[ci]: add preprocessing integration coverage

2026-07-16 18:13:49 -07:00
293 changed files with 1117 additions and 26865 deletions
@@ -1,8 +1,6 @@
 env:
  IMAGE_VERSION: "py3.12-latest"
  BUILDKITE_CLEAN_CHECKOUT: true
-  # Buildkite only launches Modal; remote jobs initialize their own submodules.
-  BUILDKITE_GIT_SUBMODULES: false

 notify:
  - github_commit_status:
@@ -79,6 +77,17 @@ steps:
            limit: 2
      agents:
        queue: "default"
+    - label: ":microscope: Preprocessing Tests"
+      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "preprocessing"
+      command: "timeout 90m .buildkite/scripts/pr_test.sh"
+      retry:
+        automatic:
+          - exit_status: 128
+            limit: 3
+          - exit_status: -1
+            limit: 2
+      agents:
+        queue: "default"
    - label: ":microscope: DreamVerse App Tests"
      if: build.env("TEST_SCOPE") == "direct" && build.env("TEST_TYPE") == "dreamverse_app"
      command: "timeout 90m .buildkite/scripts/pr_test.sh"
@@ -324,6 +333,29 @@ steps:
                  - TEST_TYPE=unit_test
                agents:
                  queue: "default"
+            - path:
+                - "fastvideo/pipelines/preprocess/**"
+                - "fastvideo/workflow/preprocess/**"
+                - "fastvideo/dataset/dataloader/parquet_io.py"
+                - "fastvideo/dataset/dataloader/record_schema.py"
+                - "fastvideo/dataset/dataloader/schema.py"
+                - "fastvideo/configs/configs.py"
+                - "fastvideo/fastvideo_args.py"
+                - "fastvideo/tests/workflow/test_t2v_preprocessing_e2e.py"
+                - "fastvideo/tests/nightly/reference_video_1_sample_v0.mp4"
+                - "fastvideo/tests/modal/pr_test.py"
+                - ".buildkite/pipeline.yml"
+                - ".buildkite/scripts/pr_test.sh"
+                - ".github/workflows/ci-slash-commands.yml"
+                - "pyproject.toml"
+                - "docker/Dockerfile"
+              config:
+                command: "timeout 20m .buildkite/scripts/pr_test.sh"
+                label: ":microscope: Preprocessing Tests"
+                env:
+                  - TEST_TYPE=preprocessing
+                agents:
+                  queue: "default"
            - path:
                - "apps/dreamverse/**"
                - "pyproject.toml"
@@ -404,7 +436,7 @@ steps:
                - "pyproject.toml"
                - "docker/Dockerfile"
              config:
-                command: "timeout 25m .buildkite/scripts/pr_test.sh"
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
                label: ":test_tube: Training Tests"
                env:
                  - TEST_TYPE=training
@@ -415,7 +447,7 @@ steps:
                - "pyproject.toml"
                - "docker/Dockerfile"
              config:
-                command: "timeout 25m .buildkite/scripts/pr_test.sh"
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
                label: ":test_tube: Distillation DMD Tests"
                env:
                  - TEST_TYPE=distillation_dmd
@@ -233,6 +233,10 @@ case "$TEST_TYPE" in
        log "Running unit tests..."
        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_unit_test"
        ;;
+    "preprocessing")
+        log "Running preprocessing integration test..."
+        MODAL_COMMAND="$MODAL_ENV HF_API_KEY=$HF_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_preprocessing_tests"
+        ;;
    "dreamverse_app")
        log "Running DreamVerse app tests..."
        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_dreamverse_app_tests"
@@ -27,25 +27,14 @@ jobs:
        ref: ${{ inputs.ref || '' }}
    # For PR events, lint the PR head — but keep the hook definitions from
    # the base branch so an untrusted PR cannot alter what gets executed.
-    # The gate scripts are saved too: the self-test step below executes them,
-    # so it must run the base-branch copies, not the PR head's.
-    - name: Save trusted hook config and gate scripts
+    - name: Save trusted hook config
      if: github.event_name == 'pull_request_target'
-      run: |
-        cp .pre-commit-config.yaml "$RUNNER_TEMP/trusted-pre-commit-config.yaml"
-        cp -a .github/scripts "$RUNNER_TEMP/trusted-scripts"
-        echo "GATE_SCRIPTS_DIR=$RUNNER_TEMP/trusted-scripts" >> "$GITHUB_ENV"
-    # allow-unsafe-pr-checkout acknowledges checkout's pull_request_target
-    # guard: the head is data for the trusted hooks to lint; nothing from it
-    # is executed (config and gate scripts are pinned to the base branch
-    # above) and credentials are not persisted. SHA-pinned to v4.4.0 because
-    # actionlint's action schema does not know the new input yet.
-    - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4.4.0
+      run: cp .pre-commit-config.yaml "$RUNNER_TEMP/trusted-pre-commit-config.yaml"
+    - uses: actions/checkout@v4
      if: github.event_name == 'pull_request_target'
      with:
        ref: ${{ github.event.pull_request.head.sha }}
        persist-credentials: false
-        allow-unsafe-pr-checkout: true
    - name: Restore trusted hook config
      if: github.event_name == 'pull_request_target'
      run: cp "$RUNNER_TEMP/trusted-pre-commit-config.yaml" .pre-commit-config.yaml
@@ -59,7 +48,5 @@ jobs:
      with:
        extra_args: --all-files --hook-stage manual
    # After pre-commit so a self-test failure cannot mask lint failures.
-    # GATE_SCRIPTS_DIR points at the base-branch copy on fork PRs (set above);
-    # push / workflow_call runs use the checked-out tree directly.
    - name: Full-suite gate self-test
-      run: bash "${GATE_SCRIPTS_DIR:-.github/scripts}/test_gate_full_suite.sh"
+      run: bash .github/scripts/test_gate_full_suite.sh
@@ -129,7 +129,7 @@ jobs:
          set -euo pipefail
          TEST_NAME=$(echo "$COMMENT" | grep -oP '(?<=/test\s)\S+' | head -1 || true)

-          VALID="encoder vae transformer kernel unit dreamverse ssim training lora-inference lora-training lora-extraction distillation self-forcing vsa vmoba performance api train-framework eval full fastcheck pre-commit"
+          VALID="encoder vae transformer kernel unit preprocessing dreamverse ssim training lora-inference lora-training lora-extraction distillation self-forcing vsa vmoba performance api train-framework eval full fastcheck pre-commit"
          if [ -z "$TEST_NAME" ] || ! echo "$VALID" | grep -qw "$TEST_NAME"; then
            echo "Unknown test: '$TEST_NAME'. Valid: $VALID"
            exit 1
@@ -137,7 +137,8 @@ jobs:

          declare -A MAP=(
            [encoder]=encoder [vae]=vae [transformer]=transformer
-            [kernel]=kernel_tests [unit]=unit_test [dreamverse]=dreamverse_app
+            [kernel]=kernel_tests [unit]=unit_test [preprocessing]=preprocessing
+            [dreamverse]=dreamverse_app
            [ssim]=ssim [training]=training
            [lora-inference]=inference_lora [lora-training]=training_lora
            [lora-extraction]=lora_extraction
@@ -13,23 +13,16 @@ on:
        required: false
        default: false
        type: boolean
-  # Auto-rebuild the CUDA images when a repository-controlled image input
-  # changes on main. This includes the trusted SM89 kernel artifact's source,
-  # metadata/key helper, ABI dependency metadata, and build orchestration.
-  # Dreamverse (apps/dreamverse/docker/Dockerfile) and the ROCm Dockerfile stay
-  # manual-dispatch only.
+  # Auto-rebuild the CUDA images when their Dockerfile changes on main. The CUDA
+  # matrix is the only lane that builds from docker/Dockerfile, so a path-scoped
+  # push trigger is a sufficient change detector on its own -- no separate
+  # detect-changes/paths-filter job is needed now that there is a single
+  # in-scope Dockerfile. Dreamverse (apps/dreamverse/docker/Dockerfile) and the
+  # rocm Dockerfile stay manual-dispatch only.
  push:
    branches: [main]
    paths:
-      - '.dockerignore'
-      - '.github/workflows/_template-build-image.yml'
-      - '.github/workflows/infra-build-image.yml'
-      - '.gitmodules'
      - 'docker/Dockerfile'
-      - 'docker/uv-excludes'
-      - 'fastvideo-kernel/**'
-      - 'fastvideo/tests/modal/kernel_build_cache.py'
-      - 'pyproject.toml'


 permissions:
@@ -57,7 +50,7 @@ jobs:
  # 2.8.3 comes from the architecture-specific prebuilt releases.
  build-cuda-images:
    # Runs on a manual dispatch when build_cuda_matrix is set, or automatically
-    # on an in-scope main push (inputs are null on push). The
+    # on a push that changed docker/Dockerfile (inputs are null on push). The
    # repository guard keeps fork syncs from auto-building; manual dispatch
    # still works in forks.
    if: ${{ (github.event_name == 'push' && github.repository == 'hao-ai-lab/FastVideo') || github.event.inputs.build_cuda_matrix == 'true' }}
@@ -11,7 +11,9 @@ on:
      - 'requirements-mkdocs.txt'
      - 'scripts/check_docs_links.py'
      - '.github/workflows/infra-docs.yml'
-  pull_request:
+  # Run the trusted base-branch workflow so fork PRs can be skipped without
+  # waiting for maintainer approval.
+  pull_request_target:
    branches: [ main ]
    paths:
      - 'docs/**'
@@ -24,21 +26,19 @@ on:

 permissions:
  contents: read
-  pages: write
-  id-token: write
-
-concurrency:
-  group: "pages"
-  cancel-in-progress: false

 jobs:
  build:
+    # MkDocs executes repository code; only trusted same-repository PRs run it.
+    if: github.event_name == 'push' || github.event.pull_request.head.repo.full_name == github.repository
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
        uses: actions/checkout@v4
        with:
+          ref: ${{ github.event.pull_request.head.sha || github.sha }}
          fetch-depth: 0
+          persist-credentials: false

      - name: Setup Python
        uses: actions/setup-python@v5
@@ -52,6 +52,7 @@ jobs:
        run: uv pip install --system -r requirements-mkdocs.txt

      - name: Setup Pages
+        if: github.event_name == 'push'
        uses: actions/configure-pages@v4

      - name: Build documentation
@@ -61,17 +62,22 @@ jobs:
        run: python scripts/check_docs_links.py

      - name: Upload artifact
+        if: github.event_name == 'push'
        uses: actions/upload-pages-artifact@v3
        with:
          path: ./site

  deploy:
+    permissions:
+      pages: write
+      id-token: write
+    concurrency: pages
    environment:
      name: github-pages
      url: ${{ steps.deployment.outputs.page_url }}
    runs-on: ubuntu-latest
    needs: build
-    if: github.ref == 'refs/heads/main'
+    if: github.event_name == 'push'
    steps:
      - name: Deploy to GitHub Pages
        id: deployment
@@ -6,7 +6,6 @@ results/
 wandb/
 *.ipynb
 *.jpg
-!examples/dataset/lingbotworld2/image.jpg
 *.safetensors
 *.mp4
 *.png
@@ -35,7 +34,6 @@ env
 *.log
 weights/
 logs/
-/Z-Image/
 official_weights/
 converted_weights/

@@ -75,7 +73,6 @@ docs/distillation/examples/
 *.pkl

 # Reference videos (negations must come after the catch-all on line below)
-!fastvideo/tests/nightly/reference_video_*.mp4

 # Static images
 !docs/assets/images/**/*.png
@@ -9,7 +9,7 @@
 **FastVideo is a unified post-training and real-time inference framework for accelerated video generation.**

 ## NEWS
- `2026/06/23`: Release FastWan-QAD: 5s of Video generated in 1.8s E2E. See the [FastWan-QAD models](https://huggingface.co/FastVideo/FastWan-QAD-FP8-1.3B), [Attn-QAT training guide](https://haoailab.com/FastVideo/training/attn_qat/), and [blog](https://haoailab.com/blogs/fastwan-qad/).
+- `2026/06/23`: Release FastWan-QAD: 5s of Video generated in 1.8s E2E. [FastWan-QAD models](https://huggingface.co/FastVideo/FastWan-QAD-FP8-1.3B), check out the [Blog](https://haoailab.com/blogs/fastwan-qad/).
 - `2026/03/17`: Release demo: Into the Dreamverse: Vibe Directing in FastVideo, check out the [Blog](https://haoailab.com/blogs/dreamverse/).
 - `2026/03/13`: Release demo: Create a 5s 1080p Video in 4.5s with FastVideo on a Single GPU, check out the [Blog](https://haoailab.com/blogs/fastvideo_realtime_1080p/).
 - `2025/11/19`: Release [CausalWan2.2 I2V A14B Preview](https://huggingface.co/FastVideo/CausalWan2.2-I2V-A14B-Preview-Diffusers) models, [Blog](https://hao-ai-lab.github.io/blogs/fastvideo_causalwan_preview/) and [Inference Code!](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py).
@@ -39,17 +39,6 @@ export FASTVIDEO_GENERATION_SEGMENT_CAP="${FASTVIDEO_GENERATION_SEGMENT_CAP:-6}"
 export FASTVIDEO_PROMPT_AUTO_SLEEP_MS="${FASTVIDEO_PROMPT_AUTO_SLEEP_MS:-120}"
 export FASTVIDEO_PROMPT_AUTO_TIMEOUT_MS="${FASTVIDEO_PROMPT_AUTO_TIMEOUT_MS:-1800}"

-if [[ "${ENABLE_TORCH_COMPILE}" == "1" ]]; then
-  # Persist Inductor, AOTAutograd, and Triton artifacts across launches.
-  export DREAMVERSE_TORCH_COMPILE_CACHE_ROOT="${DREAMVERSE_TORCH_COMPILE_CACHE_ROOT:-${HOME}/.cache/dreamverse/torch_compile}"
-  export TORCHINDUCTOR_CACHE_DIR="${TORCHINDUCTOR_CACHE_DIR:-${DREAMVERSE_TORCH_COMPILE_CACHE_ROOT}/inductor}"
-  export TRITON_CACHE_DIR="${TRITON_CACHE_DIR:-${DREAMVERSE_TORCH_COMPILE_CACHE_ROOT}/triton}"
-  export TORCHINDUCTOR_FX_GRAPH_CACHE="${TORCHINDUCTOR_FX_GRAPH_CACHE:-1}"
-  export TORCHINDUCTOR_AUTOGRAD_CACHE="${TORCHINDUCTOR_AUTOGRAD_CACHE:-1}"
-  mkdir -p "${TORCHINDUCTOR_CACHE_DIR}" "${TRITON_CACHE_DIR}"
-  echo "[launch-demo] torch.compile cache: ${DREAMVERSE_TORCH_COMPILE_CACHE_ROOT}"
-fi
-
 cd "${DREAMVERSE_ROOT}"

 if ! command -v dreamverse-server >/dev/null 2>&1; then
@@ -13,9 +13,10 @@ test.describe('create inference job', () => {
  test('creates a T2V job and shows it in the queue', async ({ page }) => {
    await page.goto('/inference');

-    // The trigger opens a real menu on click, so this path works for touch,
-    // mouse, and keyboard users.
-    await page.getByRole('button', { name: /create job/i }).click();
+    // The "Create Job" button reveals a workload menu on hover; wait for the
+    // T2V item to become visible before clicking so the CSS hover transition
+    // can't race the click.
+    await page.getByRole('button', { name: /create job/i }).hover();
    const t2vItem = page.getByRole('menuitem', { name: /T2V/i });
    await expect(t2vItem).toBeVisible();
    await t2vItem.click();
@@ -4,7 +4,7 @@ import { API_BASE, skipWithoutMock } from './helpers';

 /**
 * Gallery page: the seeded completed inference job surfaces as a media tile
- * with playback controls or an explicit media-error fallback.
+ * (an <article> wrapping a <video>) captioned with its prompt.
 */
 test.describe('gallery', () => {
  skipWithoutMock();
@@ -30,15 +30,12 @@ test.describe('gallery', () => {
      page.getByRole('heading', { level: 1, name: 'Gallery' }),
    ).toBeVisible();

-    const tile = page.locator('article').filter({ hasText: completed!.prompt });
-    await expect(tile).toBeVisible();
-    await expect(
-      tile.locator('video').or(tile.getByText('Preview unavailable')),
-    ).toBeVisible();
+    // The completed job renders as an <article> containing a <video> tile.
+    const tile = page
+      .locator('article')
+      .filter({ has: page.locator('video') });
+    await expect(tile.first()).toBeVisible();

