model",
+      "wan2.5-i2v-preview-ft-xxx",
+      "--name",
+      "my-video-lora",
+      "--plan",
+      "ptu",
+      "--input-tpm",
+      "10000",
+      "--output-tpm",
+      "1000",
+      "--aigc-prompt",
+      "a cat surfing",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stdout + stderr).not.toBe(0);
+    expect(`${stdout}\n${stderr}`).toMatch(/--aigc-\* flags are only valid for plan=lora/);
+  });
+
+  test("deploy create --plan mu --deploy-spec --dry-run 透传 deploy_spec", async () => {
+    const { stdout, stderr, exitCode } = await runCli([
+      "deploy",
+      "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 scale --dry-run 转发 capacity", async () => {
     const { stdout, stderr, exitCode } = await runCli([
       "deploy",

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

@@ -114,6 +114,35 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: finetune (offline)", () => {
     });
   });
 
+  test.each([
+    ["sft", "sft"],
+    ["sft-lora", "efficient_sft"],
+    ["dpo", "dpo_full"],
+    ["dpo-lora", "dpo_lora"],
+    ["cpt", "cpt"],
+  ])(
+    "finetune create --training-type %s 经 profile 映射为 server 类型 %s",
+    async (cliType, serverType) => {
+      const { stdout, stderr, exitCode } = await runCli([
+        "finetune",
+        "create",
+        "--model",
+        "qwen3-8b",
+        "--datasets",
+        "file-aaa",
+        "--training-type",
+        cliType,
+        "--dry-run",
+        "--output",
+        "json",
+      ]);
+      expect(exitCode, stderr).toBe(0);
+      const data = parseStdoutJson<{ action: string; body: { training_type: string } }>(stdout);
+      expect(data.action).toBe("finetune.create");
+      expect(data.body.training_type).toBe(serverType);
+    },
+  );
+
   test("finetune create --training-type 拒绝不支持的训练类型值", async () => {
     const { stdout, stderr, exitCode } = await runCli([
       "finetune",

packages/commands/src/commands/dataset/upload.ts

@@ -6,6 +6,7 @@ import {
   parseDatasetSchemaFlag,
   formatIssue,
   MAX_DATASET_BYTES,
+  MAX_MEDIA_ZIP_BYTES,
   BailianError,
   ExitCode,
   type DatasetFile,
@@ -17,7 +18,7 @@ const UPLOAD_FLAGS = {
   file: {
     type: "string",
     valueHint: "<path>",
-    description: "Local .jsonl dataset file (≤300MB)",
+    description: "Local dataset file (.jsonl or .zip; ≤300MB text, ≤1GB image/video)",
     required: true,
   },
   purpose: {
@@ -29,7 +30,7 @@ const UPLOAD_FLAGS = {
     type: "string",
     valueHint: "<s>",
     description:
-      'Record schema: "chatml" (SFT), "dpo" (chosen/rejected), or "cpt" (raw text). Default auto-detects per record.',
+      'Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record.',
   },
   noValidate: {
     type: "switch",
@@ -42,30 +43,31 @@ const UPLOAD_FLAGS = {
 } satisfies FlagsDef;
 
 export default defineCommand({
-  description: "Upload a dataset file (.jsonl) to Bailian",
+  description: "Upload a dataset file (.jsonl or .zip) to Bailian",
   auth: "apiKey",
   usageArgs:
-    "--file <path> [--purpose <name>] [--schema <chatml|dpo|cpt>] [--no-validate] [--full-validate]",
+    "--file <path> [--purpose <name>] [--schema <chatml|dpo|cpt|tts|image|video>] [--no-validate] [--full-validate]",
   flags: UPLOAD_FLAGS,
   exampleArgs: [
     "--file train.jsonl",
     "--file dpo.jsonl --schema dpo",
     "--file cpt.jsonl --schema cpt",
+    "--file audio.zip --schema tts",
     "--file eval.jsonl --purpose evaluation",
     "--file train.jsonl --full-validate",
     "--file train.jsonl --no-validate",
   ],
   notes: [
-    "Only .jsonl is supported in this release. Three record schemas are",
-    "recognized: chatml = {messages:[...]} (SFT); dpo = {messages:[...],",
-    "chosen, rejected} where chosen/rejected are single assistant messages;",
-    'cpt = {text:"..."} (continual pre-training, raw text). With no --schema,',
-    "a record carrying chosen/rejected is validated as DPO, one with text (and",
-    "no messages) as CPT, otherwise as ChatML. Pass --schema dpo / cpt to",
-    "require that shape on every record, or --schema chatml to ignore the",
-    "preference / text fields. Other purposes may carry a different schema in",
-    "the future and would be served by a purpose-specific validator.",
-    "The dataset upload cap is 300MB per file.",
+    "Supports .jsonl (text) and .zip (audio/image/video archives with a",
+    "data.jsonl manifest). Six record schemas are recognized: chatml =",
+    "{messages:[...]} (SFT); dpo = {messages:[...], chosen, rejected};",
+    'cpt = {text:"..."} (continual pre-training, raw text); tts =',
+    '{wav_fn:"train/xxx.wav", text:"..."} (audio fine-tuning); image =',
+    '{img_path:"..."} (image generation); video = {first_frame_path:"...",',
+    'video_path:"..."} (video generation). With no --schema, a record',
+    "carrying wav_fn is validated as TTS, img_path as image, video_path /",
+    "first_frame_path as video, chosen/rejected as DPO, text (no messages)",
+    "as CPT, otherwise ChatML. Upload cap: 300MB text, 1GB image/video.",
     "Upload uses the OpenAI-compatible /compatible-mode/v1/files endpoint so",
     "the purpose tag is persisted (the DashScope-native /api/v1/files drops it).",
   ],
@@ -75,9 +77,16 @@ export default defineCommand({
     const purpose = flags.purpose || "fine-tune";
     const schema = parseDatasetSchemaFlag(flags.schema);
     const format = detectOutputFormat(settings.output);
+    // Image / video schemas allow larger ZIPs (1 GB vs 300 MB for text).
+    const isMediaSchema = schema === "image" || schema === "video";
 
