feat: video finetune/deploy/invoke full pipeline + training cost calculation

- Add finetune video create subcommand (wan2.7/2.5/2.2 i2v + kf2v)
- Align video hyperparams with official docs (n_epochs=50, per-model batch_size/max_pixels)
- Add --last-frame flag to video generate for kf2v (image2video endpoint)
- Fix wan2.1-2.6 i2v input format (img_url instead of media[])
- Add training_cost field to finetune get/watch (catalog ft price, API-key domain only)
- Add --aigc-* flags to deploy create (optional, for video LoRA prompt config)

Author: gujieye

packages/cli/src/commands.ts

@@ -62,6 +62,7 @@ import {
   finetuneTextCreate,
   finetuneAudioCreate,
   finetuneImageCreate,
+  finetuneVideoCreate,
   finetuneList,
   finetuneGet,
   finetuneCancel,
@@ -183,6 +184,7 @@ export const commands: Record<string, AnyCommand> = {
   "finetune text create": finetuneTextCreate,
   "finetune audio create": finetuneAudioCreate,
   "finetune image create": finetuneImageCreate,
+  "finetune video create": finetuneVideoCreate,
   "finetune list": finetuneList,
   "finetune get": finetuneGet,
   "finetune cancel": finetuneCancel,

packages/commands/src/commands/finetune/create.ts

@@ -207,7 +207,7 @@ async function uploadResolvedLocal(
 }
 
 /** The modality a `finetune <modality> create` subcommand is bound to. */
-type CommandModality = "text" | "audio" | "image";
+type CommandModality = "text" | "audio" | "image" | "video";
 
 /**
  * Flags shared by every `finetune <modality> create` subcommand: what to train
@@ -324,6 +324,34 @@ const AUDIO_USAGE =
 const IMAGE_USAGE =
   "--base-model <model> --datasets <id|path> [--validations <id|path>] [--model-name <name>] [--suffix <text>] [--generation-type <t2i|i2i>] [--learning-rate <str>]";
 
+/**
+ * Video (Wan i2v/kf2v) flags: exposes the three hyper-parameters that the
+ * video API supports and users may want to override. Defaults are model-specific
+ * (resolved by the sft-lora profile: wan2.7 → batch_size 1 / max_pixels 102400,
+ * wan2.5 → 4 / 36864, wan2.2 → 4 / 262144).
+ */
+const VIDEO_FLAGS = {
+  ...COMMON_FLAGS,
+  nEpochs: {
+    type: "number",
+    valueHint: "<n>",
+    description: "Training epochs (default: 50)",
+  },
+  batchSize: {
+    type: "number",
+    valueHint: "<n>",
+    description: "Batch size (default: model-specific, 1 for wan2.7, 4 for wan2.5/2.2)",
+  },
+  learningRate: {
+    type: "string",
+    valueHint: "<str>",
+    description: 'Learning rate as a string to preserve precision (default: "2e-5")',
+  },
+} satisfies FlagsDef;
+
+const VIDEO_USAGE =
+  "--base-model <model> --datasets <id|path> [--validations <id|path>] [--model-name <name>] [--suffix <text>] [--n-epochs <n>] [--batch-size <n>] [--learning-rate <str>]";
+
 const COMMON_NOTES = [
   "Creating a job uploads any local datasets and consumes training quota.",
   "Use --dry-run to preview the request body without submitting.",
@@ -683,3 +711,28 @@ export const finetuneImageCreate = defineCommand({
   notes: IMAGE_NOTES,
   run: (ctx) => runCreate("image", ctx),
 });
+
+const VIDEO_NOTES = [
+  ...COMMON_NOTES,
+  "Video generation training (Wan i2v/kf2v) runs efficient_sft with model-",
+  "specific defaults: wan2.7 (batch_size=1, max_pixels=102400), wan2.5/2.2",
+  "(batch_size=4, max_pixels per model). Override with --batch-size/--n-epochs.",
+  "Datasets are .zip archives with data.jsonl + frame images + videos.",
+  "Recommended: ≥10 training samples, 20-100 for stable results.",
+];
+
+/** `bl finetune video create` — fine-tune a video generation model. Datasets are `.zip`. */
+export const finetuneVideoCreate = defineCommand({
+  description: "Create a video generation model fine-tune job (Wan i2v/kf2v, efficient_sft)",
+  auth: "apiKey",
+  usageArgs: VIDEO_USAGE,
+  flags: VIDEO_FLAGS,
+  exampleArgs: [
+    "--base-model wan2.7-i2v --datasets file-xxx",
+    "--base-model wan2.7-i2v --datasets ./i2v-data.zip",
+    "--base-model wan2.2-kf2v-flash --datasets file-xxx --n-epochs 100",
+    "--base-model wan2.7-i2v --datasets file-xxx --dry-run",
+  ],
+  notes: VIDEO_NOTES,
+  run: (ctx) => runCreate("video", ctx),
+});