-    const video = tile.locator('video');
-    if (await video.isVisible()) {
-      await expect(video).toHaveAttribute('controls', '');
-    }
+    await expect(page.getByText(completed!.prompt)).toBeVisible();
  });
 });
@@ -42,74 +42,6 @@ test.describe('app shell', () => {
      await expect(
        page.getByRole('heading', { level: 1, name: section.title }),
      ).toBeVisible();
-      await expect(page.getByRole('main')).toHaveCount(1);
    }
  });
-
-  test('keeps navigation and content usable at responsive breakpoints', async ({
-    page,
-  }) => {
-    for (const width of [320, 375, 414, 768]) {
-      await page.setViewportSize({ width, height: 800 });
-      await page.goto('/inference');
-
-      const main = page.getByRole('main');
-      await expect(main).toBeVisible();
-      await expect(
-        page.getByRole('button', { name: /Create Job/i }),
-      ).toBeVisible();
-
-      const initialBox = await main.boundingBox();
-      expect(initialBox?.x).toBe(width < 768 ? 0 : 220);
-      expect(initialBox?.width).toBe(width < 768 ? width : width - 220);
-
-      const navigation = page.getByRole('navigation', {
-        name: 'Primary navigation',
-      });
-      if (width < 768) {
-        await expect(
-          page.getByRole('button', { name: 'Open navigation' }),
-        ).toBeVisible();
-        await page.getByRole('button', { name: 'Open navigation' }).click();
-      }
-      await expect(navigation).toBeVisible();
-      await navigation.getByRole('link', { name: 'Datasets' }).click();
-
-      await expect(page).toHaveURL(/\/datasets$/);
-      expect(
-        await page.evaluate(
-          () => document.documentElement.scrollWidth <= window.innerWidth,
-        ),
-      ).toBe(true);
-    }
-  });
-
-  test('uses full-width detail drawers on mobile', async ({ page }) => {
-    await page.setViewportSize({ width: 320, height: 800 });
-    await page.goto('/inference');
-
-    await page
-      .locator('article button[aria-pressed="false"]')
-      .first()
-      .click();
-    const jobDrawer = page.getByRole('dialog', { name: 'Job details' });
-    await expect(jobDrawer).toBeVisible();
-    expect(await jobDrawer.boundingBox()).toMatchObject({ x: 0, width: 320 });
-    await jobDrawer.getByRole('button', { name: 'Close' }).click();
-
-    await page.goto('/datasets');
-    await page
-      .locator('article button[aria-pressed="false"]')
-      .first()
-      .click();
-
-    const datasetDrawer = page.getByRole('dialog', {
-      name: /dataset details$/,
-    });
-    await expect(datasetDrawer).toBeVisible();
-    expect(await datasetDrawer.boundingBox()).toMatchObject({
-      x: 0,
-      width: 320,
-    });
-  });
 });
@@ -9,7 +9,6 @@
 			"version": "0.1.0",
 			"dependencies": {
 				"@radix-ui/react-dialog": "^1.1.0",
-				"@radix-ui/react-dropdown-menu": "^2.1.24",
 				"@radix-ui/react-label": "^2.1.8",
 				"@radix-ui/react-scroll-area": "^1.2.10",
 				"@radix-ui/react-select": "^2.2.6",
@@ -1937,183 +1936,6 @@
 				}
 			}
 		},
-		"node_modules/@radix-ui/react-dropdown-menu": {
-			"version": "2.1.24",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-dropdown-menu/-/react-dropdown-menu-2.1.24.tgz",
-			"integrity": "sha512-geq8l2rJkxvkXsT9RMgtUE3P8pITFpTsvYpbySi1IH4fZEABD/Gp85myayFgxk0ktljGMJnCbeFkyTusvSvv7g==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/primitive": "1.1.7",
-				"@radix-ui/react-compose-refs": "1.1.5",
-				"@radix-ui/react-context": "1.2.2",
-				"@radix-ui/react-id": "1.1.4",
-				"@radix-ui/react-menu": "2.1.24",
-				"@radix-ui/react-primitive": "2.1.10",
-				"@radix-ui/react-use-controllable-state": "1.2.6"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/primitive": {
-			"version": "1.1.7",
-			"resolved": "https://registry.npmjs.org/@radix-ui/primitive/-/primitive-1.1.7.tgz",
-			"integrity": "sha512-rqWnm76nYT8HoNNqEjpgJ7Pw/DrBj5iBTrmEPo6HTX5+VJyBNOqTdv4g89G63HuR5g0AaENoAcH7Is5fF2kZ8Q==",
-			"license": "MIT"
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/react-compose-refs": {
-			"version": "1.1.5",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-compose-refs/-/react-compose-refs-1.1.5.tgz",
-			"integrity": "sha512-+48PbAAbq3didjJxa+OaWY2ZwgAKsNiRGyeHKszblZMQ+kcpd9pAaT11cMkGEie0vsOi3QdeTE6d5Fe3Gn61kA==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/react-context": {
-			"version": "1.2.2",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-context/-/react-context-1.2.2.tgz",
-			"integrity": "sha512-RHCUGwKHDr0hDGg4X7ma4JG4/+12qxw8rkh5QKdDldlCvtja6nUx1Ef/8HVrJze81lEsgLQlqjzjGNHantgnQA==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/react-id": {
-			"version": "1.1.4",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-id/-/react-id-1.1.4.tgz",
-			"integrity": "sha512-TMQp2llA+RYn7JcjnrMnz7wN4pcVttPZnRZo52PLQsoLVKzNlVwUeHmfePgTgRluXFvlD3GD5g5MOVVTJCO0qA==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/react-primitive": {
-			"version": "2.1.10",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-primitive/-/react-primitive-2.1.10.tgz",
-			"integrity": "sha512-MucOnzh6hR5mid6VpkbglRAMYMjKLqRnGBbjXkzjK52fuQDd1qbkx78a5P40mkcnVXJdEVxm26E9OPAiUq7nBg==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-slot": "1.3.3"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/react-slot": {
-			"version": "1.3.3",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-slot/-/react-slot-1.3.3.tgz",
-			"integrity": "sha512-qx7oqnYbxnK9kYI9m317qmFmEgo6ywqWvbTogdj7cL9p3/yx4M48p7Rnw5z3H890cL/ow/EeWJsuTykeZVXP5Q==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-compose-refs": "1.1.5"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/react-use-controllable-state": {
-			"version": "1.2.6",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-controllable-state/-/react-use-controllable-state-1.2.6.tgz",
-			"integrity": "sha512-uEQJGT97ZA/TgP/Hydw47lHu+/vQj6z/0jA+WeTbK1o9Rx45GImjpD0tc3W5ad3D6XTSR6e1yEO0FvGq6WQfVQ==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/primitive": "1.1.7",
-				"@radix-ui/react-use-effect-event": "0.0.5",
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/react-use-effect-event": {
-			"version": "0.0.5",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-effect-event/-/react-use-effect-event-0.0.5.tgz",
-			"integrity": "sha512-7cshFL8HGS/7HEiHH+9kL9HBwp2sa9yX18Knwek6KYWmXwM7pegMgta2AXMQKI+rq3JnfSj9x8wYqFMTdG1Jgg==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-dropdown-menu/node_modules/@radix-ui/react-use-layout-effect": {
-			"version": "1.1.4",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-layout-effect/-/react-use-layout-effect-1.1.4.tgz",
-			"integrity": "sha512-K20DkRkUwDnxEYMBPcg3Y6voLkEy5p5QQmszZgLngKKiC7dzBR/aEuK3w1qlx2JWDUNH6FluahYdgR3BP+QbYw==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
 		"node_modules/@radix-ui/react-focus-guards": {
 			"version": "1.1.4",
 			"resolved": "https://registry.npmjs.org/@radix-ui/react-focus-guards/-/react-focus-guards-1.1.4.tgz",
@@ -2195,494 +2017,6 @@
 				}
 			}
 		},
-		"node_modules/@radix-ui/react-menu": {
-			"version": "2.1.24",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-menu/-/react-menu-2.1.24.tgz",
-			"integrity": "sha512-uW7RVuU6Lp/ZtfeY4b3kL32zccgEWvPv1+cf17ubYzHa9cL8AHokmk36cG/XEiH/smbQvumnieXX9j/e9RqJWA==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/primitive": "1.1.7",
-				"@radix-ui/react-collection": "1.1.15",
-				"@radix-ui/react-compose-refs": "1.1.5",
-				"@radix-ui/react-context": "1.2.2",
-				"@radix-ui/react-direction": "1.1.4",
-				"@radix-ui/react-dismissable-layer": "1.1.19",
-				"@radix-ui/react-focus-guards": "1.1.6",
-				"@radix-ui/react-focus-scope": "1.1.16",
-				"@radix-ui/react-id": "1.1.4",
-				"@radix-ui/react-popper": "1.3.7",
-				"@radix-ui/react-portal": "1.1.17",
-				"@radix-ui/react-presence": "1.1.10",
-				"@radix-ui/react-primitive": "2.1.10",
-				"@radix-ui/react-roving-focus": "1.1.19",
-				"@radix-ui/react-slot": "1.3.3",
-				"@radix-ui/react-use-callback-ref": "1.1.4",
-				"aria-hidden": "^1.2.4",
-				"react-remove-scroll": "^2.7.2"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/primitive": {
-			"version": "1.1.7",
-			"resolved": "https://registry.npmjs.org/@radix-ui/primitive/-/primitive-1.1.7.tgz",
-			"integrity": "sha512-rqWnm76nYT8HoNNqEjpgJ7Pw/DrBj5iBTrmEPo6HTX5+VJyBNOqTdv4g89G63HuR5g0AaENoAcH7Is5fF2kZ8Q==",
-			"license": "MIT"
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-arrow": {
-			"version": "1.1.15",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-arrow/-/react-arrow-1.1.15.tgz",
-			"integrity": "sha512-v4zggRcjadnI+ClKDuijlQEW4tw3NoaeHc/PwpKnLoLLKNUG4InLegkstooLcRIUWCs+8L22dGURCVuFfOKfnA==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-primitive": "2.1.10"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-collection": {
-			"version": "1.1.15",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-collection/-/react-collection-1.1.15.tgz",
-			"integrity": "sha512-9W+B9NPF0NaaPh/1NJd3+KqsnlLqU9H7T2rvww+fp+T/evVXdNAyYcnfRQZFOjkR1ajQp3yORlqnI8soawLvNA==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-compose-refs": "1.1.5",
-				"@radix-ui/react-context": "1.2.2",
-				"@radix-ui/react-primitive": "2.1.10",
-				"@radix-ui/react-slot": "1.3.3"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-compose-refs": {
-			"version": "1.1.5",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-compose-refs/-/react-compose-refs-1.1.5.tgz",
-			"integrity": "sha512-+48PbAAbq3didjJxa+OaWY2ZwgAKsNiRGyeHKszblZMQ+kcpd9pAaT11cMkGEie0vsOi3QdeTE6d5Fe3Gn61kA==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-context": {
-			"version": "1.2.2",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-context/-/react-context-1.2.2.tgz",
-			"integrity": "sha512-RHCUGwKHDr0hDGg4X7ma4JG4/+12qxw8rkh5QKdDldlCvtja6nUx1Ef/8HVrJze81lEsgLQlqjzjGNHantgnQA==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-direction": {
-			"version": "1.1.4",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-direction/-/react-direction-1.1.4.tgz",
-			"integrity": "sha512-5pzg4FGQNpExhnhT2zlrP1wZFaYCd1K0nYWoFAdcYoYK868IEigqMX3B3f8yIoRlAhAeDWciLI6ZdCKHF9P4Vg==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-dismissable-layer": {
-			"version": "1.1.19",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-dismissable-layer/-/react-dismissable-layer-1.1.19.tgz",
-			"integrity": "sha512-8g4pfOL9HoKKLWGiypT+dphVqjFfmcXO5GBnhsG6zI+lxAx/8feQpr+1LSN8Re3hiZ+XkLNS4O9ztK11/LzQ6w==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/primitive": "1.1.7",
-				"@radix-ui/react-compose-refs": "1.1.5",
-				"@radix-ui/react-primitive": "2.1.10",
-				"@radix-ui/react-use-callback-ref": "1.1.4",
-				"@radix-ui/react-use-effect-event": "0.0.5"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-focus-guards": {
-			"version": "1.1.6",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-focus-guards/-/react-focus-guards-1.1.6.tgz",
-			"integrity": "sha512-RNOJjfZMTyBM6xYmV3IVGXkPjIhcBAuv48POevAXwrGJhkWZ9p1rFoIS1JFooPuT193AZmRsCPhpoVJxx6OPoQ==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-focus-scope": {
-			"version": "1.1.16",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-focus-scope/-/react-focus-scope-1.1.16.tgz",
-			"integrity": "sha512-wmRZ2WWLvmt6KHy2rNPOdPUjwq5xOHY02+m+udwJTn0aNIox/rkskAvJTyTLGhPK6KgrUjlJUJpgmx/+wFiFIQ==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-compose-refs": "1.1.5",
-				"@radix-ui/react-primitive": "2.1.10",
-				"@radix-ui/react-use-callback-ref": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-id": {
-			"version": "1.1.4",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-id/-/react-id-1.1.4.tgz",
-			"integrity": "sha512-TMQp2llA+RYn7JcjnrMnz7wN4pcVttPZnRZo52PLQsoLVKzNlVwUeHmfePgTgRluXFvlD3GD5g5MOVVTJCO0qA==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-popper": {
-			"version": "1.3.7",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-popper/-/react-popper-1.3.7.tgz",
-			"integrity": "sha512-UsJrrd7w4wuKKTdvd/DNERVlwSlUcyXzjhyDwBk+3aPOsCjOY6ZSbxuw8E6lZTjjfP8Cpd0J8VVkrYUWyGYXyg==",
-			"license": "MIT",
-			"dependencies": {
-				"@floating-ui/react-dom": "^2.0.0",
-				"@radix-ui/react-arrow": "1.1.15",
-				"@radix-ui/react-compose-refs": "1.1.5",
-				"@radix-ui/react-context": "1.2.2",
-				"@radix-ui/react-primitive": "2.1.10",
-				"@radix-ui/react-use-callback-ref": "1.1.4",
-				"@radix-ui/react-use-layout-effect": "1.1.4",
-				"@radix-ui/react-use-rect": "1.1.4",
-				"@radix-ui/react-use-size": "1.1.4",
-				"@radix-ui/rect": "1.1.3"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-portal": {
-			"version": "1.1.17",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-portal/-/react-portal-1.1.17.tgz",
-			"integrity": "sha512-vKQLcWypUnwZVvfV7UkGahH2g6ySe8M8R+zYBwPrv5byZ9QAW6cQVvNKo7GgmD+p8aYb6D9JBuvy8/WhOno2wQ==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-primitive": "2.1.10",
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-presence": {
-			"version": "1.1.10",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-presence/-/react-presence-1.1.10.tgz",
-			"integrity": "sha512-3wyzCQ6+ubRA+D4uv9m95JYLXxmOHp05qjrkjeA7uKHHtjpPggQzc6DAb0URl7j67oR0K2foO4ip27TiX037Bw==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-primitive": {
-			"version": "2.1.10",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-primitive/-/react-primitive-2.1.10.tgz",
-			"integrity": "sha512-MucOnzh6hR5mid6VpkbglRAMYMjKLqRnGBbjXkzjK52fuQDd1qbkx78a5P40mkcnVXJdEVxm26E9OPAiUq7nBg==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-slot": "1.3.3"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-roving-focus": {
-			"version": "1.1.19",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-roving-focus/-/react-roving-focus-1.1.19.tgz",
-			"integrity": "sha512-V9jI6hDjT7l3jsCQD9bLNvDLM3tH/gdbOTp7Tefp3hbbgCGQoK7tUvrWiRlcoBHIZ809ElXwNQwVo0B98LuTXQ==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/primitive": "1.1.7",
-				"@radix-ui/react-collection": "1.1.15",
-				"@radix-ui/react-compose-refs": "1.1.5",
-				"@radix-ui/react-context": "1.2.2",
-				"@radix-ui/react-direction": "1.1.4",
-				"@radix-ui/react-id": "1.1.4",
-				"@radix-ui/react-primitive": "2.1.10",
-				"@radix-ui/react-use-callback-ref": "1.1.4",
-				"@radix-ui/react-use-controllable-state": "1.2.6",
-				"@radix-ui/react-use-is-hydrated": "0.1.3",
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"@types/react-dom": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc",
-				"react-dom": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				},
-				"@types/react-dom": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-slot": {
-			"version": "1.3.3",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-slot/-/react-slot-1.3.3.tgz",
-			"integrity": "sha512-qx7oqnYbxnK9kYI9m317qmFmEgo6ywqWvbTogdj7cL9p3/yx4M48p7Rnw5z3H890cL/ow/EeWJsuTykeZVXP5Q==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-compose-refs": "1.1.5"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-use-callback-ref": {
-			"version": "1.1.4",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-callback-ref/-/react-use-callback-ref-1.1.4.tgz",
-			"integrity": "sha512-R6OUY2e2fA6Yn6s+VSx5KBV6Nx8LQEhu+cz7LCej18rQ1HLyg9PSC9jP/ZNx0o6FAIK9c0F1kHylzSxKsdlkrQ==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-use-controllable-state": {
-			"version": "1.2.6",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-controllable-state/-/react-use-controllable-state-1.2.6.tgz",
-			"integrity": "sha512-uEQJGT97ZA/TgP/Hydw47lHu+/vQj6z/0jA+WeTbK1o9Rx45GImjpD0tc3W5ad3D6XTSR6e1yEO0FvGq6WQfVQ==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/primitive": "1.1.7",
-				"@radix-ui/react-use-effect-event": "0.0.5",
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-use-effect-event": {
-			"version": "0.0.5",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-effect-event/-/react-use-effect-event-0.0.5.tgz",
-			"integrity": "sha512-7cshFL8HGS/7HEiHH+9kL9HBwp2sa9yX18Knwek6KYWmXwM7pegMgta2AXMQKI+rq3JnfSj9x8wYqFMTdG1Jgg==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-use-layout-effect": {
-			"version": "1.1.4",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-layout-effect/-/react-use-layout-effect-1.1.4.tgz",
-			"integrity": "sha512-K20DkRkUwDnxEYMBPcg3Y6voLkEy5p5QQmszZgLngKKiC7dzBR/aEuK3w1qlx2JWDUNH6FluahYdgR3BP+QbYw==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-use-rect": {
-			"version": "1.1.4",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-rect/-/react-use-rect-1.1.4.tgz",
-			"integrity": "sha512-cSOCh6JlkmfjLyNcLiu2nB4v+nm+dkZ+Q5KHWk/soo4U7ZLiEQFKHK9/YmtBHjfCEaU43IBKQOc4/uJmCaiCTQ==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/rect": "1.1.3"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/react-use-size": {
-			"version": "1.1.4",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-size/-/react-use-size-1.1.4.tgz",
-			"integrity": "sha512-D3anSY15EJoxrihpsXI6SMrmmonnQtR2ni7arO+Lfdg3O95b9hNXxONk8jA5C8ANdF/h5HMAxejgs8PWJ6rlhw==",
-			"license": "MIT",
-			"dependencies": {
-				"@radix-ui/react-use-layout-effect": "1.1.4"
-			},
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
-		"node_modules/@radix-ui/react-menu/node_modules/@radix-ui/rect": {
-			"version": "1.1.3",
-			"resolved": "https://registry.npmjs.org/@radix-ui/rect/-/rect-1.1.3.tgz",
-			"integrity": "sha512-JtyZR+mqgBibTo8xea3B6ZRmzZiM/YeVBtUkas6zMuXjAlfIFIW2FgqeM9eLyvEaYX66vr6DJMK+4U6LV0KhNw==",
-			"license": "MIT"
-		},
 		"node_modules/@radix-ui/react-popper": {
 			"version": "1.3.2",
 			"resolved": "https://registry.npmjs.org/@radix-ui/react-popper/-/react-popper-1.3.2.tgz",
@@ -3076,21 +2410,6 @@
 				}
 			}
 		},
-		"node_modules/@radix-ui/react-use-is-hydrated": {
-			"version": "0.1.3",
-			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-is-hydrated/-/react-use-is-hydrated-0.1.3.tgz",
-			"integrity": "sha512-umO/aJ+82CpOnhDZUTbILCQf7kU/g0iv+oGs/Q8jw7IkhWBzaEP4sA268PhFAJTFetbwp3ICc6ktpI4TqtxcIw==",
-			"license": "MIT",
-			"peerDependencies": {
-				"@types/react": "*",
-				"react": "^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc"
-			},
-			"peerDependenciesMeta": {
-				"@types/react": {
-					"optional": true
-				}
-			}
-		},
 		"node_modules/@radix-ui/react-use-layout-effect": {
 			"version": "1.1.2",
 			"resolved": "https://registry.npmjs.org/@radix-ui/react-use-layout-effect/-/react-use-layout-effect-1.1.2.tgz",
@@ -18,7 +18,6 @@
 	},
 	"dependencies": {
 		"@radix-ui/react-dialog": "^1.1.0",
-		"@radix-ui/react-dropdown-menu": "^2.1.24",
 		"@radix-ui/react-label": "^2.1.8",
 		"@radix-ui/react-scroll-area": "^1.2.10",
 		"@radix-ui/react-select": "^2.2.6",
@@ -1,64 +0,0 @@
-import { act, fireEvent, render, screen } from '@testing-library/react';
-import { describe, expect, it, vi } from 'vitest';
-
-import { HeaderActionsProvider } from '@/components/shell/HeaderActionsContext';
-import { getDatasets, type Dataset } from '@/lib/api';
-
-import DatasetsPage from './page';
-
-vi.mock('@/lib/api', () => ({
-  getDatasets: vi.fn(),
-}));
-
-vi.mock('@/components/datasets/AddDatasetButton', () => ({
-  default: () => null,
-}));
-
-vi.mock('@/components/datasets/CreateDatasetModal', () => ({
-  default: () => null,
-}));
-
-vi.mock('@/components/datasets/DatasetCard', () => ({
-  default: ({ dataset }: { dataset: Dataset }) => <div>{dataset.name}</div>,
-}));
-
-function renderPage() {
-  return render(
-    <HeaderActionsProvider>
-      <DatasetsPage />
-    </HeaderActionsProvider>,
-  );
-}
-
-describe('DatasetsPage', () => {
-  it('shows loading content before the initial request settles', async () => {
-    let resolveDatasets: (datasets: Dataset[]) => void = () => {};
-    vi.mocked(getDatasets).mockReturnValue(
-      new Promise<Dataset[]>((resolve) => {
-        resolveDatasets = resolve;
-      }),
-    );
-
-    renderPage();
-    expect(screen.getByLabelText('Loading datasets')).toBeInTheDocument();
-    expect(screen.queryByText('No datasets yet.')).not.toBeInTheDocument();
-
-    act(() => resolveDatasets([]));
-    expect(await screen.findByText('No datasets yet.')).toBeInTheDocument();
-  });
-
-  it('shows API failures separately from an empty list and retries', async () => {
-    vi.spyOn(console, 'error').mockImplementation(() => {});
-    vi.mocked(getDatasets).mockRejectedValueOnce(new Error('network down'));
-
-    renderPage();
-    expect(
-      await screen.findByText(/Could not load datasets from the Studio API/),
-    ).toBeInTheDocument();
-    expect(screen.queryByText('No datasets yet.')).not.toBeInTheDocument();
-
-    vi.mocked(getDatasets).mockResolvedValueOnce([]);
-    fireEvent.click(screen.getByRole('button', { name: 'Try Again' }));
-    expect(await screen.findByText('No datasets yet.')).toBeInTheDocument();
-  });
-});
@@ -1,14 +1,12 @@
 'use client';

 import * as React from 'react';
-import { AlertTriangle } from 'lucide-react';

 import AddDatasetButton from '@/components/datasets/AddDatasetButton';
 import CreateDatasetModal from '@/components/datasets/CreateDatasetModal';
 import DatasetCard from '@/components/datasets/DatasetCard';
 import { HeaderActions } from '@/components/shell/HeaderActionsContext';
 import { Card } from '@/components/ui/card';
-import { Button } from '@/components/ui/button';
 import { useStore } from '@/hooks/useStore';
 import { getDatasets } from '@/lib/api';
 import type { Dataset } from '@/lib/api';
@@ -23,28 +21,18 @@ import {

 export default function DatasetsPage() {
  const [datasets, setDatasets] = React.useState<Dataset[]>([]);
-  const [isInitialLoading, setIsInitialLoading] = React.useState(true);
  const [error, setError] = React.useState<string | null>(null);
  const { open } = useStore(createDatasetModalStore);
-  const fetchSequence = React.useRef(0);

  const fetchDatasets = React.useCallback(async () => {
-    const sequence = ++fetchSequence.current;
    try {
-      const next = await getDatasets();
-      if (sequence === fetchSequence.current) {
-        setDatasets(next);
-        setError(null);
-      }
+      setDatasets(await getDatasets());
+      setError(null);
    } catch (err) {
      console.error('Failed to fetch datasets:', err);
-      if (sequence === fetchSequence.current) {
-        setError(
-          'Could not load datasets from the Studio API. Check the server and try again.',
-        );
-      }
-    } finally {
-      if (sequence === fetchSequence.current) setIsInitialLoading(false);
+      // Distinguish an API outage from a genuinely empty list, so the user
+      // isn't told they have no datasets when the server is unreachable.
+      setError(err instanceof Error ? err.message : 'Failed to load datasets');
    }
  }, []);

@@ -62,67 +50,28 @@ export default function DatasetsPage() {
      <HeaderActions>
        <AddDatasetButton />
      </HeaderActions>
-      <div className="mx-auto flex w-full max-w-[850px] flex-col gap-6 px-4 pb-12">
+      <main className="mx-auto flex w-full max-w-[850px] flex-col gap-6 px-4 pb-12">
        <Card className="p-6">
-          <div aria-busy={isInitialLoading}>
-            {isInitialLoading ? (
-              <div
-                aria-label="Loading datasets"
-                className="flex flex-col gap-3 py-2"
-              >
-                {[0, 1, 2].map((item) => (
-                  <div
-                    key={item}
-                    className="h-24 animate-pulse rounded-lg border border-border bg-muted/50"
-                  />
-                ))}
-              </div>
-            ) : error && datasets.length === 0 ? (
-              <div
-                role="alert"
-                className="flex flex-col items-center gap-3 py-8 text-center"
-              >
-                <AlertTriangle
-                  className="size-6 text-destructive"
-                  aria-hidden
-                />
-                <p className="max-w-md text-sm text-muted-foreground">
-                  {error}
-                </p>
-                <Button type="button" variant="outline" onClick={fetchDatasets}>
-                  Try Again
-                </Button>
-              </div>
+          <div>
+            {error ? (
+              <p className="py-8 text-center text-destructive">{error}</p>
+            ) : datasets.length === 0 ? (
+              <p className="py-8 text-center text-muted-foreground">
+                No datasets yet.
+              </p>
            ) : (
-              <>
-                {error && (
-                  <p
-                    role="status"
-                    className="mb-3 rounded-lg border border-amber-500/50 bg-amber-500/10 px-3 py-2 text-sm text-foreground"
-                  >
-                    Dataset updates are temporarily unavailable. Showing the
-                    most recent results.
-                  </p>
-                )}
-                {datasets.length === 0 ? (
-                  <p className="py-8 text-center text-muted-foreground">
-                    No datasets yet.
-                  </p>
-                ) : (
-                  datasets.map((ds) => (
-                    <DatasetCard
-                      key={ds.id}
-                      dataset={ds}
-                      onUpdated={fetchDatasets}
-                      onSelect={() => handleSelectDataset(ds)}
-                    />
-                  ))
-                )}
-              </>
+              datasets.map((ds) => (
+                <DatasetCard
+                  key={ds.id}
+                  dataset={ds}
+                  onUpdated={fetchDatasets}
+                  onSelect={() => handleSelectDataset(ds)}
+                />
+              ))
            )}
          </div>
        </Card>
-      </div>
+      </main>
      <CreateDatasetModal
        isOpen={open}
        onClose={() => setCreateDatasetModalOpen(false)}
@@ -1,4 +1,4 @@
-import { fireEvent, render, screen } from '@testing-library/react';
+import { render, screen } from '@testing-library/react';
 import { describe, expect, it, vi } from 'vitest';

 import GalleryPage from './page';
@@ -43,22 +43,6 @@ describe('GalleryPage', () => {
    expect(getJobsList).toHaveBeenCalledWith('inference');
  });

-  it('provides video controls and a visible fallback when media fails', async () => {
-    vi.mocked(getJobsList).mockResolvedValue([makeJob()]);
-    renderGallery();
-
-    const video = await screen.findByLabelText(
-      'Generated video: a cat surfing a wave',
-    );
-    expect(video).toHaveAttribute('controls');
-
-    fireEvent.error(video);
-    expect(screen.getByText('Preview unavailable')).toBeInTheDocument();
-    expect(
-      screen.getByText('The generated file could not be loaded.'),
-    ).toBeInTheDocument();
-  });
-
  it('shows the empty state when no completed videos exist', async () => {
    vi.mocked(getJobsList).mockResolvedValue([
      makeJob({ status: 'running', output_path: null }),
@@ -1,6 +1,6 @@
 'use client';

-import { ImageOff, Loader2 } from 'lucide-react';
+import { Loader2 } from 'lucide-react';
 import { useEffect, useState } from 'react';

 import { Card } from '@/components/ui/card';
@@ -11,54 +11,6 @@ function isImage(job: Job): boolean {
  return job.output_path?.toLowerCase().endsWith('.png') ?? false;
 }

-function GalleryMedia({ job }: { job: Job }) {
-  const [failed, setFailed] = useState(false);
-
-  if (failed) {
-    return (
-      <div
-        role="status"
-        className="flex h-full flex-col items-center justify-center gap-2 px-4 text-center text-muted-foreground"
-      >
-        <ImageOff className="size-7" aria-hidden />
-        <span className="text-sm font-medium">Preview unavailable</span>
-        <span className="text-xs">
-          The generated file could not be loaded.
-        </span>
-      </div>
-    );
-  }
-
-  if (isImage(job)) {
-    return (
-      // eslint-disable-next-line @next/next/no-img-element
-      <img
-        src={getJobVideoUrl(job.id)}
-        alt={job.prompt}
-        className="block h-full w-full object-contain"
-        loading="lazy"
-        onError={() => setFailed(true)}
-      />
-    );
-  }
-
-  return (
-    <video
-      src={getJobVideoUrl(job.id)}
-      aria-label={
-        job.prompt ? `Generated video: ${job.prompt}` : 'Generated video'
-      }
-      className="block h-full w-full object-contain"
-      controls
-      muted
-      loop
-      playsInline
-      preload="metadata"
-      onError={() => setFailed(true)}
-    />
-  );
-}
-
 export default function GalleryPage() {
  const [jobs, setJobs] = useState<Job[]>([]);
  const [isLoading, setIsLoading] = useState(true);
@@ -125,7 +77,24 @@ export default function GalleryPage() {
                className="flex flex-col overflow-hidden rounded-lg border border-border bg-background"
              >
                <div className="relative aspect-video overflow-hidden bg-muted">
-                  <GalleryMedia job={job} />
+                  {isImage(job) ? (
+                    // eslint-disable-next-line @next/next/no-img-element
+                    <img
+                      src={getJobVideoUrl(job.id)}
+                      alt={job.prompt}
+                      className="block h-full w-full object-contain"
+                      loading="lazy"
+                    />
+                  ) : (
+                    <video
+                      src={getJobVideoUrl(job.id)}
+                      className="block h-full w-full object-contain"
+                      muted
+                      loop
+                      playsInline
+                      preload="metadata"
+                    />
+                  )}
                </div>
                <p
                  className="line-clamp-3 border-t border-border px-4 py-3 text-sm text-muted-foreground"
@@ -41,7 +41,7 @@