     if (!flags.noValidate) {
-      const result = await validateDataset(filePath, { fullValidate: flags.fullValidate, schema });
+      const maxBytes = isMediaSchema ? MAX_MEDIA_ZIP_BYTES : MAX_DATASET_BYTES;
+      const result = await validateDataset(filePath, {
+        fullValidate: flags.fullValidate,
+        schema,
+        maxBytes,
+      });
       if (!result.valid) {
         const lines = [
           `Dataset validation failed for ${filePath}`,
@@ -112,7 +121,7 @@ export default defineCommand({
           action: "dataset.upload",
           file: filePath,
           purpose,
-          max_bytes: MAX_DATASET_BYTES,
+          max_bytes: isMediaSchema ? MAX_MEDIA_ZIP_BYTES : MAX_DATASET_BYTES,
           validate: !flags.noValidate,
           schema: schema ?? "auto",
         },

packages/commands/src/commands/dataset/validate.ts

@@ -25,7 +25,7 @@ const VALIDATE_FLAGS = {
   file: {
     type: "string",
     valueHint: "<path>",
-    description: "Local .jsonl dataset file",
+    description: "Local dataset file (.jsonl or .zip)",
     required: true,
   },
   fullValidate: {
@@ -36,33 +36,38 @@ const VALIDATE_FLAGS = {
     type: "string",
     valueHint: "<s>",
     description:
-      'Record schema: "chatml" (SFT), "dpo" (chosen/rejected), or "cpt" (raw text). Default auto-detects per record.',
+      'Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record.',
   },
 } satisfies FlagsDef;
 
 export default defineCommand({
-  description: "Locally validate a dataset file (.jsonl) without uploading",
+  description: "Locally validate a dataset file (.jsonl or .zip) without uploading",
   // 纯本地校验，不触网、不需 API key（与 `pipeline validate` 一致）。
   auth: "none",
-  usageArgs: "--file <path> [--full-validate] [--schema <chatml|dpo|cpt>]",
+  usageArgs: "--file <path> [--full-validate] [--schema <chatml|dpo|cpt|tts|image|video>]",
   flags: VALIDATE_FLAGS,
   exampleArgs: [
     "--file train.jsonl",
     "--file dpo.jsonl --schema dpo",
     "--file cpt.jsonl --schema cpt",
+    "--file audio.zip --schema tts",
     "--file eval.jsonl --full-validate",
     "--file train.jsonl --output json",
   ],
   notes: [
     "Default scan: every line gets a structural check, then ~160 lines (front 50,",
     "evenly spaced 100, last 10) are JSON.parsed against the active schema.",
     "Schemas: chatml = {messages:[...]} (SFT); dpo = {messages:[...], chosen,",
-    "rejected} where chosen/rejected are single assistant messages; cpt =",
-    '{text:"..."} (continual pre-training, raw text). With no --schema, a',
-    "record carrying chosen/rejected is validated as DPO, one with text (and no",
-    "messages) as CPT, otherwise as ChatML. Pass --schema dpo / cpt to require",
-    "that shape on every record (strict), or --schema chatml to ignore the",
-    "preference / text fields. Use --full-validate to JSON.parse every line.",
+    'rejected}; cpt = {text:"..."} (continual pre-training, raw text);',
+    'tts = {wav_fn:"train/xxx.wav", text:"..."} (audio fine-tuning);',
+    'image = {img_path:"..."} (image generation); video =',
+    '{first_frame_path:"...", video_path:"..."} (video generation). With no',
+    "--schema, a record carrying wav_fn is validated as TTS, img_path as",
+    "image, video_path / first_frame_path as video, chosen/rejected as DPO,",
+    "text (no messages) as CPT, otherwise ChatML. Pass --schema to require a",
+    "specific shape on every record. ZIP archives (.zip) are validated",
+    "structurally (data.jsonl present, media references resolve) in addition",
+    "to per-record content checks. Use --full-validate to JSON.parse every line.",
   ],
   async run(ctx) {
     const { settings, flags } = ctx;