packages/commands/src/commands/finetune/fee.ts

@@ -77,9 +77,14 @@ async function fetchTrainingPrice(
 }
 
 /**
- * Compute the actual training fee from the model catalog.
+ * Compute the actual training fee from the model catalog's "ft" price entry.
  * Returns null when the price is unavailable (network error, model not in
- * catalog, or no "ft" entry in the prices array). Never throws.
+ * catalog, or no "ft" entry). Never throws.
+ *
+ * Only uses the public model catalog (model metadata) — does NOT call
+ * console-domain pricing APIs (modelCenter.getModelPrice). Models whose
+ * catalog entry lacks a "ft" price (e.g. CosyVoice) will simply omit the
+ * training_cost field until the platform adds it to the catalog.
  */
 export async function computeActualFee(
   settings: Settings,
@@ -91,7 +96,7 @@ export async function computeActualFee(
     const unitPrice = Number(ftEntry?.price);
     if (!Number.isFinite(unitPrice) || unitPrice <= 0) return null;
     const priceUnit = ftEntry?.priceUnit ?? "每百万tokens";
-    // price is yuan per million tokens.
+    // Catalog price is yuan per million tokens.
     const cost = (usageTokens / 1_000_000) * unitPrice;
     return { cost: Number(cost.toFixed(4)), unitPrice, priceUnit };
   } catch {

packages/commands/src/commands/video/generate.ts

@@ -1,6 +1,7 @@
 import {
   defineCommand,
   videoGeneratePath,
+  image2videoPath,
   taskPath,
   detectOutputFormat,
   type DashScopeVideoRequest,
@@ -43,6 +44,11 @@ export default defineCommand({
       valueHint: "<url>",
       description: "Input image URL for image-to-video generation",
     },
+    lastFrame: {
+      type: "string",
+      valueHint: "<url>",
+      description: "Last frame image URL (with --image, enables kf2v first+last frame mode)",
+    },
     negativePrompt: {
       type: "string",
       valueHint: "<text>",
@@ -110,12 +116,20 @@ export default defineCommand({
     const format = detectOutputFormat(settings.output);
 
     const imageUrl = flags.image;
+    const lastFrameUrl = flags.lastFrame as string | undefined;
 
     // Auto-upload local image file for i2v
     let resolvedImageUrl: string | undefined;
     if (imageUrl) {
       resolvedImageUrl = await ctx.client.resolveImageInput(imageUrl, model);
     }
+    let resolvedLastFrameUrl: string | undefined;
+    if (lastFrameUrl) {
+      resolvedLastFrameUrl = await ctx.client.resolveImageInput(lastFrameUrl, model);
+    }
+
+    // kf2v mode: both --image and --last-frame provided.
+    const isKf2v = Boolean(resolvedImageUrl && resolvedLastFrameUrl);
 
     const watermark = resolveWatermark(flags.watermark);
     const promptExtend = resolveBooleanFlag(flags.promptExtend, undefined, "prompt-extend");
@@ -125,10 +139,16 @@ export default defineCommand({
       input: {
         prompt: prompt,
         negative_prompt: flags.negativePrompt || undefined,
-        // i2v models (happyhorse-1.1-i2v) require input.media with type 'first_frame'
-        ...(resolvedImageUrl
-          ? { media: [{ type: "first_frame" as const, url: resolvedImageUrl }] }
-          : {}),
+        // kf2v: first+last frame flat fields via image2video endpoint.
+        // wan2.1~2.6 i2v: flat img_url via video-generation endpoint.
+        // wan2.7+ / happyhorse i2v: media[] via video-generation endpoint.
+        ...(isKf2v
+          ? { first_frame_url: resolvedImageUrl, last_frame_url: resolvedLastFrameUrl }
+          : resolvedImageUrl
+            ? /wan[x]?2\.[1-6]/i.test(model)
+              ? { img_url: resolvedImageUrl }
+              : { media: [{ type: "first_frame" as const, url: resolvedImageUrl }] }
+            : {}),
       },
       parameters: {
         resolution: flags.resolution || undefined,
@@ -141,15 +161,28 @@ export default defineCommand({
     };
 
     if (settings.dryRun) {
-      const previewBody = resolvedImageUrl
-        ? {
-            ...body,
-            input: {
-              ...body.input,
-              media: [{ type: "first_frame" as const, url: redactDataUri(resolvedImageUrl) }],
-            },
-          }
-        : body;
+      let previewBody = body;
+      if (isKf2v) {
+        previewBody = {
+          ...body,
+          input: {
+            ...body.input,
+            first_frame_url: redactDataUri(resolvedImageUrl ?? ""),
+            last_frame_url: redactDataUri(resolvedLastFrameUrl ?? ""),
+          },
+        };
+      } else if (resolvedImageUrl) {
+        const redactedUrl = redactDataUri(resolvedImageUrl);
+        previewBody = {
+          ...body,
+          input: {
+            ...body.input,
+            ...(/wan[x]?2\.[1-6]/i.test(model)
+              ? { img_url: redactedUrl }
+              : { media: [{ type: "first_frame" as const, url: redactedUrl }] }),
+          },
+        };
+      }
       emitResult({ request: previewBody }, format);
       return;
     }
@@ -162,7 +195,7 @@ export default defineCommand({
       settings,
       () =>
         ctx.client.requestJson<DashScopeAsyncResponse>({
-          path: videoGeneratePath(),
+          path: isKf2v ? image2videoPath() : videoGeneratePath(),
           method: "POST",
           body,
           async: true,

packages/commands/src/index.ts

@@ -66,6 +66,7 @@ export {
   finetuneTextCreate,
   finetuneAudioCreate,
   finetuneImageCreate,
+  finetuneVideoCreate,
 } from "./commands/finetune/create.ts";
 export { default as finetuneList } from "./commands/finetune/list.ts";
 export { default as finetuneGet } from "./commands/finetune/get.ts";