 	--border: #e2e8f0;
 	--input: #cbd5e1;
-	--ring: #1d4ed8;
+	--ring: #94a3b8;

 	--radius: 0.5rem;
 }
@@ -77,7 +77,7 @@

 	--border: #334155;
 	--input: #334155;
-	--ring: #7dd3fc;
+	--ring: #cbd5e1;
 }

@theme inline {
@@ -125,7 +125,6 @@
 html,
 body {
 	min-height: 100%;
-	overflow-x: clip;
 }

 html {
@@ -164,22 +163,6 @@ a {
 	color: inherit;
 }

-:where(
-	a,
-	button,
-	input,
-	textarea,
-	select,
-	summary,
-	[role="button"],
-	[role="menuitem"],
-	[role="slider"],
-	[tabindex]
-):focus-visible {
-	outline: 3px solid var(--ring) !important;
-	outline-offset: 2px !important;
-}
-
 summary {
 	list-style: none;
 }
@@ -1,55 +0,0 @@
-import { readFileSync } from 'node:fs';
-import { join } from 'node:path';
-import { describe, expect, it } from 'vitest';
-
-const css = readFileSync(join(process.cwd(), 'src/app/globals.css'), 'utf8');
-
-function token(block: string, name: string): string {
-  const match = block.match(new RegExp(`--${name}:\\s*(#[0-9a-fA-F]{6})`));
-  if (!match) throw new Error(`Missing --${name} token`);
-  return match[1];
-}
-
-function luminance(hex: string): number {
-  const channels = hex
-    .slice(1)
-    .match(/.{2}/g)!
-    .map((channel) => parseInt(channel, 16) / 255)
-    .map((channel) =>
-      channel <= 0.04045
-        ? channel / 12.92
-        : ((channel + 0.055) / 1.055) ** 2.4,
-    );
-  return (
-    0.2126 * channels[0] + 0.7152 * channels[1] + 0.0722 * channels[2]
-  );
-}
-
-function contrast(first: string, second: string): number {
-  const firstLuminance = luminance(first);
-  const secondLuminance = luminance(second);
-  return (
-    (Math.max(firstLuminance, secondLuminance) + 0.05) /
-    (Math.min(firstLuminance, secondLuminance) + 0.05)
-  );
-}
-
-describe('global focus styles', () => {
-  it('keeps focus tokens above 3:1 against both page themes', () => {
-    const light = css.match(/:root\s*{([\s\S]*?)\n}/)?.[1] ?? '';
-    const dark = css.match(/\.dark\s*{([\s\S]*?)\n}/)?.[1] ?? '';
-
-    expect(contrast(token(light, 'ring'), token(light, 'background'))).toBeGreaterThanOrEqual(
-      3,
-    );
-    expect(contrast(token(dark, 'ring'), token(dark, 'background'))).toBeGreaterThanOrEqual(
-      3,
-    );
-  });
-
-  it('applies a non-animated three-pixel outline to focus-visible controls', () => {
-    expect(css).toContain('):focus-visible {');
-    expect(css).toContain('outline: 3px solid var(--ring) !important;');
-    expect(css).toContain('outline-offset: 2px !important;');
-  });
-});
@@ -4,8 +4,8 @@ import GpuGrid from '@/components/system/GpuGrid';

 export default function GpusPage() {
  return (
-    <div className="mx-auto flex w-full max-w-[1100px] flex-col gap-6 px-4 pb-12 pt-6">
+    <main className="mx-auto flex w-full max-w-[1100px] flex-col gap-6 px-4 pb-12 pt-6">
      <GpuGrid />
-    </div>
+    </main>
  );
 }
@@ -52,20 +52,6 @@ describe('Settings page', () => {
    expect(updateOption).toHaveBeenCalledWith('numFrames', expect.any(Number));
  });

-  it('gives every slider an accessible name', () => {
-    renderPage();
-
-    const sliders = screen.getAllByRole('slider');
-    expect(sliders).toHaveLength(11);
-    for (const slider of sliders) {
-      expect(slider).toHaveAccessibleName();
-    }
-    expect(screen.getByRole('slider', { name: 'Frames' })).toBeInTheDocument();
-    expect(
-      screen.getByRole('slider', { name: 'Guidance Scale' }),
-    ).toBeInTheDocument();
-  });
-
  it('calls resetToDefaults when Reset to Defaults is clicked', () => {
    renderPage();
    fireEvent.click(screen.getByRole('button', { name: 'Reset to Defaults' }));
@@ -25,16 +25,6 @@ beforeEach(() => {
 });

 describe('DatasetCard', () => {
-  it('keeps selection and delete buttons as semantic siblings', () => {
-    render(<DatasetCard dataset={dataset} onUpdated={() => {}} />);
-
-    const selectButton = screen.getByRole('button', { pressed: false });
-    const deleteButton = screen.getByRole('button', { name: 'Delete' });
-
-    expect(selectButton).toHaveTextContent('My Dataset');
-    expect(selectButton).not.toContainElement(deleteButton);
-  });
-
  it('renders the name, file count and human-readable size', () => {
    render(<DatasetCard dataset={dataset} onUpdated={() => {}} />);
    expect(screen.getByText('My Dataset')).toBeInTheDocument();
@@ -81,7 +71,7 @@ describe('DatasetCard', () => {
    expect(onSelect).not.toHaveBeenCalled();
  });

-  it('keeps the selection and delete actions separate', () => {
+  it('selects on keyboard activation of the card body but not of the Delete button', () => {
    const onSelect = vi.fn();
    render(
      <DatasetCard dataset={dataset} onUpdated={() => {}} onSelect={onSelect} />,
@@ -93,12 +83,8 @@ describe('DatasetCard', () => {
    });
    expect(onSelect).not.toHaveBeenCalled();

-    // Activating the dedicated selection button selects the dataset.
-    fireEvent.click(
-      screen.getByRole('button', {
-        name: /My Dataset.*3 files.*2.0 KB/,
-      }),
-    );
+    // Activating the card body itself does select.
+    fireEvent.keyDown(screen.getByText('My Dataset'), { key: 'Enter' });
    expect(onSelect).toHaveBeenCalledTimes(1);
  });

@@ -55,33 +55,43 @@ export default function DatasetCard({
    }
  }

+  function handleKeyDown(e: React.KeyboardEvent) {
+    if ((e.target as HTMLElement).closest('button')) return;
+    if (e.key === 'Enter' || e.key === ' ') {
+      e.preventDefault();
+      onSelect();
+    }
+  }
+
  return (
-    <article
+    <div
      className={cn(
-        'mb-3 flex items-start gap-3 rounded-lg border border-border bg-background px-[1.15rem] py-4',
+        'mb-3 flex cursor-pointer flex-col gap-[0.6rem] rounded-lg border border-border bg-background px-[1.15rem] py-4',
        isSelected && 'border-accent-blue bg-accent-blue/5',
      )}
+      onClick={(e) => {
+        if ((e.target as HTMLElement).closest('button')) return;
+        onSelect();
+      }}
+      onKeyDown={handleKeyDown}
+      role="button"
+      tabIndex={0}
    >
-      <button
-        type="button"
-        aria-pressed={isSelected}
-        onClick={onSelect}
-        className="flex min-w-0 flex-1 cursor-pointer flex-col gap-[0.6rem] rounded-md text-left"
-      >
+      <div className="flex flex-wrap items-center justify-between gap-2">
        <span className="text-[0.95rem] font-semibold">{dataset.name}</span>
-        <span className="text-sm text-muted-foreground">
-          {fileCount} {fileCount === 1 ? 'file' : 'files'} · {sizeLabel}
-        </span>
-      </button>
-      <Button
-        type="button"
-        variant="destructive"
-        size="sm"
-        onClick={handleDelete}
-        disabled={isLoading}
-      >
-        Delete
-      </Button>
-    </article>
+        <Button
+          type="button"
+          variant="destructive"
+          size="sm"
+          onClick={handleDelete}
+          disabled={isLoading}
+        >
+          Delete
+        </Button>
+      </div>
+      <div className="text-sm text-muted-foreground">
+        {fileCount} {fileCount === 1 ? 'file' : 'files'} · {sizeLabel}
+      </div>
+    </div>
  );
 }
@@ -6,9 +6,6 @@ import * as api from '@/lib/api';
 import type { Dataset } from '@/lib/api';

 vi.mock('@/lib/api');
-vi.mock('sonner', () => ({
-  toast: { error: vi.fn() },
-}));

 const mockedApi = vi.mocked(api);

@@ -30,27 +27,6 @@ beforeEach(() => {
 });

 describe('DatasetSidebar', () => {
-  it('fills the mobile viewport without reserving main-content width', async () => {
-    const onWidthChange = vi.fn();
-
-    render(
-      <DatasetSidebar
-        dataset={dataset}
-        isMobile
-        onClose={() => {}}
-        onWidthChange={onWidthChange}
-      />,
-    );
-
-    const drawer = screen.getByRole('dialog', {
-      name: 'My Dataset dataset details',
-    });
-    expect(drawer).toHaveStyle({ width: '100%', maxWidth: 'none' });
-    expect(drawer).toHaveAttribute('aria-modal', 'true');
-    expect(drawer).toHaveFocus();
-    expect(onWidthChange).toHaveBeenCalledWith(0);
-  });
-
  it('lists dataset files after loading', async () => {
    render(<DatasetSidebar dataset={dataset} onClose={() => {}} />);

@@ -67,16 +43,6 @@ describe('DatasetSidebar', () => {
    expect(mockedApi.getDatasetMediaUrl).toHaveBeenCalledWith('ds-1', 'b.mp4');
  });

-  it('shows a fallback when a dataset preview cannot load', async () => {
-    render(<DatasetSidebar dataset={dataset} onClose={() => {}} />);
-
-    const preview = await screen.findByLabelText('Preview of a.mp4');
-    fireEvent.error(preview);
-
-    expect(screen.getByText('Preview unavailable')).toBeInTheDocument();
-    expect(screen.queryByLabelText('Preview of a.mp4')).not.toBeInTheDocument();
-  });
-
  it('debounces caption save by 500ms', async () => {
    render(<DatasetSidebar dataset={dataset} onClose={() => {}} />);
    const textarea = await screen.findByDisplayValue('cap a');
@@ -133,39 +99,6 @@ describe('DatasetSidebar', () => {
    }
  });

-  it('shows a failed save and lets the user retry it', async () => {
-    vi.spyOn(console, 'error').mockImplementation(() => {});
-    mockedApi.updateDatasetCaption
-      .mockRejectedValueOnce(new Error('network down'))
-      .mockResolvedValueOnce(undefined);
-    render(<DatasetSidebar dataset={dataset} onClose={() => {}} />);
-    const textarea = await screen.findByDisplayValue('cap a');
-
-    vi.useFakeTimers();
-    try {
-      fireEvent.change(textarea, { target: { value: 'needs retry' } });
-      await act(async () => {
-        await vi.advanceTimersByTimeAsync(500);
-      });
-
-      expect(screen.getByText(/Not saved/)).toBeInTheDocument();
-      fireEvent.click(screen.getByRole('button', { name: 'Retry' }));
-      await act(async () => {
-        await Promise.resolve();
-      });
-
-      expect(mockedApi.updateDatasetCaption).toHaveBeenCalledTimes(2);
-      expect(mockedApi.updateDatasetCaption).toHaveBeenLastCalledWith(
-        'ds-1',
-        'a.mp4',
-        'needs retry',
-      );
-      expect(screen.getByText('Saved')).toBeInTheDocument();
-    } finally {
-      vi.useRealTimers();
-    }
-  });
-
  it('debounces per file: editing another caption does not cancel a pending save', async () => {
    render(<DatasetSidebar dataset={dataset} onClose={() => {}} />);
    await screen.findByDisplayValue('cap a');
@@ -1,8 +1,7 @@
 'use client';

 import * as React from 'react';
-import { ImageOff, X } from 'lucide-react';
-import { toast } from 'sonner';
+import { X } from 'lucide-react';

 import DownloadCaptions from '@/components/datasets/DownloadCaptions';
 import { Textarea } from '@/components/ui/textarea';
@@ -14,14 +13,12 @@ import {
  type Dataset,
 } from '@/lib/api';
 import { cn } from '@/lib/utils';
-import { useDrawerFocus } from '@/hooks/useDrawerFocus';

 const SIDEBAR_MIN_WIDTH = 320;
 const SIDEBAR_MAX_WIDTH = 900;
 const INITIAL_PAGE_SIZE = 24;
 const PAGE_SIZE = 24;
 const SCROLL_THRESHOLD = 200;
-type CaptionSaveState = 'idle' | 'saving' | 'saved' | 'error';

 // Memoized so a caption keystroke re-renders only the edited card, not every
 // visible <video> in the grid (visibleCount grows unbounded with scrolling).
@@ -30,101 +27,54 @@ const DatasetFileCard = React.memo(function DatasetFileCard({
  mediaUrl,
  caption,
  thumbLoaded,
-  saveState,
  onCaptionChange,
-  onCaptionRetry,
  onThumbLoaded,
 }: {
  fileName: string;
  mediaUrl: string;
  caption: string;
  thumbLoaded: boolean;
-  saveState: CaptionSaveState;
  onCaptionChange: (fileName: string, value: string) => void;
-  onCaptionRetry: (fileName: string, value: string) => void;
  onThumbLoaded: (fileName: string) => void;
 }) {
-  const [mediaFailed, setMediaFailed] = React.useState(false);
-
-  React.useEffect(() => {
-    setMediaFailed(false);
-  }, [mediaUrl]);
-
  return (
    <div className="relative flex flex-col overflow-hidden rounded-lg border border-border bg-background">
-      {!thumbLoaded && !mediaFailed && (
+      {!thumbLoaded && (
        <div className="pointer-events-none absolute inset-0 flex items-center justify-center bg-background/70">
          <div className="h-6 w-6 animate-spin rounded-full border-2 border-muted-foreground/40 border-t-accent-blue" />
        </div>
      )}
-      {mediaFailed ? (
-        <div
-          role="status"
-          className="flex aspect-video w-full flex-col items-center justify-center gap-1 bg-muted px-2 text-center text-muted-foreground"
-        >
-          <ImageOff className="size-5" aria-hidden />
-          <span className="text-xs">Preview unavailable</span>
-        </div>
-      ) : (
-        // eslint-disable-next-line jsx-a11y/media-has-caption
-        <video
-          src={mediaUrl}
-          aria-label={`Preview of ${fileName}`}
-          className="aspect-video w-full bg-border object-cover"
-          muted
-          autoPlay
-          loop
-          playsInline
-          onLoadedData={() => onThumbLoaded(fileName)}
-          onError={() => {
-            setMediaFailed(true);
-            onThumbLoaded(fileName);
-          }}
-        />
-      )}
+      {/* eslint-disable-next-line jsx-a11y/media-has-caption */}
+      <video
+        src={mediaUrl}
+        className="aspect-video w-full bg-border object-cover"
+        muted
+        autoPlay
+        loop
+        playsInline
+        onLoadedData={() => onThumbLoaded(fileName)}
+        onError={() => onThumbLoaded(fileName)}
+      />
      <Textarea
-        aria-label={`Caption for ${fileName}`}
        value={caption}
        onChange={(e) => onCaptionChange(fileName, e.target.value)}
        placeholder="Caption"
        rows={2}
        className="min-h-[2.5rem] resize-y rounded-none border-0 bg-transparent p-1.5 text-xs shadow-none focus-visible:border-transparent focus-visible:ring-0"
      />
-      <div
-        aria-live="polite"
-        className="flex min-h-6 items-center px-1.5 pb-1 text-[0.7rem] text-muted-foreground"
-      >
-        {saveState === 'saving' && <span>Saving…</span>}
-        {saveState === 'saved' && <span>Saved</span>}
-        {saveState === 'error' && (
-          <span role="alert" className="text-destructive">
-            Not saved.{' '}
-            <button
-              type="button"
-              onClick={() => onCaptionRetry(fileName, caption)}
-              className="inline-flex min-h-11 items-center font-medium underline underline-offset-2"
-            >
-              Retry
-            </button>
-          </span>
-        )}
-      </div>
    </div>
  );
 });

 export default function DatasetSidebar({
  dataset,
-  isMobile = false,
  onClose,
  onWidthChange,
 }: {
  dataset: Dataset;
-  isMobile?: boolean;
  onClose: () => void;
  onWidthChange?: (w: number) => void;
 }) {
-  const drawerRef = useDrawerFocus<HTMLElement>(isMobile);
  const [width, setWidth] = React.useState(400);
  const [isDragging, setIsDragging] = React.useState(false);
  const [fileNames, setFileNames] = React.useState<string[]>([]);
@@ -134,21 +84,17 @@ export default function DatasetSidebar({
  const [thumbLoaded, setThumbLoaded] = React.useState<
    Record<string, boolean>
  >({});
-  const [captionSaveStates, setCaptionSaveStates] = React.useState<
-    Record<string, CaptionSaveState>
-  >({});

  // Pending debounced caption saves, keyed per file so editing one caption
  // can't cancel another file's pending save.
  const pendingSaves = React.useRef(
    new Map<string, { timer: ReturnType<typeof setTimeout>; save: () => void }>(),
  );
-  const captionVersions = React.useRef(new Map<string, number>());
  const scrollRef = React.useRef<HTMLDivElement>(null);

  React.useEffect(() => {
-    onWidthChange?.(isMobile ? 0 : width);
-  }, [isMobile, width, onWidthChange]);
+    onWidthChange?.(width);
+  }, [width, onWidthChange]);

  React.useEffect(() => {
    let cancelled = false;
@@ -160,8 +106,6 @@ export default function DatasetSidebar({
        setCaptions(data.captions);
        setVisibleCount(INITIAL_PAGE_SIZE);
        setThumbLoaded({});
-        setCaptionSaveStates({});
-        captionVersions.current.clear();
      })
      .catch((err) => console.error('Failed to load dataset files:', err))
      .finally(() => {
@@ -194,51 +138,23 @@ export default function DatasetSidebar({
  });

  const datasetId = dataset.id;
-  const persistCaption = React.useCallback(
-    (fileName: string, value: string, version: number) => {
-      setCaptionSaveStates((prev) => ({ ...prev, [fileName]: 'saving' }));
-      void updateDatasetCaption(datasetId, fileName, value)
-        .then(() => {
-          if (captionVersions.current.get(fileName) !== version) return;
-          setCaptionSaveStates((prev) => ({ ...prev, [fileName]: 'saved' }));
-        })
-        .catch((error) => {
-          if (captionVersions.current.get(fileName) !== version) return;
-          console.error('Failed to save caption:', error);
-          setCaptionSaveStates((prev) => ({ ...prev, [fileName]: 'error' }));
-          toast.error('Caption was not saved', {
-            description: `${fileName}: check the Studio API, then retry.`,
-          });
-        });
-    },
-    [datasetId],
-  );
-
  const handleCaptionChange = React.useCallback(
    (fileName: string, value: string) => {
      setCaptions((prev) => ({ ...prev, [fileName]: value }));
-      setCaptionSaveStates((prev) => ({ ...prev, [fileName]: 'idle' }));
      const pending = pendingSaves.current.get(fileName);
      if (pending) clearTimeout(pending.timer);
-      const version = (captionVersions.current.get(fileName) ?? 0) + 1;
-      captionVersions.current.set(fileName, version);
-      const save = () => persistCaption(fileName, value, version);
+      const save = () => {
+        updateDatasetCaption(datasetId, fileName, value).catch((err) =>
+          console.error('Failed to save caption:', err),
+        );
+      };
      const timer = setTimeout(() => {
        pendingSaves.current.delete(fileName);
        save();
      }, 500);
      pendingSaves.current.set(fileName, { timer, save });
    },
-    [persistCaption],
-  );
-
-  const handleCaptionRetry = React.useCallback(
-    (fileName: string, value: string) => {
-      const version = (captionVersions.current.get(fileName) ?? 0) + 1;
-      captionVersions.current.set(fileName, version);
-      persistCaption(fileName, value, version);
-    },
-    [persistCaption],
+    [datasetId],
  );

  function handleScroll() {
@@ -273,16 +189,8 @@ export default function DatasetSidebar({

  return (
    <aside
-      ref={drawerRef}
-      tabIndex={-1}
-      role="dialog"
-      aria-label={`${dataset.name} dataset details`}
-      aria-modal={isMobile || undefined}
-      className="fixed bottom-0 right-0 top-[var(--header-height)] z-50 flex max-h-[calc(100dvh-var(--header-height))] min-w-0 shrink-0 flex-col border-l border-border bg-card md:min-w-[320px]"
-      style={{
-        width: isMobile ? '100%' : width,
-        maxWidth: isMobile ? 'none' : SIDEBAR_MAX_WIDTH,
-      }}
+      className="fixed bottom-0 right-0 top-[var(--header-height)] z-50 flex max-h-[calc(100vh-var(--header-height))] min-w-[320px] shrink-0 flex-col border-l border-border bg-card"
+      style={{ width, maxWidth: SIDEBAR_MAX_WIDTH }}
    >
      <div className="flex shrink-0 items-center justify-between border-b border-border px-5 py-4">
        <h2 className="m-0 min-w-0 truncate text-base font-semibold text-foreground">
@@ -295,7 +203,7 @@ export default function DatasetSidebar({
            onClick={onClose}
            title="Close"
            aria-label="Close"
-            className="flex size-11 items-center justify-center rounded-lg text-muted-foreground transition-colors hover:bg-accent hover:text-foreground"
+            className="flex items-center justify-center rounded-lg p-1.5 text-muted-foreground transition-colors hover:bg-accent hover:text-foreground"
          >
            <X className="h-[18px] w-[18px]" />
          </button>
@@ -323,9 +231,7 @@ export default function DatasetSidebar({
                  mediaUrl={getDatasetMediaUrl(dataset.id, fileName)}
                  caption={captions[fileName] ?? ''}
                  thumbLoaded={!!thumbLoaded[fileName]}
-                  saveState={captionSaveStates[fileName] ?? 'idle'}
                  onCaptionChange={handleCaptionChange}
-                  onCaptionRetry={handleCaptionRetry}
                  onThumbLoaded={markThumbLoaded}
                />
              ))}
@@ -334,14 +240,14 @@ export default function DatasetSidebar({
        </div>
      </div>

-      {!isMobile && <div
+      <div
        role="presentation"
        onMouseDown={onMouseDown}
        className={cn(
          'absolute bottom-0 left-0 top-0 z-[1] w-1.5 cursor-col-resize hover:bg-accent-blue/25',
          isDragging && 'bg-accent-blue/25',
        )}
-      />}
+      />
    </aside>
  );
 }
@@ -1,50 +0,0 @@
-import { render, screen } from '@testing-library/react';
-import userEvent from '@testing-library/user-event';
-import { describe, expect, it, vi } from 'vitest';
-
-import { downloadBlob } from '@/lib/utils';
-
-import DownloadCaptions from './DownloadCaptions';
-
-vi.mock('@/lib/utils', async (importOriginal) => {
-  const actual = await importOriginal<typeof import('@/lib/utils')>();
-  return { ...actual, downloadBlob: vi.fn() };
-});
-
-const props = {
-  fileNames: ['b.mp4', 'a.mp4'],
-  captions: { 'a.mp4': 'cap a', 'b.mp4': 'cap b' },
-};
-
-describe('DownloadCaptions', () => {
-  it('opens the format menu on click and downloads the selection', async () => {
-    const user = userEvent.setup();
-    render(<DownloadCaptions {...props} />);
-
-    await user.click(
-      screen.getByRole('button', { name: 'Download Captions' }),
-    );
-    await user.click(screen.getByRole('menuitem', { name: 'JSON' }));
-
-    expect(vi.mocked(downloadBlob)).toHaveBeenCalledWith(
-      expect.any(Blob),
-      'videos2caption.json',
-    );
-  });
-
-  it('operates entirely from the keyboard', async () => {
-    const user = userEvent.setup();
-    render(<DownloadCaptions {...props} />);
-
-    screen.getByRole('button', { name: 'Download Captions' }).focus();
-    await user.keyboard('{Enter}');
-    const first = await screen.findByRole('menuitem', { name: 'JSON' });
-    expect(first).toHaveFocus();
-    await user.keyboard('{ArrowDown}{Enter}');
-
-    expect(vi.mocked(downloadBlob)).toHaveBeenCalledWith(
-      expect.any(Blob),
-      'videos.txt',
-    );
-  });
-});
@@ -1,13 +1,12 @@
 'use client';

 import { ChevronDown } from 'lucide-react';
-import * as DropdownMenu from '@radix-ui/react-dropdown-menu';

 import { Button } from '@/components/ui/button';
 import { downloadBlob } from '@/lib/utils';

 const MENU_ITEM =
-  'block min-h-11 w-full cursor-pointer select-none px-4 py-2 text-left text-sm font-medium text-foreground outline-none transition-colors data-[highlighted]:bg-muted';
+  'block w-full cursor-pointer px-4 py-2 text-left text-sm font-medium text-foreground transition-colors hover:bg-muted disabled:cursor-not-allowed disabled:opacity-50';

 export default function DownloadCaptions({
  fileNames,
@@ -63,40 +62,52 @@ export default function DownloadCaptions({
    downloadBlob(new Blob([csv], { type: 'text/csv' }), 'captions.csv');
  }

-  // Same click/keyboard-accessible menu idiom as CreateJobButton — hover-only
-  // menus exclude touch and keyboard users.
  return (
-    <DropdownMenu.Root>
-      <DropdownMenu.Trigger asChild>
-        <Button
-          type="button"
-          variant="outline"
-          size="sm"
-          disabled={disabled}
-          className="gap-1.5"
+    <div className="group relative inline-block">
+      <Button
+        type="button"
+        variant="outline"
+        size="sm"
+        disabled={disabled}
+        aria-haspopup="menu"
+        className="gap-1.5"
+      >
+        Download Captions
+        <ChevronDown className="h-3.5 w-3.5 opacity-85" />
+      </Button>
+      {!disabled && (
+        <div
+          role="menu"
+          className="invisible absolute right-0 top-full z-[200] min-w-full -translate-y-1 pt-1 opacity-0 transition-all group-focus-within:visible group-focus-within:translate-y-0 group-focus-within:opacity-100 group-hover:visible group-hover:translate-y-0 group-hover:opacity-100"
        >
-          Download Captions
-          <ChevronDown className="h-3.5 w-3.5 opacity-85" aria-hidden />
-        </Button>
-      </DropdownMenu.Trigger>
-      <DropdownMenu.Portal>
-        <DropdownMenu.Content
-          align="end"
-          sideOffset={4}
-          collisionPadding={8}
-          className="z-[200] min-w-40 overflow-hidden rounded-lg border border-border bg-popover py-1 text-popover-foreground shadow-lg"
-        >
-          <DropdownMenu.Item className={MENU_ITEM} onSelect={handleDownloadJson}>
-            JSON
-          </DropdownMenu.Item>
-          <DropdownMenu.Item className={MENU_ITEM} onSelect={handleDownloadTxt}>
-            TXT
-          </DropdownMenu.Item>
-          <DropdownMenu.Item className={MENU_ITEM} onSelect={handleDownloadCsv}>
-            CSV
-          </DropdownMenu.Item>
-        </DropdownMenu.Content>
-      </DropdownMenu.Portal>
-    </DropdownMenu.Root>
+          <div className="overflow-hidden rounded-lg border border-border bg-popover py-1 shadow-lg">
+            <button
+              type="button"
+              role="menuitem"
+              className={MENU_ITEM}
+              onClick={handleDownloadJson}
+            >
+              JSON
+            </button>
+            <button
+              type="button"
+              role="menuitem"
+              className={MENU_ITEM}
+              onClick={handleDownloadTxt}
+            >
+              TXT
+            </button>
+            <button
+              type="button"
+              role="menuitem"
+              className={MENU_ITEM}
+              onClick={handleDownloadCsv}
+            >
+              CSV
+            </button>
+          </div>
+        </div>
+      )}
+    </div>
  );
 }
@@ -1,53 +0,0 @@
-import { render, screen } from '@testing-library/react';
-import userEvent from '@testing-library/user-event';
-import { describe, expect, it, vi } from 'vitest';
-
-import CreateJobButton from './CreateJobButton';
-
-vi.mock('./CreateJobModal', () => ({
-  default: ({
-    isOpen,
-    workloadType,
-  }: {
-    isOpen: boolean;
-    workloadType: string;
-  }) =>
-    isOpen ? (
-      <div role="dialog" data-workload-type={workloadType}>
-        Create job form
-      </div>
-    ) : null,
-}));
-
-describe('CreateJobButton', () => {
-  it('opens the workload menu on click and selects an item', async () => {
-    const user = userEvent.setup();
-    render(<CreateJobButton jobType="inference" />);
-
-    await user.click(screen.getByRole('button', { name: 'Create Job' }));
-    await user.click(screen.getByRole('menuitem', { name: /I2V/i }));
-
-    expect(screen.getByRole('dialog')).toHaveAttribute(
-      'data-workload-type',
-      'i2v',
-    );
-  });
-
-  it('opens and operates the workload menu from the keyboard', async () => {
-    const user = userEvent.setup();
-    render(<CreateJobButton jobType="inference" />);
-
-    const trigger = screen.getByRole('button', { name: 'Create Job' });
-    trigger.focus();
-    await user.keyboard('{Enter}');
-
-    const firstItem = await screen.findByRole('menuitem', { name: /T2V/i });
-    expect(firstItem).toHaveFocus();
-    await user.keyboard('{Enter}');
-
-    expect(screen.getByRole('dialog')).toHaveAttribute(
-      'data-workload-type',
-      't2v',
-    );
-  });
-});
@@ -2,7 +2,6 @@

 import * as React from 'react';
 import { ChevronDown } from 'lucide-react';
-import * as DropdownMenu from '@radix-ui/react-dropdown-menu';

 import CreateJobModal from '@/components/jobs/CreateJobModal';
 import { Button } from '@/components/ui/button';
@@ -34,35 +33,31 @@ export default function CreateJobButton({ jobType }: CreateJobButtonProps) {

  return (
    <>
-      <DropdownMenu.Root>
-        <DropdownMenu.Trigger asChild>
-          <Button type="button" className="gap-1.5">
-            Create Job
-            <ChevronDown className="size-3.5 opacity-85" aria-hidden />
-          </Button>
-        </DropdownMenu.Trigger>
-        <DropdownMenu.Portal>
-          <DropdownMenu.Content
-            align="end"
-            sideOffset={4}
-            collisionPadding={8}
-            className="z-[200] min-w-48 overflow-hidden rounded-lg border border-border bg-popover py-1 text-popover-foreground shadow-lg"
-          >
-            {options.map((opt) => (
-              <DropdownMenu.Item
-                key={opt.type}
-                onSelect={() => openModal(opt.type)}
-                className="flex min-h-11 cursor-pointer select-none flex-col justify-center px-4 py-2 text-left text-sm font-medium outline-none data-[highlighted]:bg-secondary"
-              >
-                {opt.label}
-                <span className="mt-0.5 block text-xs font-normal text-muted-foreground">
-                  {opt.desc}
-                </span>
-              </DropdownMenu.Item>
-            ))}
-          </DropdownMenu.Content>
-        </DropdownMenu.Portal>
-      </DropdownMenu.Root>
+      <div className="group relative inline-block">
+        <Button type="button" className="gap-1.5">
+          Create Job
+          <ChevronDown className="size-3.5 opacity-85" aria-hidden />
+        </Button>
+        <div
+          role="menu"
+          className="invisible absolute right-0 top-full z-[200] mt-1 min-w-full -translate-y-1 rounded-lg border border-border bg-popover py-1 opacity-0 shadow-lg transition-all duration-150 group-hover:visible group-hover:translate-y-0 group-hover:opacity-100"
+        >
+          {options.map((opt) => (
+            <button
+              key={opt.type}
+              type="button"
+              role="menuitem"
+              onClick={() => openModal(opt.type)}
+              className="block w-full whitespace-nowrap px-4 py-2 text-left text-sm font-medium text-popover-foreground transition-colors hover:bg-secondary"
+            >
+              {opt.label}
+              <span className="mt-0.5 block text-xs font-normal text-muted-foreground">
+                {opt.desc}
+              </span>
+            </button>
+          ))}
+        </div>
+      </div>
      <CreateJobModal
        isOpen={modalOpen}
        onClose={() => setModalOpen(false)}
@@ -50,57 +50,6 @@ function renderModal(
 }

 describe('CreateJobModal', () => {
-  it('shows a model loading error instead of an empty model list', async () => {
-    vi.spyOn(console, 'error').mockImplementation(() => {});
-    vi.mocked(getModels).mockRejectedValueOnce(new Error('network down'));
-
-    renderModal();
-
-    expect(
-      await screen.findByText(/Models could not be loaded/),
-    ).toBeInTheDocument();
-    expect(screen.getByLabelText('Model')).toHaveAttribute(
-      'aria-invalid',
-      'true',
-    );
-  });
-
-  it('keeps the form open and reports job creation failures', async () => {
-    vi.spyOn(console, 'error').mockImplementation(() => {});
-    vi.mocked(createJob).mockRejectedValueOnce(new Error('API rejected job'));
-    const user = userEvent.setup();
-    const { onClose, onSuccess } = renderModal();
-
-    await screen.findByRole('option', { name: 'Wan T2V (wan/t2v-1.3b)' });
-    await user.type(screen.getByLabelText('Prompt'), 'a careful test prompt');
-    await user.click(screen.getByRole('button', { name: 'Create Job' }));
-
-    expect(
-      await screen.findByText(/API rejected job.*then try again/),
-    ).toBeInTheDocument();
-    expect(onSuccess).not.toHaveBeenCalled();
-    expect(onClose).not.toHaveBeenCalled();
-  });
-
-  it('reports image upload failures next to the file input', async () => {
-    vi.spyOn(console, 'error').mockImplementation(() => {});
-    vi.mocked(uploadImage).mockRejectedValueOnce(new Error('Upload failed'));
-    const user = userEvent.setup();
-    renderModal({ workloadType: 'i2v' });
-
-    await screen.findByRole('option', { name: 'Wan T2V (wan/t2v-1.3b)' });
-    const input = screen.getByLabelText('Image');
-    await user.upload(
-      input,
-      new File(['image'], 'input.png', { type: 'image/png' }),
-    );
-
-    expect(
-      await screen.findByText(/Upload failed.*Choose the image again/),
-    ).toBeInTheDocument();
-    expect(input).toHaveAttribute('aria-invalid', 'true');
-  });
-
  it('renders the form fields for an inference job', async () => {
    renderModal();

@@ -103,17 +103,6 @@ export default function CreateJobModal({
  const [fakeScoreModelPath, setFakeScoreModelPath] = React.useState('');
  const [isSubmitting, setIsSubmitting] = React.useState(false);
  const [isLoadingModels, setIsLoadingModels] = React.useState(false);
-  const [isLoadingDatasets, setIsLoadingDatasets] = React.useState(false);
-  const [modelLoadError, setModelLoadError] = React.useState<string | null>(
-    null,
-  );
-  const [datasetLoadError, setDatasetLoadError] = React.useState<string | null>(
-    null,
-  );
-  const [imageUploadError, setImageUploadError] = React.useState<string | null>(
-    null,
-  );
-  const [submitError, setSubmitError] = React.useState<string | null>(null);
  const imageInputRef = React.useRef<HTMLInputElement>(null);

  // Seed field values from the persisted default options each time the modal
@@ -155,10 +144,6 @@ export default function CreateJobModal({
    setImageFileName('');
    setSelectedDatasetId('');
    setSelectedValidationDatasetId('');
-    setModelLoadError(null);
-    setDatasetLoadError(null);
-    setImageUploadError(null);
-    setSubmitError(null);
    if (workloadType === 'dmd_t2v') {
      setDmdUseVsa(false);
      setDmdVsaSparsity(0.8);
@@ -177,7 +162,6 @@ export default function CreateJobModal({
    // can't overwrite the current workload's model list/selection.
    let stale = false;
    setIsLoadingModels(true);
-    setModelLoadError(null);
    getModels(inferenceWorkload)
      .then((list) => {
        if (stale) return;
@@ -196,13 +180,7 @@ export default function CreateJobModal({
        }
      })
      .catch((e) => {
-        if (stale) return;
-        console.error('Failed to load models:', e);
-        setModels([]);
-        setModelId('');
-        setModelLoadError(
-          'Models could not be loaded. Check the Studio API and reopen this form to try again.',
-        );
+        if (!stale) console.error('Failed to load models:', e);
      })
      .finally(() => {
        if (!stale) setIsLoadingModels(false);
@@ -215,22 +193,11 @@ export default function CreateJobModal({
  // Training jobs need a dataset; load the ready datasets when relevant.
  React.useEffect(() => {
    if (isOpen && !isInference) {
-      setIsLoadingDatasets(true);
-      setDatasetLoadError(null);
      getDatasets()
        .then(setReadyDatasets)
-        .catch((error) => {
-          console.error('Failed to load datasets:', error);
-          setReadyDatasets([]);
-          setDatasetLoadError(
-            'Datasets could not be loaded. Check the Studio API and reopen this form to try again.',
-          );
-        })
-        .finally(() => setIsLoadingDatasets(false));
+        .catch(() => setReadyDatasets([]));
    } else {
      setReadyDatasets([]);
-      setIsLoadingDatasets(false);
-      setDatasetLoadError(null);
    }
  }, [isOpen, isInference]);

@@ -239,24 +206,16 @@ export default function CreateJobModal({
    if (!file) {
      setImagePath('');
      setImageFileName('');
-      setImageUploadError(null);
      return;
    }
    setIsUploadingImage(true);
    setImageFileName(file.name);
-    setImageUploadError(null);
    try {
      const { path } = await uploadImage(file);
      setImagePath(path);
-    } catch (error) {
-      console.error('Failed to upload image:', error);
+    } catch {
      setImagePath('');
      setImageFileName('');
-      setImageUploadError(
-        error instanceof Error
-          ? `${error.message}. Choose the image again to retry.`
-          : 'The image could not be uploaded. Choose it again to retry.',
-      );
    } finally {
      setIsUploadingImage(false);
    }
@@ -265,7 +224,6 @@ export default function CreateJobModal({
  function clearImage() {
    setImagePath('');
    setImageFileName('');
-    setImageUploadError(null);
    if (imageInputRef.current) imageInputRef.current.value = '';
  }

@@ -281,7 +239,6 @@ export default function CreateJobModal({
      workloadType === 'lora_t2v' ? 'lora' : jobType
    ) as JobType;
    setIsSubmitting(true);
-    setSubmitError(null);
    try {
      const payload: CreateJobRequest = {
        model_id: modelId,
@@ -339,11 +296,6 @@ export default function CreateJobModal({
      onClose();
    } catch (err) {
      console.error('Failed to create job:', err);
-      setSubmitError(
-        err instanceof Error
-          ? `${err.message}. Check the form and Studio API, then try again.`
-          : 'The job could not be created. Check the form and Studio API, then try again.',
-      );
    } finally {
      setIsSubmitting(false);
    }
@@ -391,11 +343,7 @@ export default function CreateJobModal({
              value={modelId}
              onChange={(e) => setModelId(e.target.value)}
              required
-              aria-describedby={
-                modelLoadError ? 'modal-model-error' : undefined
-              }
-              aria-invalid={modelLoadError ? true : undefined}
-              disabled={isSubmitting || isLoadingModels || !!modelLoadError}
+              disabled={isSubmitting || isLoadingModels}
            >
              <option value="" disabled>
                {isLoadingModels
@@ -410,15 +358,6 @@ export default function CreateJobModal({
                </option>
              ))}
            </NativeSelect>
-            {modelLoadError && (
-              <p
-                id="modal-model-error"
-                role="alert"
-                className="text-sm text-destructive"
-              >
-                {modelLoadError}
-              </p>
-            )}
          </FieldRow>

          {isInference && workloadType === 'i2v' && (
@@ -430,10 +369,6 @@ export default function CreateJobModal({
                accept=".png,.jpg,.jpeg,.webp,.bmp"
                onChange={handleImageChange}
                disabled={isSubmitting || isUploadingImage}
-                aria-describedby={
-                  imageUploadError ? 'modal-image-error' : undefined
-                }
-                aria-invalid={imageUploadError ? true : undefined}
                required
                className="h-auto py-2 file:mr-3 file:cursor-pointer file:rounded-md file:border-0 file:bg-secondary file:px-2 file:py-1 file:text-sm file:text-secondary-foreground"
              />
@@ -450,15 +385,6 @@ export default function CreateJobModal({
                  </button>
                </span>
              )}
-              {imageUploadError && (
-                <p
-                  id="modal-image-error"
-                  role="alert"
-                  className="text-sm text-destructive"
-                >
-                  {imageUploadError}
-                </p>
-              )}
            </FieldRow>
          )}

@@ -506,22 +432,12 @@ export default function CreateJobModal({
                    id="modal-dataset"
                    value={selectedDatasetId}
                    onChange={(e) => setSelectedDatasetId(e.target.value)}
-                    aria-describedby={
-                      datasetLoadError ? 'modal-dataset-error' : undefined
-                    }
-                    aria-invalid={datasetLoadError ? true : undefined}
-                    disabled={
-                      isSubmitting || isLoadingDatasets || !!datasetLoadError
-                    }
+                    disabled={isSubmitting}
                  >
                    <option value="" disabled>
-                      {isLoadingDatasets
-                        ? 'Loading datasets…'
-                        : datasetLoadError
-                          ? 'Datasets unavailable'
-                          : readyDatasets.length === 0
-                            ? 'No datasets (add in Datasets tab)'
-                            : 'Select a dataset…'}
+                      {readyDatasets.length === 0
+                        ? 'No datasets (add in Datasets tab)'
+                        : 'Select a dataset…'}
                    </option>
                    {readyDatasets.map((d) => (
                      <option key={d.id} value={d.id}>
@@ -529,15 +445,6 @@ export default function CreateJobModal({
                      </option>
                    ))}
                  </NativeSelect>
-                  {datasetLoadError && (
-                    <p
-                      id="modal-dataset-error"
-                      role="alert"
-                      className="text-sm text-destructive"
-                    >
-                      {datasetLoadError}
-                    </p>
-                  )}
                </FieldRow>
                <FieldRow
                  htmlFor="modal-validation-dataset"
@@ -550,9 +457,7 @@ export default function CreateJobModal({
                    onChange={(e) =>
                      setSelectedValidationDatasetId(e.target.value)
                    }
-                    disabled={
-                      isSubmitting || isLoadingDatasets || !!datasetLoadError
-                    }
+                    disabled={isSubmitting}
                  >
                    <option value="">None</option>
                    {readyDatasets.map((d) => (
@@ -906,22 +811,9 @@ export default function CreateJobModal({
            </details>
          )}

-          <div className="flex flex-col items-start gap-2">
-            {submitError && (
-              <p role="alert" className="text-sm text-destructive">
-                {submitError}
-              </p>
-            )}
-            <Button
-              type="submit"
-              disabled={
-                isSubmitting ||
-                isUploadingImage ||
-                !!modelLoadError ||
-                !!datasetLoadError
-              }
-            >
-              {isSubmitting ? 'Creating…' : 'Create Job'}
+          <div>
+            <Button type="submit" disabled={isSubmitting}>
+              {isSubmitting ? 'Creating...' : 'Create Job'}
            </Button>
          </div>
        </form>
@@ -20,14 +20,6 @@ vi.mock('@/lib/api', () => ({
  downloadJobVideo: vi.fn(),
 }));

-vi.mock('@/lib/utils', async (importOriginal) => {
-  const actual = await importOriginal<typeof import('@/lib/utils')>();
-  return {
-    ...actual,
-    downloadBlob: vi.fn(),
-  };
-});
-
 const makeJob = (overrides: Partial<Job> = {}): Job =>
  makeBaseJob({
    model_id: 'Wan2.1-T2V',
@@ -54,16 +46,6 @@ beforeEach(() => {
 });

 describe('JobCard', () => {
-  it('keeps selection and job action buttons as semantic siblings', () => {
-    render(<JobCard job={makeJob()} />);
-
-    const selectButton = screen.getByRole('button', { pressed: false });
-    const deleteButton = screen.getByRole('button', { name: 'Delete' });
-
-    expect(selectButton).toHaveTextContent('Wan2.1-T2V');
-    expect(selectButton).not.toContainElement(deleteButton);
-  });
-
  it('renders the model, prompt, status and inference meta', () => {
    render(<JobCard job={makeJob()} />);
    expect(screen.getByText('Wan2.1-T2V')).toBeInTheDocument();
@@ -119,10 +119,18 @@ export default function JobCard({ job, onJobUpdated }: JobCardProps) {
    }
  }

-  function handleSelectJob() {
+  function handleSelectJob(e: React.MouseEvent | React.KeyboardEvent) {
+    if ((e.target as HTMLElement).closest('button')) return;
    setActiveJobId(isSelected ? null : job.id);
  }