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

@@ -365,6 +365,104 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: finetune (offline)", () => {
     expect(data.action).toBe("finetune.create");
     expect(data.body.training_type).toBe("efficient_sft");
   });
+
+  test("finetune video create --help 暴露视频超参且不含文本超参", async () => {
+    // Video exposes --n-epochs / --batch-size / --learning-rate; the text-only
+    // --training-type / --max-length surface is not offered.
+    const { stderr, exitCode } = await runCommandE2e(FINETUNE_ROUTES, [
+      "finetune",
+      "video",
+      "create",
+      "--help",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    expect(stderr).toMatch(/--base-model/);
+    expect(stderr).toMatch(/--n-epochs/);
+    expect(stderr).toMatch(/--batch-size/);
+    expect(stderr).toMatch(/--learning-rate/);
+    expect(stderr).not.toMatch(/--training-type|--max-length/);
+  });
+
+  test("finetune video create --datasets 缺失时退出为用法错误 (2)", async () => {
+    const { stderr, exitCode } = await runCommandE2e(FINETUNE_ROUTES, [
+      "finetune",
+      "video",
+      "create",
+      "--base-model",
+      "wan2.7-i2v",
+      "--quiet",
+    ]);
+    expect(exitCode).toBe(2);
+    expect(stderr).toMatch(/--datasets|Missing required/i);
+  });
+
+  test.each([
+    // Model-family-specific defaults resolved by the sft-lora video profile.
+    ["wan2.7-i2v", 1, 102400],
+    ["wan2.5-i2v-preview", 4, 36864],
+    ["wan2.2-kf2v-flash", 4, 262144],
+  ])(
+    "finetune video create --dry-run %s 解析 batch_size=%i / max_pixels=%i",
+    async (baseModel, batchSize, maxPixels) => {
+      const { stdout, stderr, exitCode } = await runCommandE2e(FINETUNE_ROUTES, [
+        "finetune",
+        "video",
+        "create",
+        "--base-model",
+        baseModel,
+        "--datasets",
+        "file-video",
+        "--dry-run",
+        "--output",
+        "json",
+      ]);
+      expect(exitCode, stderr).toBe(0);
+      const data = parseStdoutJson<{
+        action: string;
+        body: {
+          model: string;
+          training_type: string;
+          hyper_parameters: Record<string, unknown>;
+        };
+      }>(stdout);
+      expect(data.action).toBe("finetune.create");
+      expect(data.body.model).toBe(baseModel);
+      expect(data.body.training_type).toBe("efficient_sft");
+      expect(data.body.hyper_parameters.batch_size).toBe(batchSize);
+      expect(data.body.hyper_parameters.max_pixels).toBe(maxPixels);
+      expect(data.body.hyper_parameters.learning_rate).toBe("2e-5");
+      expect(data.body.hyper_parameters.lora_rank).toBe(32);
+    },
+  );
+
+  test("finetune video create --dry-run 转发超参覆盖且不做 clamp", async () => {
+    const { stdout, stderr, exitCode } = await runCommandE2e(FINETUNE_ROUTES, [
+      "finetune",
+      "video",
+      "create",
+      "--base-model",
+      "wan2.7-i2v",
+      "--datasets",
+      "file-video",
+      "--n-epochs",
+      "100",
+      "--batch-size",
+      "2",
+      "--learning-rate",
+      "1e-5",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<{
+      body: { hyper_parameters: Record<string, unknown> };
+    }>(stdout);
+    // Video overrides are forwarded verbatim (no [8, 1024] text clamp).
+    expect(data.body.hyper_parameters.n_epochs).toBe(100);
+    expect(data.body.hyper_parameters.batch_size).toBe(2);
+    expect(data.body.hyper_parameters.learning_rate).toBe("1e-5");
+  });
 });
 
 describe.skipIf(!isDashScopeE2EReady())("e2e: finetune (DashScope)", () => {

packages/commands/tests/e2e/topic-routes.ts

@@ -135,6 +135,7 @@ export const FINETUNE_ROUTES: E2eRouteExports = {
   "finetune text create": "finetuneTextCreate",
   "finetune audio create": "finetuneAudioCreate",
   "finetune image create": "finetuneImageCreate",
+  "finetune video create": "finetuneVideoCreate",
   "finetune list": "finetuneList",
   "finetune get": "finetuneGet",
   "finetune cancel": "finetuneCancel",