+  function handleKeyDown(e: React.KeyboardEvent) {
+    if (e.key === 'Enter' || e.key === ' ') {
+      e.preventDefault();
+      handleSelectJob(e);
+    }
+  }
+
  async function handleDownloadVideo(e: React.MouseEvent) {
    e.preventDefault();
    e.stopPropagation();
@@ -140,7 +148,11 @@ export default function JobCard({ job, onJobUpdated }: JobCardProps) {
  }

  return (
-    <article
+    <div
+      role="button"
+      tabIndex={0}
+      onClick={handleSelectJob}
+      onKeyDown={handleKeyDown}
      className={cn(
        'mb-3 flex cursor-pointer flex-col gap-2.5 rounded-lg border bg-background p-4 transition-colors last:mb-0',
        isSelected
@@ -148,42 +160,35 @@ export default function JobCard({ job, onJobUpdated }: JobCardProps) {
          : 'border-border hover:border-muted-foreground/40',
      )}
    >
-      <button
-        type="button"
-        aria-pressed={isSelected}
-        onClick={handleSelectJob}
-        className="flex w-full flex-col gap-2.5 rounded-md text-left"
-      >
-        <span className="flex flex-wrap items-center justify-between gap-2">
-          <span className="text-[0.95rem] font-semibold text-foreground">
-            {job.model_id}
-          </span>
-          <Badge variant={BADGE_VARIANTS[job.status] ?? 'secondary'}>
-            {job.status}
-          </Badge>
+      <div className="flex flex-wrap items-center justify-between gap-2">
+        <span className="text-[0.95rem] font-semibold text-foreground">
+          {job.model_id}
        </span>
-        <span className="max-w-full overflow-hidden text-ellipsis whitespace-nowrap text-sm text-muted-foreground">
-          {job.prompt}
-        </span>
-        <span className="flex flex-wrap items-center gap-4 text-xs text-muted-foreground">
-          {job.job_type === 'inference' ? (
-            <>
-              <span>{job.num_frames} frames</span>
-              <span>
-                {job.height}×{job.width}
-              </span>
-            </>
-          ) : (
-            <span>{job.workload_type?.replace(/_/g, ' ') ?? job.job_type}</span>
-          )}
-          {elapsedTime && (
-            <span className="inline-flex items-center gap-1">
-              <Timer className="size-3.5" aria-hidden />
-              {elapsedTime}
+        <Badge variant={BADGE_VARIANTS[job.status] ?? 'secondary'}>
+          {job.status}
+        </Badge>
+      </div>
+      <p className="max-w-full overflow-hidden text-ellipsis whitespace-nowrap text-sm text-muted-foreground">
+        {job.prompt}
+      </p>
+      <div className="flex flex-wrap items-center gap-4 text-xs text-muted-foreground">
+        {job.job_type === 'inference' ? (
+          <>
+            <span>{job.num_frames} frames</span>
+            <span>
+              {job.height}×{job.width}
            </span>
-          )}
-        </span>
-      </button>
+          </>
+        ) : (
+          <span>{job.workload_type?.replace(/_/g, ' ') ?? job.job_type}</span>
+        )}
+        {elapsedTime && (
+          <span className="inline-flex items-center gap-1">
+            <Timer className="size-3.5" aria-hidden />
+            {elapsedTime}
+          </span>
+        )}
+      </div>
      <div className="flex flex-wrap items-center gap-1.5">
        {job.status === 'running' ? (
          <Button
@@ -235,6 +240,6 @@ export default function JobCard({ job, onJobUpdated }: JobCardProps) {
          Delete
        </Button>
      </div>
-    </article>
+    </div>
  );
 }
@@ -19,32 +19,6 @@ const makeJob = (overrides: Partial<Job> = {}): Job =>
  });

 describe('JobDetailsSidebar', () => {
-  it('fills the mobile viewport without reserving main-content width', async () => {
-    vi.mocked(getJobLogs).mockResolvedValue({
-      lines: [],
-      total: 0,
-      progress: 0,
-      progress_msg: '',
-      phase: '',
-    });
-    const onWidthChange = vi.fn();
-
-    render(
-      <JobDetailsSidebar
-        job={makeJob({ status: 'completed' })}
-        isMobile
-        onClose={vi.fn()}
-        onWidthChange={onWidthChange}
-      />,
-    );
-
-    const drawer = screen.getByRole('dialog', { name: 'Job details' });
-    expect(drawer).toHaveStyle({ width: '100%', maxWidth: 'none' });
-    expect(drawer).toHaveAttribute('aria-modal', 'true');
-    expect(drawer).toHaveFocus();
-    expect(onWidthChange).toHaveBeenCalledWith(0);
-  });
-
  it('renders log lines streamed from the job log poll', async () => {
    vi.mocked(getJobLogs).mockResolvedValue({
      lines: ['boot sequence started', 'loading model weights'],
@@ -4,7 +4,6 @@ import * as React from 'react';
 import { X } from 'lucide-react';

 import { Button } from '@/components/ui/button';
-import { useDrawerFocus } from '@/hooks/useDrawerFocus';
 import { useResizable } from '@/hooks/useResizable';
 import { downloadJobLog, getJobLogs } from '@/lib/api';
 import type { Job } from '@/lib/types';
@@ -16,16 +15,13 @@ const POLL_INTERVAL_MS = 2000;

 export default function JobDetailsSidebar({
  job,
-  isMobile = false,
  onClose,
  onWidthChange,
 }: {
  job: Job;
-  isMobile?: boolean;
  onClose: () => void;
  onWidthChange?: (w: number) => void;
 }) {
-  const drawerRef = useDrawerFocus<HTMLElement>(isMobile);
  const [width, setWidth] = React.useState(360);
  const [isDragging, setIsDragging] = React.useState(false);
  const [isLoading, setIsLoading] = React.useState(false);
@@ -57,8 +53,8 @@ export default function JobDetailsSidebar({
  });

  React.useEffect(() => {
-    onWidthChange?.(isMobile ? 0 : width);
-  }, [isMobile, width, onWidthChange]);
+    onWidthChange?.(width);
+  }, [width, onWidthChange]);

  // Auto-scroll the console to the bottom whenever new lines land. Runs after
  // commit so scrollHeight reflects the freshly-rendered output.
@@ -141,16 +137,8 @@ export default function JobDetailsSidebar({

  return (
    <aside
-      ref={drawerRef}
-      tabIndex={-1}
-      role="dialog"
-      aria-label="Job details"
-      aria-modal={isMobile || undefined}
-      className="fixed bottom-0 right-0 top-[var(--header-height)] z-50 flex max-h-[calc(100dvh-var(--header-height))] min-w-0 shrink-0 flex-col border-l border-border bg-card md:min-w-[280px]"
-      style={{
-        width: isMobile ? '100%' : width,
-        maxWidth: isMobile ? 'none' : SIDEBAR_MAX_WIDTH,
-      }}
+      className="fixed bottom-0 right-0 top-[var(--header-height)] z-50 flex max-h-[calc(100vh-var(--header-height))] min-w-[280px] shrink-0 flex-col border-l border-border bg-card"
+      style={{ width, maxWidth: SIDEBAR_MAX_WIDTH }}
    >
      <div className="flex items-center justify-between border-b border-border px-5 py-4">
        <h2 className="m-0 text-base font-semibold text-foreground">
@@ -207,14 +195,14 @@ export default function JobDetailsSidebar({
        </pre>
      </div>

-      {!isMobile && <div
+      <div
        role="presentation"
        onMouseDown={onMouseDown}
        className={cn(
          'absolute bottom-0 left-0 top-0 z-[1] w-1.5 cursor-col-resize hover:bg-accent-blue/25',
          isDragging && 'bg-accent-blue/25',
        )}
-      />}
+      />
    </aside>
  );
 }
@@ -1,4 +1,4 @@
-import { act, fireEvent, render, screen, waitFor } from '@testing-library/react';
+import { act, render, screen, waitFor } from '@testing-library/react';
 import { beforeEach, describe, expect, it, vi } from 'vitest';

 import JobQueue from '@/components/jobs/JobQueue';
@@ -37,46 +37,6 @@ beforeEach(() => {
 });

 describe('JobQueue', () => {
-  it('shows a loading placeholder before the initial request settles', async () => {
-    let resolveJobs: (jobs: Job[]) => void = () => {};
-    vi.mocked(getJobsList).mockReturnValue(
-      new Promise<Job[]>((resolve) => {
-        resolveJobs = resolve;
-      }),
-    );
-
-    render(<JobQueue jobType="inference" />);
-
-    expect(screen.getByLabelText('Loading jobs')).toBeInTheDocument();
-    expect(
-      screen.queryByText('No inference jobs yet. Create one above.'),
-    ).not.toBeInTheDocument();
-
-    act(() => resolveJobs([]));
-    expect(
-      await screen.findByText('No inference jobs yet. Create one above.'),
-    ).toBeInTheDocument();
-  });
-
-  it('shows request failures separately from an empty queue and retries', async () => {
-    vi.spyOn(console, 'error').mockImplementation(() => {});
-    vi.mocked(getJobsList).mockRejectedValueOnce(new Error('network down'));
-    render(<JobQueue jobType="inference" />);
-
-    expect(
-      await screen.findByText(/Could not load jobs from the Studio API/),
-    ).toBeInTheDocument();
-    expect(
-      screen.queryByText('No inference jobs yet. Create one above.'),
-    ).not.toBeInTheDocument();
-
-    vi.mocked(getJobsList).mockResolvedValueOnce([]);
-    fireEvent.click(screen.getByRole('button', { name: 'Try Again' }));
-    expect(
-      await screen.findByText('No inference jobs yet. Create one above.'),
-    ).toBeInTheDocument();
-  });
-
  it('shows an empty placeholder and fetches for the single job type', async () => {
    render(<JobQueue jobType="inference" />);
    expect(
@@ -1,10 +1,8 @@
 'use client';

 import * as React from 'react';
-import { AlertTriangle } from 'lucide-react';

 import JobCard from '@/components/jobs/JobCard';
-import { Button } from '@/components/ui/button';
 import { useStore } from '@/hooks/useStore';
 import { getJobsList } from '@/lib/api';
 import type { Job, JobType } from '@/lib/types';
@@ -34,8 +32,6 @@ function jobsShallowEqual(a: Job | null, b: Job | null): boolean {

 export default function JobQueue({ jobType, jobTypesForList }: JobQueueProps) {
  const [jobs, setJobs] = React.useState<Job[]>([]);
-  const [isInitialLoading, setIsInitialLoading] = React.useState(true);
-  const [error, setError] = React.useState<string | null>(null);
  const { nonce } = useStore(jobsRefreshStore);
  const { activeJobId } = useStore(activeJobStore);

@@ -75,22 +71,11 @@ export default function JobQueue({ jobType, jobTypesForList }: JobQueueProps) {
              new Date(a.created_at ?? 0).getTime(),
          );
      }
-      if (seq === fetchSeq.current) {
-        setJobs(next);
-        setError(null);
-      }
+      if (seq === fetchSeq.current) setJobs(next);
    } catch (e) {
      console.error('Failed to fetch jobs:', e);
-      if (seq === fetchSeq.current) {
-        setError(
-          'Could not load jobs from the Studio API. Check the server and try again.',
-        );
-      }
    } finally {
-      if (seq === fetchSeq.current) {
-        inFlight.current = false;
-        setIsInitialLoading(false);
-      }
+      if (seq === fetchSeq.current) inFlight.current = false;
    }
  }, [typesKey]);

@@ -137,59 +122,20 @@ export default function JobQueue({ jobType, jobTypesForList }: JobQueueProps) {
  const multiType = typesToFetch.length > 1;

  return (
-    <div className="mx-auto flex w-full max-w-[850px] flex-col gap-6 px-4 pb-12">
+    <main className="mx-auto flex w-full max-w-[850px] flex-col gap-6 px-4 pb-12">
      <section className="p-6">
-        <div aria-busy={isInitialLoading}>
-          {isInitialLoading ? (
-            <div
-              aria-label="Loading jobs"
-              className="flex flex-col gap-3 py-2"
-            >
-              {[0, 1, 2].map((item) => (
-                <div
-                  key={item}
-                  className="h-32 animate-pulse rounded-lg border border-border bg-muted/50"
-                />
-              ))}
-            </div>
-          ) : error && jobs.length === 0 ? (
-            <div
-              role="alert"
-              className="flex flex-col items-center gap-3 py-8 text-center"
-            >
-              <AlertTriangle
-                className="size-6 text-destructive"
-                aria-hidden
-              />
-              <p className="max-w-md text-sm text-muted-foreground">{error}</p>
-              <Button type="button" variant="outline" onClick={fetchJobs}>
-                Try Again
-              </Button>
-            </div>
-          ) : (
-            <>
-              {error && (
-                <p
-                  role="status"
-                  className="mb-3 rounded-lg border border-amber-500/50 bg-amber-500/10 px-3 py-2 text-sm text-foreground"
-                >
-                  Job updates are temporarily unavailable. Showing the most
-                  recent results.
-                </p>
-              )}
-              {jobs.length === 0 ? (
+        <div>
+          {jobs.length === 0 ? (
            <p className="py-8 text-center text-muted-foreground">
              No {multiType ? 'jobs' : `${jobType} jobs`} yet. Create one above.
            </p>
-              ) : (
-                jobs.map((job) => (
-                  <JobCard key={job.id} job={job} onJobUpdated={fetchJobs} />
-                ))
-              )}
-            </>
+          ) : (
+            jobs.map((job) => (
+              <JobCard key={job.id} job={job} onJobUpdated={fetchJobs} />
+            ))
          )}
        </div>
      </section>
-    </div>
+    </main>
  );
 }
@@ -9,7 +9,6 @@ import { HeaderActionsProvider } from '@/components/shell/HeaderActionsContext';
 import PrimarySidebar from '@/components/shell/PrimarySidebar';
 import JobDetailsSidebar from '@/components/jobs/JobDetailsSidebar';
 import { Toaster } from '@/components/ui/sonner';
-import { useMediaQuery } from '@/hooks/useMediaQuery';
 import { useStore } from '@/hooks/useStore';
 import {
  activeDatasetStore,
@@ -24,80 +23,46 @@ export function AppShell({ children }: { children: React.ReactNode }) {
  const pathname = usePathname();
  const { activeJob } = useStore(activeJobStore);
  const { activeDataset } = useStore(activeDatasetStore);
-  const isMobile = useMediaQuery('(max-width: 767px)');

  const [primaryWidth, setPrimaryWidth] = React.useState(220);
  const [secondaryWidth, setSecondaryWidth] = React.useState(0);
-  const [primaryOpen, setPrimaryOpen] = React.useState(false);

  const jobSidebarOpen = JOB_ROUTES.includes(pathname) && activeJob != null;
  const datasetSidebarOpen =
    pathname === '/datasets' && activeDataset != null;
  const secondaryOpen = jobSidebarOpen || datasetSidebarOpen;
-  // Mobile detail drawers claim aria-modal, so everything behind them must
-  // actually be inert — the platform enforces what the ARIA claims.
-  const drawerModal = isMobile && secondaryOpen;

  React.useEffect(() => {
    initDefaultOptions();
  }, []);

-  React.useEffect(() => {
-    setPrimaryOpen(false);
-  }, [pathname]);
-
  React.useEffect(() => {
    function handleKeyDown(e: KeyboardEvent) {
-      if (e.key !== 'Escape' || document.querySelector('[data-modal]')) return;
-      if (primaryOpen) {
-        setPrimaryOpen(false);
-        return;
+      if (e.key === 'Escape' && !document.querySelector('[data-modal]')) {
+        if (activeJobStore.get().activeJob) setActiveJobId(null);
+        if (activeDatasetStore.get().activeDataset) setActiveDatasetId(null);
      }
-      if (activeJobStore.get().activeJob) setActiveJobId(null);
-      if (activeDatasetStore.get().activeDataset) setActiveDatasetId(null);
    }
    document.addEventListener('keydown', handleKeyDown);
    return () => document.removeEventListener('keydown', handleKeyDown);
-  }, [primaryOpen]);
+  }, []);

  return (
    <HeaderActionsProvider>
-      <div
-        style={{ display: 'contents' }}
-        inert={drawerModal ? true : undefined}
-      >
-        <Header
-          navigationOpen={primaryOpen}
-          onNavigationToggle={() => setPrimaryOpen((open) => !open)}
-        />
-      </div>
+      <Header />
      <div
        className="flex overflow-hidden"
        style={{
          marginTop: 'var(--header-height)',
-          height: 'calc(100dvh - var(--header-height))',
+          height: 'calc(100vh - var(--header-height))',
        }}
      >
-        <PrimarySidebar
-          isMobile={isMobile}
-          mobileOpen={primaryOpen}
-          onMobileClose={() => setPrimaryOpen(false)}
-          onWidthChange={setPrimaryWidth}
-        />
-        {primaryOpen && (
-          <button
-            type="button"
-            aria-label="Close navigation"
-            onClick={() => setPrimaryOpen(false)}
-            className="fixed inset-x-0 bottom-0 top-[var(--header-height)] z-40 bg-black/55 md:hidden"
-          />
-        )}
+        <PrimarySidebar onWidthChange={setPrimaryWidth} />
        <main
          className="flex min-w-0 flex-1 flex-col overflow-auto"
-          inert={drawerModal ? true : undefined}
          style={{
-            marginLeft: isMobile ? 0 : primaryWidth,
-            marginRight: isMobile || !secondaryOpen ? 0 : secondaryWidth,
+            marginLeft: primaryWidth,
+            marginRight: secondaryOpen ? secondaryWidth : 0,
          }}
        >
          {children}
@@ -105,7 +70,6 @@ export function AppShell({ children }: { children: React.ReactNode }) {
        {jobSidebarOpen && activeJob && (
          <JobDetailsSidebar
            job={activeJob}
-            isMobile={isMobile}
            onClose={() => setActiveJobId(null)}
            onWidthChange={setSecondaryWidth}
          />
@@ -113,7 +77,6 @@ export function AppShell({ children }: { children: React.ReactNode }) {
        {datasetSidebarOpen && activeDataset && (
          <DatasetSidebar
            dataset={activeDataset}
-            isMobile={isMobile}
            onClose={() => setActiveDatasetId(null)}
            onWidthChange={setSecondaryWidth}
          />
@@ -1,10 +1,8 @@
 'use client';

-import { Menu, X } from 'lucide-react';
 import { usePathname } from 'next/navigation';

 import { useHeaderActions } from '@/components/shell/HeaderActionsContext';
-import { Button } from '@/components/ui/button';
 import { ThemeToggle } from '@/components/ui/theme-toggle';

 const TAB_TITLES: Record<string, string> = {
@@ -17,47 +15,25 @@ const TAB_TITLES: Record<string, string> = {
  '/settings': 'Settings',
 };

-export default function Header({
-  navigationOpen,
-  onNavigationToggle,
-}: {
-  navigationOpen: boolean;
-  onNavigationToggle: () => void;
-}) {
+export default function Header() {
  const pathname = usePathname();
  const { actions } = useHeaderActions();
  const title = TAB_TITLES[pathname] ?? 'FastVideo';

  return (
-    <header className="fixed inset-x-0 top-0 z-[100] flex h-[var(--header-height)] items-center gap-2 border-b border-border bg-background/80 px-2 backdrop-blur sm:px-4 md:gap-6 md:px-6">
-      <Button
-        type="button"
-        variant="outline"
-        size="icon"
-        aria-label={navigationOpen ? 'Close navigation' : 'Open navigation'}
-        aria-controls="primary-navigation"
-        aria-expanded={navigationOpen}
-        onClick={onNavigationToggle}
-        className="shrink-0 md:hidden"
-      >
-        {navigationOpen ? (
-          <X className="size-5" aria-hidden />
-        ) : (
-          <Menu className="size-5" aria-hidden />
-        )}
-      </Button>
+    <header className="fixed inset-x-0 top-0 z-[100] flex h-[var(--header-height)] items-center gap-6 border-b border-border bg-background/80 px-6 backdrop-blur">
      {/* eslint-disable-next-line @next/next/no-img-element */}
      <img
        src="/logo.svg"
        alt="FastVideo Logo"
        width={100}
        height={42}
-        className="hidden h-[42px] w-[78px] shrink-0 object-contain min-[361px]:block md:w-[100px]"
+        className="block h-[42px] w-[100px]"
      />
-      <h1 className="sr-only m-0 flex-1 text-xl font-semibold tracking-tight md:not-sr-only">
+      <h1 className="m-0 flex-1 text-xl font-semibold tracking-tight">
        {title}
      </h1>
-      <div className="ml-auto flex min-w-0 items-center gap-2 md:gap-3">
+      <div className="flex items-center gap-3">
        {actions}
        <ThemeToggle />
      </div>
@@ -1,7 +1,6 @@
 'use client';

 import * as React from 'react';
-import { X } from 'lucide-react';
 import Link from 'next/link';
 import { usePathname } from 'next/navigation';

@@ -20,18 +19,12 @@ const JOB_ROUTES = [
 ] as const;

 const TAB_BASE =
-  'block min-h-11 px-5 py-[0.65rem] text-left text-sm text-muted-foreground transition-colors hover:bg-accent/60 hover:text-foreground';
+  'block px-5 py-[0.65rem] text-left text-sm text-muted-foreground transition-colors hover:bg-accent/60 hover:text-foreground';
 const TAB_ACTIVE = 'bg-accent-blue/10 font-medium text-accent-blue';

 export default function PrimarySidebar({
-  isMobile,
-  mobileOpen,
-  onMobileClose,
  onWidthChange,
 }: {
-  isMobile: boolean;
-  mobileOpen: boolean;
-  onMobileClose: () => void;
  onWidthChange?: (w: number) => void;
 }) {
  const pathname = usePathname();
@@ -45,8 +38,8 @@ export default function PrimarySidebar({
  const isJobsActive = JOB_ROUTES.some((r) => pathname === r.href);

  React.useEffect(() => {
-    onWidthChange?.(isMobile ? 0 : layoutWidth);
-  }, [isMobile, layoutWidth, onWidthChange]);
+    onWidthChange?.(layoutWidth);
+  }, [layoutWidth, onWidthChange]);

  React.useEffect(() => {
    if (JOB_ROUTES.some((r) => pathname === r.href)) {
@@ -65,37 +58,11 @@ export default function PrimarySidebar({

  return (
    <aside
-      id="primary-navigation"
-      aria-hidden={isMobile && !mobileOpen}
-      inert={isMobile && !mobileOpen ? true : undefined}
-      className={cn(
-        'fixed bottom-0 left-0 top-[var(--header-height)] z-50 flex max-h-[calc(100dvh-var(--header-height))] shrink-0 flex-col border-r border-border bg-card transition-transform duration-200 md:translate-x-0',
-        mobileOpen ? 'translate-x-0' : '-translate-x-full',
-      )}
-      style={{
-        width: isMobile
-          ? 'min(18rem, calc(100vw - 3rem))'
-          : effectiveWidth,
-      }}
+      className="fixed bottom-0 left-0 top-[var(--header-height)] z-50 flex max-h-[calc(100vh-var(--header-height))] shrink-0 flex-col border-r border-border bg-card"
+      style={{ width: effectiveWidth }}
    >
-      {isMobile && (
-        <div className="flex h-14 items-center justify-between border-b border-border px-4">
-          <span className="text-sm font-semibold">Navigation</span>
-          <button
-            type="button"
-            onClick={onMobileClose}
-            aria-label="Close navigation"
-            className="flex size-11 items-center justify-center rounded-lg text-muted-foreground hover:bg-accent hover:text-foreground"
-          >
-            <X className="size-5" aria-hidden />
-          </button>
-        </div>
-      )}
      {!isCollapsed && (
-        <nav
-          aria-label="Primary navigation"
-          className="flex flex-col overflow-y-auto py-2"
-        >
+        <nav className="flex flex-col py-2">
          <div className="flex flex-col">
            <button
              type="button"
@@ -128,8 +95,6 @@ export default function PrimarySidebar({
                  <Link
                    key={route.href}
                    href={route.href}
-                    aria-current={pathname === route.href ? 'page' : undefined}
-                    onClick={onMobileClose}
                    className={cn(
                      TAB_BASE,
                      'px-4 py-2 text-[0.85rem]',
@@ -144,32 +109,24 @@ export default function PrimarySidebar({
          </div>
          <Link
            href="/datasets"
-            aria-current={pathname === '/datasets' ? 'page' : undefined}
-            onClick={onMobileClose}
            className={cn(TAB_BASE, pathname === '/datasets' && TAB_ACTIVE)}
          >
            Datasets
          </Link>
          <Link
            href="/gallery"
-            aria-current={pathname === '/gallery' ? 'page' : undefined}
-            onClick={onMobileClose}
            className={cn(TAB_BASE, pathname === '/gallery' && TAB_ACTIVE)}
          >
            Gallery
          </Link>
          <Link
            href="/gpus"
-            aria-current={pathname === '/gpus' ? 'page' : undefined}
-            onClick={onMobileClose}
            className={cn(TAB_BASE, pathname === '/gpus' && TAB_ACTIVE)}
          >
            GPUs
          </Link>
          <Link
            href="/settings"
-            aria-current={pathname === '/settings' ? 'page' : undefined}
-            onClick={onMobileClose}
            className={cn(TAB_BASE, pathname === '/settings' && TAB_ACTIVE)}
          >
            Settings
@@ -177,7 +134,7 @@ export default function PrimarySidebar({
        </nav>
      )}

-      {!isMobile && <div
+      <div
        className={cn(
          'absolute bottom-0 p-2',
          isCollapsed ? '-right-[60px] top-0' : 'right-0',
@@ -188,7 +145,8 @@ export default function PrimarySidebar({
          onClick={() => setIsCollapsed((v) => !v)}
          title={isCollapsed ? 'Expand sidebar' : 'Collapse sidebar'}
          className={cn(
-            'flex size-11 items-center justify-center rounded-lg text-muted-foreground transition-colors hover:bg-accent hover:text-foreground',
+            'flex items-center justify-center rounded-lg text-muted-foreground transition-colors hover:bg-accent hover:text-foreground',
+            isCollapsed ? 'p-3' : 'p-2',
          )}
        >
          <svg
@@ -201,9 +159,9 @@ export default function PrimarySidebar({
            <path d={isCollapsed ? 'M9 18l6-6-6-6' : 'M15 18l-6-6 6-6'} />
          </svg>
        </button>
-      </div>}
+      </div>

-      {!isMobile && !isCollapsed && (
+      {!isCollapsed && (
        <div
          role="presentation"
          onMouseDown={onMouseDown}
@@ -1,4 +1,4 @@
-import { act, render, screen } from '@testing-library/react';
+import { render, screen } from '@testing-library/react';
 import { beforeEach, describe, expect, it, vi } from 'vitest';

 import GpuGrid from './GpuGrid';
@@ -73,30 +73,4 @@ describe('GpuGrid', () => {
      await screen.findByText(/Could not reach the API server/),
    ).toBeInTheDocument();
  });
-
-  it('keeps the last snapshot visible and warns when a refresh fails', async () => {
-    vi.useFakeTimers();
-    try {
-      vi.mocked(getGpus)
-        .mockResolvedValueOnce(SNAPSHOT)
-        .mockRejectedValueOnce(new Error('network down'));
-
-      render(<GpuGrid />);
-      await act(async () => {
-        await vi.advanceTimersByTimeAsync(0);
-      });
-      expect(screen.getAllByText('NVIDIA B200')).toHaveLength(2);
-
-      await act(async () => {
-        await vi.advanceTimersByTimeAsync(3000);
-      });
-
-      expect(
-        screen.getByText(/values below may be stale/),
-      ).toBeInTheDocument();
-      expect(screen.getAllByText('NVIDIA B200')).toHaveLength(2);
-    } finally {
-      vi.useRealTimers();
-    }
-  });
 });
@@ -1,9 +1,7 @@
 'use client';

 import * as React from 'react';
-import { AlertTriangle } from 'lucide-react';

-import { Button } from '@/components/ui/button';
 import { Card, CardContent } from '@/components/ui/card';
 import { getGpus, type GpuInfo, type GpuSnapshot } from '@/lib/api';
 import { cn } from '@/lib/utils';
@@ -99,8 +97,7 @@ function GpuCard({ gpu }: { gpu: GpuInfo }) {

 export default function GpuGrid() {
  const [snapshot, setSnapshot] = React.useState<GpuSnapshot | null>(null);
-  const [fetchError, setFetchError] = React.useState<string | null>(null);
-  const [retryToken, setRetryToken] = React.useState(0);
+  const [fetchError, setFetchError] = React.useState(false);

  React.useEffect(() => {
    let mounted = true;
@@ -113,14 +110,10 @@ export default function GpuGrid() {
        const next = await getGpus();
        if (mounted) {
          setSnapshot(next);
-          setFetchError(null);
+          setFetchError(false);
        }
      } catch {
-        if (mounted) {
-          setFetchError(
-            'GPU status could not be refreshed. The values below may be stale.',
-          );
-        }
+        if (mounted) setFetchError(true);
      } finally {
        inFlight = false;
      }
@@ -132,27 +125,14 @@ export default function GpuGrid() {
      mounted = false;
      clearInterval(interval);
    };
-  }, [retryToken]);
+  }, []);

  if (fetchError && !snapshot) {
    return (
-      <div
-        role="alert"
-        className="flex flex-col items-center gap-3 py-8 text-center"
-      >
-        <AlertTriangle className="size-6 text-destructive" aria-hidden />
-        <p className="text-muted-foreground">
-          Could not reach the API server. GPU status needs the Studio API server
-          running.
-        </p>
-        <Button
-          type="button"
-          variant="outline"
-          onClick={() => setRetryToken((token) => token + 1)}
-        >
-          Try Again
-        </Button>
-      </div>
+      <p className="py-8 text-center text-muted-foreground">
+        Could not reach the API server. GPU status needs the studio API server
+        running.
+      </p>
    );
  }
  if (!snapshot) {
@@ -170,22 +150,9 @@ export default function GpuGrid() {
  return (
    <div className="flex flex-col gap-4">
      {fetchError && (
-        <div
-          role="status"
-          aria-live="polite"
-          className="flex flex-wrap items-center gap-3 rounded-lg border border-amber-500/50 bg-amber-500/10 px-3 py-2 text-sm"
-        >
-          <AlertTriangle className="size-4 text-amber-600" aria-hidden />
-          <span className="min-w-0 flex-1">{fetchError}</span>
-          <Button
-            type="button"
-            variant="outline"
-            size="sm"
-            onClick={() => setRetryToken((token) => token + 1)}
-          >
-            Refresh Now
-          </Button>
-        </div>
+        <p className="rounded-md border border-amber-500/40 bg-amber-500/10 px-3 py-2 text-sm text-amber-600 dark:text-amber-400">
+          Lost contact with the API server — showing the last known values.
+        </p>
      )}
      <div className="grid gap-4 [grid-template-columns:repeat(auto-fill,minmax(280px,1fr))]">
        {snapshot.gpus.map((gpu) => (
@@ -1,48 +0,0 @@
-import { render, screen } from '@testing-library/react';
-import { describe, expect, it } from 'vitest';
-
-import { Button } from './button';
-import { Input } from './input';
-import { NativeSelect } from './native-select';
-import { Slider } from './slider';
-import { Switch } from './switch';
-
-describe('shared control accessibility', () => {
-  it('keeps button, input, and select targets at least 44px tall', () => {
-    render(
-      <>
-        <Button size="sm">Small action</Button>
-        <Input aria-label="Text value" />
-        <NativeSelect aria-label="Choice" defaultValue="one">
-          <option value="one">One</option>
-        </NativeSelect>
-      </>,
-    );
-
-    expect(screen.getByRole('button', { name: 'Small action' })).toHaveClass(
-      'h-11',
-    );
-    expect(screen.getByRole('textbox', { name: 'Text value' })).toHaveClass(
-      'h-11',
-    );
-    expect(screen.getByRole('combobox', { name: 'Choice' })).toHaveClass(
-      'h-11',
-    );
-  });
-
-  it('uses 44px switch and slider interaction surfaces', () => {
-    render(
-      <>
-        <Switch aria-label="Enabled" />
-        <Slider aria-label="Amount" defaultValue={[50]} />
-      </>,
-    );
-
-    expect(screen.getByRole('switch', { name: 'Enabled' })).toHaveClass(
-      'size-11',
-    );
-    expect(screen.getByRole('slider', { name: 'Amount' })).toHaveClass(
-      'size-11',
-    );
-  });
-});
@@ -29,12 +29,11 @@ const badgeVariants = cva(
 );

 export interface BadgeProps
-  extends React.HTMLAttributes<HTMLSpanElement>,
+  extends React.HTMLAttributes<HTMLDivElement>,
    VariantProps<typeof badgeVariants> {}

-// A span (phrasing content), so badges stay valid inside buttons and links.
 function Badge({ className, variant, ...props }: BadgeProps) {
-  return <span className={cn(badgeVariants({ variant }), className)} {...props} />;
+  return <div className={cn(badgeVariants({ variant }), className)} {...props} />;
 }

 export { Badge, badgeVariants };
@@ -7,7 +7,7 @@ import { cva, type VariantProps } from "class-variance-authority";
 import { cn } from "@/lib/utils";

 const buttonVariants = cva(
-	"inline-flex items-center justify-center gap-2 whitespace-nowrap rounded-xl border !text-sm !font-semibold transition-colors duration-150 disabled:pointer-events-none disabled:cursor-not-allowed disabled:opacity-50",
+	"inline-flex items-center justify-center gap-2 whitespace-nowrap rounded-xl border !text-sm !font-semibold transition-colors duration-150 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-sky-400/40 disabled:pointer-events-none disabled:cursor-not-allowed disabled:opacity-50",
 	{
 		variants: {
 			variant: {
@@ -18,11 +18,11 @@ const buttonVariants = cva(
 				destructive: "border-rose-500/60 bg-rose-600/90 text-white hover:bg-rose-500",
 			},
 			size: {
-				default: "h-11 px-4 py-2",
-				sm: "h-11 rounded-lg px-3 !text-xs",
-				lg: "h-12 px-5 !text-sm",
-				icon: "size-11",
-				"icon-sm": "size-11",
+				default: "h-10 px-4 py-2",
+				sm: "h-9 rounded-lg px-3 !text-xs",
+				lg: "h-11 px-5 !text-sm",
+				icon: "size-10",
+				"icon-sm": "size-8",
 			},
 		},
 		defaultVariants: {
@@ -42,7 +42,7 @@ const DialogContent = React.forwardRef<
      {...props}
    >
      {children}
-      <DialogPrimitive.Close className="absolute right-2 top-2 flex size-11 items-center justify-center rounded-lg text-muted-foreground opacity-70 transition-opacity hover:bg-secondary hover:opacity-100 disabled:pointer-events-none sm:right-4 sm:top-4">
+      <DialogPrimitive.Close className="absolute right-4 top-4 rounded-lg p-1 text-muted-foreground opacity-70 transition-opacity hover:bg-secondary hover:opacity-100 focus:outline-none focus:ring-2 focus:ring-sky-400/40 disabled:pointer-events-none">
        <X className="h-4 w-4" />
        <span className="sr-only">Close</span>
      </DialogPrimitive.Close>
@@ -9,7 +9,7 @@ const Input = React.forwardRef<HTMLInputElement, React.ComponentProps<"input">>(
 		<input
 			type={type}
 			className={cn(
-				"flex h-11 w-full rounded-xl border border-input bg-card/60 px-3 py-2 text-sm text-foreground shadow-sm backdrop-blur-md transition-colors placeholder:text-muted-foreground focus-visible:border-ring disabled:cursor-not-allowed disabled:opacity-50",
+				"flex h-10 w-full rounded-xl border border-input bg-card/60 px-3 py-2 text-sm text-foreground shadow-sm backdrop-blur-md transition-colors placeholder:text-muted-foreground focus-visible:border-sky-400/70 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-sky-400/25 disabled:cursor-not-allowed disabled:opacity-50",
 				className,
 			)}
 			ref={ref}
@@ -11,7 +11,7 @@ const NativeSelect = React.forwardRef<
  <select
    ref={ref}
    className={cn(
-      'flex h-11 w-full appearance-none rounded-xl border border-input bg-card px-3 py-2 text-sm text-foreground shadow-sm transition-colors focus-visible:border-ring disabled:cursor-not-allowed disabled:opacity-50',
+      'flex h-10 w-full appearance-none rounded-xl border border-input bg-card px-3 py-2 text-sm text-foreground shadow-sm transition-colors focus-visible:border-sky-400/70 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-sky-400/25 disabled:cursor-not-allowed disabled:opacity-50',
      className,
    )}
    {...props}
@@ -17,7 +17,7 @@ const SelectTrigger = React.forwardRef<
  <SelectPrimitive.Trigger
    ref={ref}
    className={cn(
-      'flex h-11 w-full items-center justify-between gap-2 rounded-xl border border-input bg-card px-3 py-2 text-sm text-foreground shadow-sm transition-colors placeholder:text-muted-foreground focus-visible:border-ring disabled:cursor-not-allowed disabled:opacity-50 [&>span]:line-clamp-1',
+      'flex h-10 w-full items-center justify-between gap-2 rounded-xl border border-input bg-card px-3 py-2 text-sm text-foreground shadow-sm outline-none transition-colors placeholder:text-muted-foreground focus:border-sky-400/70 focus:ring-2 focus:ring-sky-400/25 disabled:cursor-not-allowed disabled:opacity-50 [&>span]:line-clamp-1',
      className,
    )}
    {...props}
@@ -117,7 +117,7 @@ const SelectItem = React.forwardRef<
  <SelectPrimitive.Item
    ref={ref}
    className={cn(
-      'relative flex min-h-11 w-full cursor-default select-none items-center rounded-xl py-2 pl-8 pr-3 text-sm text-foreground outline-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 data-[highlighted]:bg-accent data-[highlighted]:text-accent-foreground',
+      'relative flex w-full cursor-default select-none items-center rounded-xl py-2 pl-8 pr-3 text-sm text-foreground outline-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 data-[highlighted]:bg-accent data-[highlighted]:text-accent-foreground',
      className,
    )}
    {...props}
@@ -8,39 +8,21 @@ import { cn } from '@/lib/utils';
 const Slider = React.forwardRef<
  React.ElementRef<typeof SliderPrimitive.Root>,
  React.ComponentPropsWithoutRef<typeof SliderPrimitive.Root>
->(
-  (
-    {
+>(({ className, ...props }, ref) => (
+  <SliderPrimitive.Root
+    ref={ref}
+    className={cn(
+      'relative flex w-full touch-none select-none items-center',
      className,
-      id,
-      'aria-label': ariaLabel,
-      'aria-labelledby': ariaLabelledBy,
-      'aria-describedby': ariaDescribedBy,
-      ...props
-    },
-    ref,
-  ) => (
-    <SliderPrimitive.Root
-      ref={ref}
-      className={cn(
-        'relative flex h-11 w-full touch-none select-none items-center',
-        className,
-      )}
-      {...props}
-    >
-      <SliderPrimitive.Track className="relative h-1.5 w-full grow overflow-hidden rounded-full bg-border">
-        <SliderPrimitive.Range className="absolute h-full bg-accent-blue" />
-      </SliderPrimitive.Track>
-      <SliderPrimitive.Thumb
-        id={id}
-        aria-label={ariaLabel}
-        aria-labelledby={ariaLabelledBy}
-        aria-describedby={ariaDescribedBy}
-        className="relative block size-11 rounded-full bg-transparent after:absolute after:left-1/2 after:top-1/2 after:size-4 after:-translate-x-1/2 after:-translate-y-1/2 after:rounded-full after:border-2 after:border-accent-blue after:bg-accent-blue after:shadow after:content-[''] hover:after:border-accent-blue/80 disabled:pointer-events-none disabled:opacity-50"
-      />
-    </SliderPrimitive.Root>
-  ),
-);
+    )}
+    {...props}
+  >
+    <SliderPrimitive.Track className="relative h-1.5 w-full grow overflow-hidden rounded-full bg-border">
+      <SliderPrimitive.Range className="absolute h-full bg-accent-blue" />
+    </SliderPrimitive.Track>
+    <SliderPrimitive.Thumb className="block h-4 w-4 rounded-full border-2 border-accent-blue bg-accent-blue shadow transition-colors hover:border-accent-blue/80 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-sky-400/40 disabled:pointer-events-none disabled:opacity-50" />
+  </SliderPrimitive.Root>
+));
 Slider.displayName = SliderPrimitive.Root.displayName;

 export { Slider };
@@ -12,14 +12,14 @@ const Switch = React.forwardRef<
  <SwitchPrimitives.Root
    ref={ref}
    className={cn(
-      'peer relative inline-flex size-11 shrink-0 cursor-pointer items-center justify-center rounded-xl bg-transparent transition-colors before:absolute before:h-5 before:w-9 before:rounded-full before:border before:border-border before:bg-background data-[state=checked]:before:border-accent-blue data-[state=checked]:before:bg-accent-blue disabled:cursor-not-allowed disabled:opacity-50',
+      'peer inline-flex h-5 w-9 shrink-0 cursor-pointer items-center rounded-full border border-border transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-sky-400/40 disabled:cursor-not-allowed disabled:opacity-50 data-[state=checked]:border-accent-blue data-[state=checked]:bg-accent-blue data-[state=unchecked]:bg-background',
      className,
    )}
    {...props}
  >
    <SwitchPrimitives.Thumb
      className={cn(
-        'pointer-events-none absolute left-1.5 top-[15px] z-[1] block h-3.5 w-3.5 rounded-full bg-muted-foreground shadow-lg ring-0 transition-transform data-[state=checked]:translate-x-4 data-[state=checked]:bg-white',
+        'pointer-events-none block h-3.5 w-3.5 rounded-full bg-muted-foreground shadow-lg ring-0 transition-transform data-[state=checked]:translate-x-4 data-[state=checked]:bg-white data-[state=unchecked]:translate-x-0.5',
      )}
    />
  </SwitchPrimitives.Root>
@@ -29,7 +29,7 @@ const TabsTrigger = React.forwardRef<
  <TabsPrimitive.Trigger
    ref={ref}
    className={cn(
-      'inline-flex items-center justify-center whitespace-nowrap rounded-md px-3 py-1.5 text-sm font-medium text-muted-foreground transition-colors hover:text-foreground disabled:pointer-events-none disabled:opacity-50 data-[state=active]:bg-card data-[state=active]:text-foreground data-[state=active]:shadow-sm',
+      'inline-flex items-center justify-center whitespace-nowrap rounded-md px-3 py-1.5 text-sm font-medium text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-sky-400/40 disabled:pointer-events-none disabled:opacity-50 data-[state=active]:bg-card data-[state=active]:text-foreground data-[state=active]:shadow-sm',
      className,
    )}
    {...props}
@@ -8,7 +8,7 @@ const Textarea = React.forwardRef<HTMLTextAreaElement, React.ComponentProps<"tex
 	return (
 		<textarea
 			className={cn(
-				"flex min-h-24 w-full resize-y rounded-xl border border-input bg-card px-3 py-2 text-sm text-foreground shadow-sm transition-colors placeholder:text-muted-foreground focus-visible:border-ring disabled:cursor-not-allowed disabled:opacity-50",
+				"flex min-h-24 w-full resize-y rounded-xl border border-input bg-card px-3 py-2 text-sm text-foreground shadow-sm transition-colors placeholder:text-muted-foreground focus-visible:border-sky-400/70 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-sky-400/25 disabled:cursor-not-allowed disabled:opacity-50",
 				className,
 			)}
 			ref={ref}
@@ -1,23 +0,0 @@
-'use client';
-
-import * as React from 'react';
-
-/**
- * Move focus into a drawer when it opens as a modal (mobile), and hand it
- * back to the previously focused element on close. Pairs with `inert` on
- * the background content — together they make `aria-modal` truthful.
- */
-export function useDrawerFocus<T extends HTMLElement>(active: boolean) {
-  const ref = React.useRef<T | null>(null);
-
-  React.useEffect(() => {
-    if (!active) return;
-    const previous = document.activeElement;
-    ref.current?.focus();
-    return () => {
-      if (previous instanceof HTMLElement) previous.focus();
-    };
-  }, [active]);
-
-  return ref;
-}
@@ -1,18 +0,0 @@
-'use client';
-
-import * as React from 'react';
-
-export function useMediaQuery(query: string): boolean {
-  const [matches, setMatches] = React.useState(false);
-
-  React.useEffect(() => {
-    const mediaQuery = window.matchMedia(query);
-    const updateMatch = () => setMatches(mediaQuery.matches);
-
-    updateMatch();
-    mediaQuery.addEventListener('change', updateMatch);
-    return () => mediaQuery.removeEventListener('change', updateMatch);
-  }, [query]);
-
-  return matches;
-}
@@ -186,40 +186,15 @@ RUN --mount=type=cache,target=/opt/uv/cache \

 COPY . .

-# Build immutable FastVideo kernel wheels for the published image. The requested
-# architecture remains installed for normal image users; amd64 images also carry
-# an SM89 artifact so the predominant L40S Modal lanes can reuse it exactly.
-ARG FASTVIDEO_KERNEL_PREBUILT_DIR=/opt/fastvideo-kernel-prebuilt
+# Install FastVideo Unified Kernel exactly once. build.sh initializes only its
+# CUTLASS/ThunderKittens submodules, then compiles for TORCH_CUDA_ARCH_LIST
+# (default Hopper sm_90a) without requiring a live GPU.
 RUN --mount=type=cache,target=/opt/uv/cache \
    source $HOME/.local/bin/env && \
    source /opt/venv/bin/activate && \
-    default_arch="${TORCH_CUDA_ARCH_LIST}" && \
-    default_wheel_dir="${FASTVIDEO_KERNEL_PREBUILT_DIR}/${default_arch}" && \
-    export TORCH_CUDA_ARCH_LIST="${default_arch}" && \
    cd fastvideo-kernel && \
-    CMAKE_ARGS= CMAKE_BUILD_PARALLEL_LEVEL=${CMAKE_BUILD_PARALLEL_LEVEL} \
-    ./build.sh --wheel-dir "${default_wheel_dir}" && \
-    cd /FastVideo && \
-    CMAKE_ARGS= python fastvideo/tests/modal/kernel_build_cache.py write-build-info \
-        --wheel-dir "${default_wheel_dir}" \
-        --output "${default_wheel_dir}/metadata.json" && \
-    if [[ "${TARGETARCH:-amd64}" == "amd64" && "${default_arch}" != "8.9" ]]; then \
-        export TORCH_CUDA_ARCH_LIST=8.9 && \
-        l40s_wheel_dir="${FASTVIDEO_KERNEL_PREBUILT_DIR}/8.9" && \
-        cd /FastVideo/fastvideo-kernel && \
-        CMAKE_ARGS= CMAKE_BUILD_PARALLEL_LEVEL=${CMAKE_BUILD_PARALLEL_LEVEL} \
-        ./build.sh --wheel-dir "${l40s_wheel_dir}" && \
-        cd /FastVideo && \
-        CMAKE_ARGS= python fastvideo/tests/modal/kernel_build_cache.py write-build-info \
-            --wheel-dir "${l40s_wheel_dir}" \
-            --output "${l40s_wheel_dir}/metadata.json"; \
-    fi && \
-    export TORCH_CUDA_ARCH_LIST="${default_arch}" && \
-    default_wheel="$(find "${default_wheel_dir}" -maxdepth 1 -type f \
-        \( -name 'fastvideo_kernel-*.whl' -o -name 'fastvideo-kernel-*.whl' \) \
-        | sort | tail -n 1)" && \
-    uv pip install "${default_wheel}" \
-        --reinstall-package fastvideo-kernel --no-deps
+    CMAKE_BUILD_PARALLEL_LEVEL=${CMAKE_BUILD_PARALLEL_LEVEL} \
+    TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST} ./build.sh

 # Install FastVideo itself (editable) now that the source is present, and set up
 # shell configuration. Dependencies and the local kernel are already installed,
@@ -8,15 +8,13 @@ FastVideo provides highly optimized custom attention kernels to accelerate video
 * **[Sliding Tile Attention (STA)](sta/index.md)**: STA kernel support is kept in
  `fastvideo-kernel`; full FastVideo STA pipeline workflow is archived in
  `sta_do_not_delete`.
-* **[Attn-QAT Training](../training/attn_qat.md)**: Runtime-JIT Triton forward
-  and backward kernels for role-local quantization-aware training.
 * **Backend development guide**: See the developer guide at
  [Attention Backend Development](../contributing/attention_backend.md).

 ## General Build Instructions

 These instructions apply to building the `fastvideo-kernel` package from
-source, which includes STA, VSA, and Attn-QAT kernels.
+source, which includes both STA and VSA kernels.

 ### Prerequisites

@@ -85,6 +85,7 @@ status.
 | Transformer Tests | `transformer` | `fastvideo/models/dits/**`, `fastvideo/models/loader/**`, `fastvideo/tests/transformers/**`, `fastvideo/layers/**`, `fastvideo/attention/**`, `pyproject.toml`, `docker/Dockerfile` |
 | Kernel Tests | `kernel_tests` | `fastvideo-kernel/**`, `pyproject.toml`, `docker/Dockerfile` |
 | Unit Tests | `unit_test` | `fastvideo/**`, `.buildkite/**`, `.github/**`, `pyproject.toml`, `docker/Dockerfile` |
+| Preprocessing Tests | `preprocessing` | Preprocessing pipelines/workflows, Parquet schema/writer, the integration test, and its CI entrypoints |
 | DreamVerse App Tests | `dreamverse_app` | `apps/dreamverse/**`, `pyproject.toml` |

 ### Tier 3: Full Suite
@@ -147,6 +148,7 @@ Valid direct test names:
 | `/test transformer` | `transformer` |
 | `/test kernel` | `kernel_tests` |
 | `/test unit` | `unit_test` |
+| `/test preprocessing` | `preprocessing` |
 | `/test dreamverse` | `dreamverse_app` |
 | `/test ssim` | `ssim` |
 | `/test training` | `training` |
@@ -255,9 +257,10 @@ If you add a new CI test category:

 ### Documentation

-`.github/workflows/infra-docs.yml` builds documentation for PRs that touch
-`docs/**`, `mkdocs.yml`, `requirements-mkdocs.txt`, or the workflow itself. On
-pushes to `main`, it also deploys the built site to GitHub Pages.
+`.github/workflows/infra-docs.yml` builds documentation for same-repository PRs
+that touch `docs/**`, `mkdocs.yml`, `requirements-mkdocs.txt`, or the workflow
+itself. Fork PRs skip this executable build instead of waiting for maintainer
+approval. On pushes to `main`, it also deploys the built site to GitHub Pages.

 The docs job:

@@ -269,26 +272,16 @@ The docs job:
 ### Docker Images

 `.github/workflows/infra-build-image.yml` supports manual `workflow_dispatch`
-runs and automatically rebuilds the CUDA matrix when a repository-controlled
-image input changes on `main` in the canonical repository. Those inputs include
-the CUDA Dockerfile and reusable workflow, dependency metadata, Docker context
-policy, `fastvideo-kernel/**`, and the kernel artifact metadata/key helper.
-Manual runs let maintainers choose which image families to build. The
-`fastvideo-dev` matrix builds Python 3.12 images for CUDA 12.6 and CUDA 13 on
-native `amd64` and `arm64` runners, then publishes one multi-platform manifest
-per CUDA version. CUDA 12.6 owns the `py3.12-latest` and global `latest` tags, as
-well as the explicit `py3.12-cuda12.6.3-latest` alias. CUDA 13 is published under
-the explicit `py3.12-cuda13.0.0-latest` tag. This publication policy does not
-change the unparameterized `docker/Dockerfile` build defaults, which remain CUDA
-13 and `cu130`.
-
-Published amd64 development images keep their configured Hopper kernel wheel
-installed and also carry an immutable SM89 wheel under
-`/opt/fastvideo-kernel-prebuilt`. Modal PR and SSIM jobs select the exact
-source, ABI, and GPU-architecture match from that directory, so L40S jobs reuse
-the trusted image artifact while kernel-changing PRs still build locally. Once
-a kernel or artifact-key change reaches `main`, the image workflow republishes
-the matching trusted artifact before later jobs consume the updated image tag.
+runs and automatically rebuilds the CUDA matrix when `docker/Dockerfile`
+changes on `main` in the canonical repository. Manual runs let maintainers
+choose which image families to build. The `fastvideo-dev` matrix builds Python
+3.12 images for CUDA 12.6 and CUDA 13 on native `amd64` and `arm64` runners,
+then publishes one multi-platform manifest per CUDA version. CUDA 12.6 owns the
+`py3.12-latest` and global `latest` tags, as well as the explicit
+`py3.12-cuda12.6.3-latest` alias. CUDA 13 is published under the explicit
+`py3.12-cuda13.0.0-latest` tag. This publication policy does not change the
+unparameterized `docker/Dockerfile` build defaults, which remain CUDA 13 and
+`cu130`.

 The optional Dreamverse matrix builds backend and UI images for CUDA 12.6 and
 CUDA 13 on `amd64`. Dreamverse remains `amd64`-only because its FA4 dependency
@@ -333,25 +333,15 @@ surfaces:
      use_distill:
        sources: [fastvideo.configs.pipelines.longcat.LongCatT2V480PConfig, fastvideo.configs.pipelines.longcat.LongCatT2V704PConfig]
      scheduler_arch:
-        sources:
-          - fastvideo.configs.pipelines.sd35.SD35Config
-          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
      text_encoder_archs:
-        sources:
-          - fastvideo.configs.pipelines.sd35.SD35Config
-          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
      tokenizer_archs:
-        sources:
-          - fastvideo.configs.pipelines.sd35.SD35Config
-          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
      transformer_arch:
-        sources:
-          - fastvideo.configs.pipelines.sd35.SD35Config
-          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
      vae_arch:
-        sources:
-          - fastvideo.configs.pipelines.sd35.SD35Config
-          - fastvideo.configs.pipelines.zimage.ZImagePipelineConfig
+        sources: [fastvideo.configs.pipelines.sd35.SD35Config]
      expand_timesteps:
        sources:
          - fastvideo.configs.pipelines.wan.FastWan2_2_TI2V_5B_Config
@@ -431,8 +421,6 @@ surfaces:
      frame_receptive_field: "MagiHuman internal data-proxy receptive-field setting."
      image_conditioning: "MagiHuman preset variant marker for reference-image conditioning."
      ref_audio_offset: "MagiHuman internal data-proxy audio alignment offset."
-      scheduler_sigma_min: "Z-Image scheduler parity invariant; not part of the public typed inference API."
-      scheduler_use_reference_discrete_timesteps: "Z-Image scheduler parity invariant; not part of the public typed inference API."
      sr_local_attn_layers: "MagiHuman SR internal sparse-attention layer selection."
      text_offset: "MagiHuman internal data-proxy text alignment offset."
      vae_stride: "MagiHuman internal VAE/data-proxy stride setting."
@@ -450,7 +438,6 @@ surfaces:
      grid_sizes: request.inputs.grid_sizes
      pose: request.inputs.pose
      c2ws_plucker_emb: request.inputs.c2ws_plucker_emb
-      action_path: request.inputs.action_path
      refine_from: request.inputs.refine_from
      stage1_video: request.inputs.stage1_video
      prompt: request.prompt
@@ -460,7 +447,6 @@ surfaces:
      output_video_name: request.output.output_video_name
      num_videos_per_prompt: request.sampling.num_videos_per_prompt
      seed: request.sampling.seed
-      max_sequence_length: request.sampling.max_sequence_length
      num_frames: request.sampling.num_frames
      height: request.sampling.height
      width: request.sampling.width
@@ -470,10 +456,7 @@ surfaces:
      num_inference_steps: request.sampling.num_inference_steps
      num_inference_steps_sr: request.sampling.num_inference_steps_sr
      guidance_scale: request.sampling.guidance_scale
-      batch_cfg: request.sampling.batch_cfg
      guidance_scale_2: request.sampling.guidance_scale_2
-      cfg_normalization: request.sampling.cfg_normalization
-      cfg_truncation: request.sampling.cfg_truncation
      guidance_rescale: request.sampling.guidance_rescale
      use_embedded_guidance: request.sampling.use_embedded_guidance
      true_cfg_scale: request.sampling.true_cfg_scale
@@ -525,6 +508,7 @@ surfaces:
    internal_only:
      data_type: "Derived from the request shape and not a public input."
      latents: "Pre-generated diffusion latents supplied by parity/debug harnesses; not a public input."
+      max_sequence_length: "Model-specific text-encoder sequence cap; not part of the public typed inference API."

  sampling_param_extensions: {}

@@ -2,12 +2,6 @@

 We introduce a new finetuning strategy - **Sparse-distill**, which jointly integrates **[DMD](https://arxiv.org/abs/2405.14867)** and **[VSA](https://arxiv.org/abs/2505.13389)** in a single training process. This approach combines the benefits of both distillation to shorten diffusion steps and sparse attention to reduce attention computation, enabling much faster video generation.

-!!! tip "Attn-QAT DMD2 workflow"
-    The modular trainer also provides a Wan2.1 MixKit recipe that first
-    fine-tunes with fake-quantized attention, then distills the student to
-    timesteps `[1000, 757, 522]` while teacher and critic remain on Flash
-    Attention. See [Attn-QAT Training](../training/attn_qat.md).
-
 ## 📊 Model Overview

 We provide two distilled models:
@@ -134,4 +134,4 @@ If you're planning to contribute to FastVideo please see the following page:

 If you encounter any issues during installation, please open an issue on our [GitHub repository](https://github.com/hao-ai-lab/FastVideo).

-You can also join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ) for additional support.
+You can also join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ) for additional support.
@@ -100,4 +100,4 @@ If you're planning to contribute to FastVideo please see the following page:

 If you encounter any issues during installation, please open an issue on our [GitHub repository](https://github.com/hao-ai-lab/FastVideo).

-You can also join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ) for additional support.
+You can also join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ) for additional support.
@@ -134,7 +134,7 @@ uv pip install "https://github.com/mjun0812/flash-attention-prebuild-wheels/rele

 If you hit other issues, please open an issue on our
 [GitHub repository](https://github.com/hao-ai-lab/FastVideo). You can also join
-our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ)
+our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ)
 for additional support.

 ## Development Environment Setup
@@ -11,7 +11,7 @@ This guide explains how to implement a custom diffusion pipeline in FastVideo, l
 4. **Register Your Pipeline** - Make it discoverable by the framework
 5. **Configure Your Pipeline** - (Coming soon)

-Need help? Join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ).
+Need help? Join our [Slack community](https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ).

 ## Step 1: Pipeline Modules

@@ -134,4 +134,5 @@ If the generated video doesn't match your prompt:
 - Learn about [Advanced Inference Configurations](configuration.md)
 - Learn about using [Optimizations](optimizations.md)
 - See [Examples](examples/examples_inference_index.md) for more usage scenarios
- Join our [Community Slack](https://join.slack.com/t/fastvideo/shared_invite/zt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ).
+- Join our [Community Discord](https://discord.gg/JA7cksDz86).
+- Join our [Community Slack](https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ).
@@ -119,22 +119,6 @@ pip install "nvidia-cutlass-dsl>=4.5.2" apache-tvm-ffi flashinfer-python
 The `--no-deps` flag prevents upgrading torch/torchvision. Use the supported
 PyTorch 2.12.0 and CUDA 13 environment for this kernel.

-Branch-to-`nvidia-cutlass-dsl` compatibility (the fork tracks the CuTe DSL API
-surface closely):
-
-| fork branch | cutlass-dsl | notes |
-|---|---|---|
-| `fp4` | `==4.4.2` (+ `nvidia-cutlass-dsl-libs-base==4.4.2`) | validated set on GB200: `quack-kernels==0.4.1`, `flashinfer-python==0.6.8`, `CUTE_DSL_ENABLE_TVM_FFI=1`, `FASTVIDEO_FA4=1` |
-| `fix/cutlass-dsl-4.5` | `>=4.5.2` | carries the `cute.core.ThrMma` -> `cute.ThrMma` fix |
-| any | 4.6-era | unsupported: `cute.make_fragment` was removed at module level; fails at CuTe JIT trace |
-
-`FASTVIDEO_FA4=1` is required alongside the fork: it ships no compiled
-FlashAttention-2, so dense attention paths raise ImportError without the FA4
-opt-in. The same kernel also serves `ATTN_QAT_INFER` on sm_100a/sm_103a
-(datacenter Blackwell) — the selection log's receipt line
-(`ATTN_QAT_INFER resolved: ...`) records the arch, kernel, and quantization
-mode that actually bound.
-
 #### Usage

 Enable FP4 attention via the `--nvfp4_fa4` flag:
@@ -13,92 +13,6 @@ For the canonical, code-level list of model IDs recognized by
    We do this because we believe VSA is strictly better than STA for the
    actively maintained `main` inference path.

-## Registered Model IDs
-
-Every Hugging Face model ID registered in `fastvideo/registry.py` on `main`
-(commit `8d89f30d`), grouped by family. Any ID below can
-be passed to `VideoGenerator.from_pretrained(...)`; FastVideo resolves the
-matching pipeline and sampling defaults. The **Family** column is a
-documentation grouping: it follows each registration's declared `model_family`,
-except `black-forest-labs/FLUX.1-dev`, which declares none and is listed under
-`flux` for readability. The **Workloads** column shows each
-registration's declared `workload_types`; `—` means the entry is registered
-without a UI workload option but is still loadable by ID. The **Example**
-column links a runnable script in `examples/inference/basic/` where one exists.
-
-| Family | HuggingFace Model ID | Workloads | Example |
-|--------|----------------------|-----------|---------|
-| cosmos | `nvidia/Cosmos-Predict2-2B-Video2World` | T2V | — |
-| cosmos25 | `KyleShao/Cosmos-Predict2.5-2B-Diffusers` | T2V | [basic_cosmos2_5_t2w.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_cosmos2_5_t2w.py) |
-| cosmos25 | `nvidia/Cosmos-Predict2.5-14B` | T2V | [basic_cosmos2_5_t2w.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_cosmos2_5_t2w.py) |
-| dreamx_world | `FastVideo/DreamX-World-5B-Cam-Diffusers` | I2V | [basic_dreamx_world.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_dreamx_world.py) |
-| dreamx_world | `FastVideo/DreamX-World-5B-Diffusers` | I2V | [basic_dreamx_world.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_dreamx_world.py) |
-| flux | `black-forest-labs/FLUX.1-dev` | T2I | [basic_flux_dev.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_flux_dev.py) |
-| flux2 | `black-forest-labs/FLUX.2-klein-4B`<br>`black-forest-labs/FLUX.2-klein-9B` | T2I | [basic_flux2_klein.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_flux2_klein.py) |
-| flux2 | `black-forest-labs/FLUX.2-dev` | T2I | [basic_flux2.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_flux2.py) |
-| gamecraft | `FastVideo/HunyuanGameCraft-Diffusers` | I2V | [basic_gamecraft.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_gamecraft.py) |
-| gen3c | `FastVideo/GEN3C-Cosmos-7B-Diffusers` | T2V | [basic_gen3c.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_gen3c.py) |
-| glm_image | `zai-org/GLM-Image` | T2I | [basic_glm_image.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_glm_image.py) |
-| hunyuan | `hunyuanvideo-community/HunyuanVideo` | T2V | — |
-| hunyuan | `FastVideo/FastHunyuan-diffusers` | T2V | — |
-| hunyuan15 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v` | T2V | [basic_hy15.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_hy15.py) |
-| hunyuan15 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled` | I2V | [basic_hy15.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_hy15.py) |
-| hunyuan15 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | T2V | [basic_hy15.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_hy15.py) |
-| hunyuan15 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v_distilled` | I2V | [basic_hy15.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_hy15.py) |
-| hunyuan15 | `weizhou03/HunyuanVideo-1.5-Diffusers-1080p`<br>`weizhou03/HunyuanVideo-1.5-Diffusers-1080p-2SR` | — | [basic_hy15_1080p.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_hy15_1080p.py) |
-| hyworld | `FastVideo/HY-WorldPlay-Bidirectional-Diffusers` | — | [basic_hyworld.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_hyworld.py) |
-| kandinsky5 | `kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers` | T2V | [basic_kandinsky5_t2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_t2v.py) |
-| kandinsky5 | `kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers` | T2V | [basic_kandinsky5_t2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_t2v.py) |
-| kandinsky5 | `kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers` | T2V | [basic_kandinsky5_t2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_t2v.py) |
-| kandinsky5 | `kandinskylab/Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers` | T2V | [basic_kandinsky5_t2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_t2v.py) |
-| kandinsky5 | `kandinskylab/Kandinsky-5.0-I2V-Lite-5s-Diffusers` | I2V | [basic_kandinsky5_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_i2v.py) |
-| kandinsky5 | `kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers` | I2V | [basic_kandinsky5_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_i2v.py) |
-| kandinsky5 | `kandinskylab/Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers` | I2V | [basic_kandinsky5_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_kandinsky5_i2v.py) |
-| lingbot_video | `FastVideo/LingBot-Video-MoE-30B-A3B-Diffusers` | T2V | [basic_lingbot_video.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lingbot_video.py) |
-| lingbot_video | `FastVideo/LingBot-Video-Dense-1.3B-Diffusers` | T2V | [basic_lingbot_video.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lingbot_video.py) |
-| lingbotworld | `FastVideo/LingBot-World-Base-Cam-Diffusers` | I2V | [basic_lingbotworld_base_cam.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lingbotworld_base_cam.py) |
-| lingbotworld2 | `robbyant/lingbot-world-v2-14b-causal-fast` | I2V | [basic_lingbotworld2_causal_fast.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lingbotworld2_causal_fast.py) |
-| longcat | `FastVideo/LongCat-Video-T2V-Diffusers` | T2V | [basic_longcat_t2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_longcat_t2v.py) |
-| longcat | `FastVideo/LongCat-Video-I2V-Diffusers` | I2V | [basic_longcat_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_longcat_i2v.py) |
-| longcat | `FastVideo/LongCat-Video-VC-Diffusers` | — | [basic_longcat_vc.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_longcat_vc.py) |
-| ltx2 | `FastVideo/LTX2-Distilled-Diffusers`<br>`FastVideo/LTX2.3-Distilled-Diffusers`<br>`FastVideo/LTX-2.3-Distilled-Diffusers` | T2V | [basic_ltx2_distilled.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_ltx2_distilled.py) |
-| ltx2 | `Lightricks/LTX-2.3`<br>`FastVideo/LTX2.3-base`<br>`FastVideo/LTX2.3-Diffusers` | T2V | [basic_ltx2.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_ltx2.py) |
-| ltx2 | `Lightricks/LTX-2`<br>`FastVideo/LTX2-base`<br>`FastVideo/LTX2-Diffusers` | T2V | [basic_ltx2.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_ltx2.py) |
-| matrixgame | `FastVideo/Matrix-Game-2.0-Base-Distilled-Diffusers`<br>`FastVideo/Matrix-Game-2.0-GTA-Distilled-Diffusers`<br>`FastVideo/Matrix-Game-2.0-TempleRun-Distilled-Diffusers`<br>`FastVideo/Matrix-Game-2.0-Base-Diffusers`<br>`FastVideo/Matrix-Game-2.0-GTA-Diffusers`<br>`FastVideo/Matrix-Game-2.0-TempleRun-Diffusers`<br>`mignonjia/mg_longtuning_distilled_zelda`<br>`mignonjia/mg_sf_distilled_zelda_1k_steps`<br>`mignonjia/mg_sf_distilled_zelda`<br>`mignonjia/mg_causal_zelda`<br>`mignonjia/mg_bidirectional_zelda` | I2V | [basic_matrixgame2.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_matrixgame2.py) |
-| matrixgame | `FastVideo/Matrix-Game-3.0-Base-Distilled-Diffusers` | I2V | [basic_matrixgame3.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_matrixgame3.py) |
-| sd35 | `stabilityai/stable-diffusion-3.5-medium` | T2I | [basic_sd35_t2i.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_sd35_t2i.py) |
-| stable_audio | `FastVideo/stable-audio-open-1.0-Diffusers` | T2V | [basic_stable_audio.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_stable_audio.py) |
-| stable_audio | `FastVideo/stable-audio-open-small-Diffusers` | T2V | [basic_stable_audio_small.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_stable_audio_small.py) |
-| turbodiffusion | `loayrashid/TurboWan2.1-T2V-1.3B-Diffusers` | T2V | [basic_turbodiffusion.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_turbodiffusion.py) |
-| turbodiffusion | `loayrashid/TurboWan2.1-T2V-14B-Diffusers` | T2V | [basic_turbodiffusion_14b.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_turbodiffusion_14b.py) |
-| turbodiffusion | `loayrashid/TurboWan2.2-I2V-A14B-Diffusers` | I2V | [basic_turbodiffusion_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_turbodiffusion_i2v.py) |
-| wan | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | T2V | [basic.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic.py) |
-| wan | `Wan-AI/Wan2.1-T2V-14B-Diffusers`<br>`FastVideo/Wan2.1-VSA-T2V-14B-720P-Diffusers` | T2V | — |
-| wan | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | I2V | — |
-| wan | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` | I2V | — |
-| wan | `weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers` | I2V | — |
-| wan | `IRMChen/Wan2.1-Fun-1.3B-Control-Diffusers` | — | [basic_wan2_2_Fun.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_wan2_2_Fun.py) |
-| wan | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers`<br>`FastVideo/FastWan2.1-T2V-14B-480P-Diffusers` | T2V | [basic_dmd.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_dmd.py) |
-| wan | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | T2V, I2V | [basic_wan2_2_ti2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_wan2_2_ti2v.py) |
-| wan | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers`<br>`FastVideo/FastWan2.2-TI2V-5B-Diffusers` | T2V, I2V | [basic_dmd.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_dmd.py) |
-| wan | `decart-ai/Lucy-Edit-Dev`<br>`decart-ai/Lucy-Edit-1.1-Dev` | — | [basic_lucy_edit.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_lucy_edit.py) |
-| wan | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | T2V | [basic_wan2_2.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_wan2_2.py) |
-| wan | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | I2V | [basic_wan2_2_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_wan2_2_i2v.py) |
-| wan | `wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers` | T2V | [basic_self_forcing_causal.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal.py) |
-| wan | `rand0nmr/SFWan2.2-T2V-A14B-Diffusers` | T2V | [basic_self_forcing_causal_wan2_2_t2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_t2v.py) |
-| wan | `FastVideo/SFWan2.2-I2V-A14B-Preview-Diffusers` | I2V | [basic_self_forcing_causal_wan2_2_i2v.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_self_forcing_causal_wan2_2_i2v.py) |
-| zimage | `Tongyi-MAI/Z-Image-Turbo` | T2I | [basic_zimage.py](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_zimage.py) |
-
-**Note (stable_audio)**: the Stable Audio Open pipelines generate audio
-(`StableAudioT2AConfig` / `StableAudioOpenSmallConfig`); they are registered
-under the generic T2V workload option in the registry.
-
-**Note (Wan-VACE)**: not currently supported — no VACE pipeline or registered
-model ID exists on `main`
-([#1435](https://github.com/hao-ai-lab/FastVideo/issues/1435)). The closest
-supported path is the Wan2.1-Fun control pipeline
-(`IRMChen/Wan2.1-Fun-1.3B-Control-Diffusers`).
-
 The symbols used have the following meanings:

 - ✅ = Full compatibility
@@ -111,9 +25,6 @@ The `HuggingFace Model ID` can be passed directly to
 `from_pretrained()`. FastVideo then uses model-specific default settings for
 pipeline initialization and sampling.

-Registered models absent from this table have not been validated against these
-optimizations: absence means **untested**, not incompatible.
-
 <style>
  /* Target tables in this section */
  #models-x-optimization + p + table {
@@ -145,7 +56,7 @@ optimizations: absence means **untested**, not incompatible.
 | Model Name | HuggingFace Model ID | Resolutions | TeaCache | Sliding Tile Attn (Legacy Branch) | Sage Attn | VSA | BSA |
 |------------|---------------------|-------------|----------|-------------------|-----------|-----|-----|
 | FastWan2.1 T2V 1.3B | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
-| FastWan2.2 TI2V 5B Full Attn | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
+| FastWan2.2 TI2V 5B Full Attn* | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
 | Wan2.2 TI2V 5B | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 720P | ⭕ | ⭕ | ✅ | ⭕ | ⭕ |
 | DreamX-World 5B Cam | `FastVideo/DreamX-World-5B-Cam-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | DreamX-World 5B AR | `FastVideo/DreamX-World-5B-Diffusers` | 704px1280p | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
@@ -154,14 +65,14 @@ optimizations: absence means **untested**, not incompatible.
 | Wan2.2 I2V A14B | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |
 | HunyuanVideo | `hunyuanvideo-community/HunyuanVideo` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ | ⭕ |
 | FastHunyuan | `FastVideo/FastHunyuan-diffusers` | 720px1280p<br>544px960p | ❌ | ✅ | ✅ | ⭕ | ⭕ |
-| Wan2.1 T2V 1.3B | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | 480P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
-| Wan2.1 T2V 14B | `Wan-AI/Wan2.1-T2V-14B-Diffusers` | 480P, 720P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
-| Wan2.1 I2V 480P | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | 480P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
+| Wan2.1 T2V 1.3B | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` | 480P | ✅ | ✅* | ✅ | ⭕ | ⭕ |
+| Wan2.1 T2V 14B | `Wan-AI/Wan2.1-T2V-14B-Diffusers` | 480P, 720P | ✅ | ✅* | ✅ | ⭕ | ⭕ |
+| Wan2.1 I2V 480P | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers` | 480P | ✅ | ✅* | ✅ | ⭕ | ⭕ |
 | Wan2.1 I2V 720P | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers` | 720P | ✅ | ✅ | ✅ | ⭕ | ⭕ |
 | TurboWan2.1 T2V 1.3B | `loayrashid/TurboWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | TurboWan2.1 T2V 14B | `loayrashid/TurboWan2.1-T2V-14B-Diffusers` | 480P, 720P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | TurboWan2.2 I2V A14B | `loayrashid/TurboWan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
-| LongCat T2V 13.6B | `FastVideo/LongCat-Video-T2V-Diffusers` | 480P<br>720P | ❌ | ❌ | ❌ | ⭕ | ✅ |
+| LongCat T2V 13.6B | See note** | 480P<br>720P | ❌ | ❌ | ❌ | ⭕ | ✅ |
 | Matrix Game 2.0 Base Distilled | `FastVideo/Matrix-Game-2.0-Base-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | Matrix Game 2.0 GTA Distilled | `FastVideo/Matrix-Game-2.0-GTA-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | Matrix Game 2.0 TempleRun Distilled | `FastVideo/Matrix-Game-2.0-TempleRun-Distilled-Diffusers` | 352x640 | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
@@ -187,21 +98,6 @@ resolve default pipeline and sampling configuration for it.
 `FastVideo/GEN3C-Cosmos-7B-Diffusers`) or convert locally with
 `scripts/checkpoint_conversion/convert_gen3c_to_fastvideo.py`.

-## Hardware and OS
-
-Per the installation guides:
-
- **NVIDIA GPU (x86_64)** — CUDA 12.6 or 13.0; see the
-  [GPU install guide](../getting_started/installation/gpu.md).
- **NVIDIA DGX Spark (GB10, aarch64)** — CUDA 13, from-source kernel build; see
-  the [DGX Spark install guide](../getting_started/installation/spark.md).
- **Apple silicon (MPS)** — macOS 14 or newer; see the
-  [MPS install guide](../getting_started/installation/mps.md) and
-  [`basic_mps.py`](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_mps.py).
-
-Optimization-specific hardware constraints (e.g. STA requiring Hopper) are
-listed under [Special requirements](#special-requirements).
-
 ## Special requirements

 ### Sliding Tile Attention
@@ -1,147 +0,0 @@
-# Attn-QAT Training
-
-Attn-QAT simulates low-bit attention during training while keeping the rest of
-the training method unchanged. In the modular `fastvideo/train` framework it is
-a per-role model option, not a separate training method: supervised fine-tuning
-and DMD2 still own their losses and optimizer cadence.
-
-This guide covers the QAD Wan2.1-T2V-1.3B MixKit workflow:
-
-1. run a 4,000-step supervised Attn-QAT fine-tune;
-2. export the stage-1 DCP checkpoint to Diffusers format; and
-3. distill the student to three denoising steps with DMD2.
-
-The ready-to-run configs and wrappers are in
-`examples/train/scenario/qad_wan2_1_mixkit/`.
-
-## Role-local attention backends
-
-A DMD2 run owns three independent model roles. Configure the attention backend
-on each role so fake quantization is applied only to the student:
-
-```yaml
-models:
-  student:
-    attention_backend: ATTN_QAT_TRAIN
-  teacher:
-    attention_backend: FLASH_ATTN
-  critic:
-    attention_backend: FLASH_ATTN
-```
-
-The override is active only while that role's transformer is constructed, then
-the previous process-wide backend is restored. This lets student, teacher, and
-critic use different implementations in one process. Invalid role-level names
-fail during configuration instead of silently selecting another backend.
-
-See [Training Infrastructure](train_infra.md) for the complete model-role
-configuration reference.
-
-## Prerequisites
-
- Install FastVideo and make the `fastvideo-kernel` Python package importable.
-  `ATTN_QAT_TRAIN` intentionally fails instead of falling back to dense
-  attention when its kernel cannot be loaded.
- Prepare the precomputed MixKit VAE latents and text embeddings.
- Run the commands below from the repository root. The supplied recipe expects
-  four GPUs by default; set `NUM_GPUS` to override it.
-
-Download the published preprocessed dataset:
-
-```bash
-bash examples/training/finetune/wan_t2v_1.3B/mixkit/download_mixkit_data.sh
-```
-
-## Stage 1: supervised Attn-QAT fine-tuning
-
-The stage-1 config uses `ATTN_QAT_TRAIN` on the student, sequence parallelism
-across four GPUs, FP32 master weights, and 4,000 optimizer steps:
-
-```bash
-NUM_GPUS=4 \
-  bash examples/train/scenario/qad_wan2_1_mixkit/run_stage1.sh
-```
-
-Pass a dataset directory as the first positional argument when it differs from
-the default:
-
-```bash
-NUM_GPUS=4 \
-  bash examples/train/scenario/qad_wan2_1_mixkit/run_stage1.sh \
-  /path/to/combined_parquet_dataset
-```
-
-The wrapper calls `examples/train/run.sh`; the YAML file remains the source of
-truth for optimizer, validation, checkpointing, and distributed settings.
-
-## Export the stage-1 checkpoint
-
-Modular training checkpoints use Distributed Checkpoint (DCP) format. Export
-the student before using it to initialize stage 2:
-
-```bash
-bash examples/train/scenario/qad_wan2_1_mixkit/export_stage1.sh \
-  checkpoints/wan_t2v_qat_finetune/checkpoint-4000 \
-  checkpoints/wan_t2v_qat_finetune/diffusers
-```
-
-Both arguments are optional; the command above shows their defaults.
-
-## Stage 2: three-step DMD2 distillation
-
-Stage 2 loads the exported student weights, keeps Attn-QAT on the student, and
-uses Flash Attention for the teacher and critic:
-
-```bash
-NUM_GPUS=4 \
-  bash examples/train/scenario/qad_wan2_1_mixkit/run_stage2.sh \
-  data/HD-Mixkit-Finetune-Wan/combined_parquet_dataset \
-  checkpoints/wan_t2v_qat_finetune/diffusers/transformer/model.safetensors
-```
-
-The migrated recipe preserves these behaviors:
-
-| Behavior | Modular configuration |
-|---|---|
-| Student fake-quantized attention | `models.student.attention_backend: ATTN_QAT_TRAIN` |
-| Teacher and critic full-precision attention | Role-local `FLASH_ATTN` |
-| Generator update every five critic steps | `method.generator_update_interval: 5` |
-| Three-step rollout | `method.dmd_denoising_steps: [1000, 757, 522]` |
-| Score timestep range | `method.min_timestep_ratio: 0.02`, `max_timestep_ratio: 0.98` |
-| Legacy guidance `cond + 2(cond - uncond)` | Standard CFG scale `3.0` |
-| Stage handoff | DCP checkpoint to Diffusers export to student override weights |
-
-The timestep ratios apply to randomly sampled teacher and critic score
-timesteps; `dmd_denoising_steps` separately controls the student rollout. See
-[DMD Distillation](../distillation/dmd.md) for general DMD concepts.
-
-## Architecture-specific Triton routing
-
-The training kernel is runtime-JIT-compiled Triton code and selects its route on
-every call. It supports different query and key/value sequence lengths for
-cross-attention; key and value must have the same sequence length.
-
-| Hardware/configuration | Route |
-|---|---|
-| SM100, validated non-causal BF16 QAT configuration with head dimension 128 | Large-tile forward and split 64x64 backward; optimized backward requires a 16-aligned KV length |
-| SM120, including RTX 5090 | Previous forward tiling with joined quantized/STE P@V operations and a shallower backward pipeline for long sequences |
-| Unsupported configurations | Previous Triton implementation |
-
-Warp specialization is disabled automatically on SM100 and SM120 because the
-Triton 3.7 NVWS compiler pass aborts for this kernel on Blackwell. No user
-setting is required.
-
-The available tuning and comparison controls are:
-
-| Environment variable | Default | Effect |
-|---|---|---|
-| `FASTVIDEO_ATTN_QAT_FWD_MODE` | `fast` | Selects `fast`, `balanced`, or `reference` forward tiling on the SM100 optimized route |
-| `FASTVIDEO_ATTN_QAT_FWD_EXACT_M` | `0` | Set to `1` to recompute reference-order softmax statistics and keep `dV` bitwise-compatible on the SM100 optimized route |
-| `FASTVIDEO_ATTN_QAT_SM100_OPTIMIZED` | `1` | Set to `0` to force the previous SM100 forward and backward for comparison |
-| `FASTVIDEO_ATTN_QAT_SM120_JOIN_QAT_PV` | `1` | Set to `0` to compare SM120 against the split P@V path |
-
-The first invocation JIT-compiles the selected configuration; later calls reuse
-the Triton cache. To measure the production shape, run
-`python benchmarks/benchmark_attn_qat_train.py` from `fastvideo-kernel/`.
-
-For import and backend-selection failures, see [Debugging](../utilities/debugging.md).
@@ -62,18 +62,6 @@ bash examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh
 - Gradient checkpointing: `--enable_gradient_checkpointing_type "full"`
 - Memory scaling: Increase `--sp_size` or reduce `--num_latent_t` to fit in memory

-## Attention Quantization-Aware Training
-
-Attn-QAT fine-tunes a model while simulating low-bit attention in the forward
-and backward passes. The modular trainer can select the backend per model role,
-so a later DMD2 stage can keep fake quantization on the student while the
-teacher and critic use Flash Attention.
-
-The ready-to-run Wan2.1 MixKit workflow includes supervised fine-tuning,
-checkpoint export, and three-step DMD2 distillation:
-
-**→ [Follow the Attn-QAT training guide](attn_qat.md)**
-
 ## LoRA Finetuning

 LoRA (Low-Rank Adaptation) trains lightweight adapters while keeping the base model frozen. This significantly reduces memory usage and training time.
@@ -178,7 +166,6 @@ Ready-to-run training scripts are available for multiple models:
 | Wan2.1 I2V 14B | I2V | `examples/training/finetune/wan_i2v_14B_480p/crush_smol/` |
 | Wan2.1-Fun 1.3B InP | I2V | `examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/` |
 | Wan2.1 VSA | T2V/I2V | `examples/training/finetune/Wan2.1-VSA/Wan-Syn-Data/` |
-| Wan2.1 T2V 1.3B Attn-QAT | QAT SFT + DMD2 | `examples/train/scenario/qad_wan2_1_mixkit/` |

 Each example includes:

@@ -43,9 +43,6 @@ Ready-to-run examples with preprocessing scripts, training launchers, and valida

 **→ [Browse all training examples](examples/examples_training_index.md)**

-For the complete two-stage Wan2.1 MixKit quantization-aware workflow, see
-**[Attn-QAT Training](attn_qat.md)**.
-
 Each example includes:

 - `download_dataset.sh` — download sample data
@@ -62,11 +59,9 @@ FastVideo supports several training approaches:
 | **Full finetune** | Adapt entire model to a new domain or style |
 | **LoRA finetune** | Lightweight adaptation with frozen base weights |
 | **VSA finetune** | Finetune with Variable Sparse Attention for efficiency |
-| **Attn-QAT** | Train with fake-quantized attention, optionally followed by DMD2 distillation |

 ## Next Steps

 1. **Get started**: Pick an example from the [training examples index](examples/examples_training_index.md)
 2. **Prepare data**: Follow [data preprocessing](data_preprocess.md) for your own dataset
-3. **Train with quantized attention**: Follow the [Attn-QAT two-stage recipe](attn_qat.md)
-4. **Run inference**: After training, see [inference examples](../inference/examples/examples_inference_index.md)
+3. **Run inference**: After training, see [inference examples](../inference/examples/examples_inference_index.md)
@@ -81,7 +81,6 @@ Common model parameters:
 | `disable_custom_init_weights` | `false` | Skip custom weight initialization (use for teacher/critic) |
 | `flow_shift` | `3.0` | Timestep shifting factor |
 | `enable_gradient_checkpointing_type` | `null` | Gradient checkpointing (`"full"` or `null`) |
-| `attention_backend` | `null` | Optional role-local backend for Wan models (for example `ATTN_QAT_TRAIN`); overrides the process default only while this role's transformer is built |

 Which roles are needed depends on the training method:

@@ -212,29 +211,6 @@ pipeline:
  flow_shift: 8
 ```

-Registered transformer linear-quantization configs can also be selected by
-name. For example, the LTX-2 NVFP4-QAT recipe applies real FP4 forward GEMMs
-with a straight-through-estimator backward to its deployment-targeted
-attention/FFN projections:
-
-```yaml
-pipeline:
-  dit_config:
-    quant_config: nvfp4_qat_train
-```
-
-The LTX-2 recipe in
-`examples/train/configs/overfit_ltx2_t2v_nvfp4_qat.yaml` combines that linear
-configuration with `models.student.attention_backend: ATTN_QAT_TRAIN` for
-video-attention forward/backward. On sm120, its validation callback temporarily
-switches those layers to `ATTN_QAT_INFER`.
-On GB200, set `callbacks.validation.attn_qat_infer: false` to keep validation on
-the train-time QAT backend; the inference kernel is sm120-only.
-
-User-adaptable LTX-2 fine-tuning recipes (full, LoRA, and NVFP4 QAT) live in
-`examples/train/configs/fine_tuning/ltx2/`, alongside the other model
-families under `examples/train/configs/fine_tuning/`.
-
 ---

 ## Training Methods
@@ -322,8 +298,6 @@ method:
 | `dmd_denoising_steps` | *(required)* | Timestep schedule for student rollout |
 | `generator_update_interval` | `1` | Update student every N critic steps |
 | `real_score_guidance_scale` | `1.0` | CFG scale for teacher predictions |
-| `min_timestep_ratio` | `0.0` | Lower bound for randomly sampled teacher/critic score timesteps |
-| `max_timestep_ratio` | `1.0` | Upper bound for randomly sampled teacher/critic score timesteps |
 | `fake_score_learning_rate` | *(required)* | Critic optimizer learning rate |
 | `fake_score_betas` | *(required)* | Critic optimizer Adam betas |
 | `fake_score_lr_scheduler` | *(required)* | Critic LR scheduler type |
@@ -77,13 +77,8 @@ If forcing a backend fails, verify optional dependencies are installed:
 - `SAGE_ATTN`: SageAttention package
 - `SAGE_ATTN_THREE`: upstream `sageattn3` package
 - `ATTN_QAT_INFER`: `fastvideo-kernel` checkout/source install that exposes
-  `attn_qat_infer`, AND a consumer-Blackwell (sm_120/sm_121) GPU -- on any
-  other device the backend reports unavailable (even if a CUDA 13 wheel
-  bundles the extension) and selection falls back to FlashAttention
- `ATTN_QAT_TRAIN`: `fastvideo-kernel`; its runtime-JIT Triton implementation
-  selects an optimized route on SM100, joins the quantized and STE P@V paths on
-  SM120, and retains the previous route for unsupported configurations. See
-  [Attn-QAT Training](../training/attn_qat.md) for architecture controls.
+  `attn_qat_infer`
+- `ATTN_QAT_TRAIN`: `fastvideo-kernel` install exposing `fastvideo_kernel`

 As a fallback, use:

@@ -1,17 +0,0 @@
-# LingBot World 2 Example Dataset
-
-These files were copied unchanged from the LingBot World 2 repository for the
-FastVideo causal-fast inference example.
-
- Repository: `https://github.com/Robbyant/lingbot-world-v2.git`
- Source commit: `94f43115de8d4a4f9f282126528c300a0b232c5f`
- Source directory: `examples/03`
-
-## Files
-
- `image.jpg`: source image for image-to-video generation. SHA-256:
-  `6ee3dacfef32cfef504dd698adb8a660cf15f686535c52fed4903fef27c0edd0`
- `poses.npy`: camera-to-world trajectory matrices. SHA-256:
-  `bd0a23a696e184b0b43e7767eb432bfe644690560fe327fa96961affc941c404`
- `intrinsics.npy`: camera intrinsic parameters. SHA-256:
-  `821fca6cf957ae8fbb1181307f02479efb1705e04c9e05734cd02fb43462e082`
@@ -1,52 +0,0 @@
-"""Generate a five-second Dense LingBot-Video clip with the official defaults."""
-
-import argparse
-from pathlib import Path
-
-from fastvideo import VideoGenerator
-
-
-def parse_args() -> argparse.Namespace:
-    """Parse the converted checkpoint and output paths for the sample."""
-    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument(
-        "--model-path",
-        type=Path,
-        required=True,
-        help="Path to a converted Dense LingBot-Video checkpoint.",
-    )
-    parser.add_argument(
-        "--output-path",
-        type=Path,
-        default=Path("outputs/lingbot-video/dense-t2v"),
-        help="Directory for the generated video.",
-    )
-    return parser.parse_args()
-
-
-def main() -> None:
-    """Load the converted Dense checkpoint and generate the default T2V sample."""
-    args = parse_args()
-    generator = VideoGenerator.from_pretrained(
-        str(args.model_path),
-        num_gpus=1,
-        use_fsdp_inference=False,
-        text_encoder_cpu_offload=True,
-        vae_cpu_offload=False,
-        pin_cpu_memory=True,
-    )
-    try:
-        generator.generate({
-            "prompt": "A red fox runs through fresh snow at sunrise.",
-            "output": {
-                "output_path": str(args.output_path),
-                "save_video": True,
-                "return_frames": False,
-            },
-        })
-    finally:
-        generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()
@@ -1,52 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Run LingBot World 2 14B causal-fast I2V generation with FastVideo."""
-
-import os
-from pathlib import Path
-
-from fastvideo import VideoGenerator
-
-
-REPO_ROOT = Path(__file__).resolve().parents[3]
-DATASET_DIR = REPO_ROOT / "examples" / "dataset" / "lingbotworld2"
-OUTPUT_PATH = REPO_ROOT / "outputs" / "lingbotworld2_causal_fast.mp4"
-
-
-def main() -> None:
-    """Load the native FastVideo LingBot World 2 causal-fast pipeline and generate one video."""
-    generator = VideoGenerator.from_pretrained(
-        os.environ["LINGBOTWORLD2_MODEL_PATH"],
-        num_gpus=8,
-        sp_size=8,
-        hsdp_shard_dim=8,
-        use_fsdp_inference=True,
-        dit_layerwise_offload=False,
-        dit_cpu_offload=False,
-        vae_cpu_offload=False,
-        text_encoder_cpu_offload=False,
-        pin_cpu_memory=True,
-        override_pipeline_cls_name="LingBotWorld2CausalFastPipeline",
-    )
-
-    try:
-        generator.generate_video(
-            "A serene lakeside scene with a lone tree standing in calm water, surrounded by distant snow-capped mountains under a bright blue sky with drifting white clouds; gentle ripples reflect the tree and sky, creating a tranquil, meditative atmosphere.",
-            image_path=str(DATASET_DIR / "image.jpg"),
-            action_path=str(DATASET_DIR),
-            output_path=str(OUTPUT_PATH),
-            save_video=True,
-            height=480,
-            width=832,
-            num_frames=65,
-            num_inference_steps=4,
-            guidance_scale=1.0,
-            negative_prompt="",
-            fps=16,
-            seed=42,
-        )
-    finally:
-        generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()
@@ -191,22 +191,6 @@ def main() -> None:
    print(f"Using refine upsampler: {refine_upsampler_path}")

    pipeline_config = PipelineConfig.from_pretrained(model_root)
-    # LTX-2 NVFP4 deploy contract (train==deploy surface):
-    #   * Linears: NVFP4 block-scaled GEMMs (per-16 E2M1 + E4M3 SFs) on every
-    #     arch, via flashinfer.
-    #   * ATTN_QAT_INFER attention differs per arch: sm_120a/sm_121a use the
-    #     fastvideo-kernel CUTLASS (SageAttention3-FP4) scheme that
-    #     ATTN_QAT_TRAIN simulates; sm_100a (GB200) / sm_103a (GB300) use the
-    #     FP4 FA4 kernel (flash-attention-fp4) with per-16 block-scaled NVFP4
-    #     Q/K and BF16 P/V -- a train-sim mismatch that is gated by MS-SSIM
-    #     measurement, not assumed equal. The selection receipt is logged at
-    #     backend resolution ("ATTN_QAT_INFER resolved: ...").
-    # Original-weight retention: the default purges the always-FP4 layers'
-    # bf16 originals after conversion. Refine-only layers (the cross-modal
-    # AV projections) always keep theirs: the base stage profile runs them
-    # dense by deployment contract -- in the two-stage fast profile AND the
-    # distilled single-stage deploy. retain_original_weights=True keeps
-    # everything (debugging).
    pipeline_config.dit_config.quant_config = NVFP4Config()
    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
    torch_compile_kwargs = {
@@ -1,96 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Run Z-Image-Turbo text-to-image generation through FastVideo.
-
-User story:
-    "I want the official Z-Image-Turbo defaults and a deterministic PNG from
-    a local or Hugging Face checkpoint."
-"""
-
-import argparse
-from pathlib import Path
-
-from fastvideo import VideoGenerator
-from fastvideo.api import (
-    EngineConfig,
-    GenerationRequest,
-    GeneratorConfig,
-    OutputConfig,
-    ParallelismConfig,
-    PipelineSelection,
-    SamplingConfig,
-)
-
-
-DEFAULT_PROMPT = (
-    "Young Chinese woman in red Hanfu, intricate embroidery. Impeccable makeup, red floral forehead pattern. "
-    "Elaborate high bun, golden phoenix headdress, red flowers, beads. Holds round folding fan with lady, trees, bird. "
-    "Neon lightning-bolt lamp (⚡️), bright yellow glow, above extended left palm. Soft-lit outdoor night background, "
-    "silhouetted tiered pagoda (西安大雁塔), blurred colorful distant lights."
-)
-DEFAULT_REVISION = "f332072aa78be7aecdf3ee76d5c247082da564a6"
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Run Z-Image-Turbo text-to-image generation.")
-    parser.add_argument("--model-path", default="Tongyi-MAI/Z-Image-Turbo")
-    parser.add_argument("--revision", default=DEFAULT_REVISION)
-    parser.add_argument("--output", default="outputs/zimage/zimage_turbo.png")
-    parser.add_argument("--prompt", default=DEFAULT_PROMPT)
-    parser.add_argument("--negative-prompt", default="")
-    parser.add_argument("--height", type=int, default=1024)
-    parser.add_argument("--width", type=int, default=1024)
-    parser.add_argument("--steps", type=int, default=8)
-    parser.add_argument("--guidance-scale", type=float, default=0.0)
-    parser.add_argument("--max-sequence-length", type=int, default=512)
-    parser.add_argument("--cfg-normalization", action=argparse.BooleanOptionalAction, default=False)
-    parser.add_argument("--cfg-truncation", type=float, default=1.0)
-    parser.add_argument("--seed", type=int, default=42)
-    return parser.parse_args()
-
-
-def main() -> None:
-    args = parse_args()
-    output = Path(args.output)
-    output.parent.mkdir(parents=True, exist_ok=True)
-
-    generator = VideoGenerator.from_config(
-        GeneratorConfig(
-            model_path=args.model_path,
-            revision=args.revision,
-            engine=EngineConfig(
-                num_gpus=1,
-                parallelism=ParallelismConfig(tp_size=1, sp_size=1),
-                use_fsdp_inference=False,
-            ),
-            # The model registry selects the native zimage_turbo preset.
-            pipeline=PipelineSelection(workload_type="t2i"),
-        ))
-    try:
-        generator.generate(
-            GenerationRequest(
-                prompt=args.prompt,
-                negative_prompt=args.negative_prompt,
-                sampling=SamplingConfig(
-                    height=args.height,
-                    width=args.width,
-                    num_frames=1,
-                    fps=1,
-                    num_inference_steps=args.steps,
-                    guidance_scale=args.guidance_scale,
-                    max_sequence_length=args.max_sequence_length,
-                    cfg_normalization=args.cfg_normalization,
-                    cfg_truncation=args.cfg_truncation,
-                    seed=args.seed,
-                ),
-                output=OutputConfig(
-                    output_path=str(output),
-                    save_video=True,
-                    return_frames=False,
-                ),
-            ))
-    finally:
-        generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()
@@ -6,6 +6,10 @@ original single-file version to make each concern independently reviewable.
 The folder name matches the sibling `gradio_local_demo*.py` demos in this
 directory so both flat and packaged demos read consistently.

+> **Status: draft.** This package is structurally in place but will not run
+> against the current upstream `fastvideo` package. See *Blocking prereqs*
+> below.
+
 ## Layout

 | File | Purpose |
@@ -34,21 +38,27 @@ python -m gradio_local_demo_ltx2_3 --port 7860
 GPU requirement: a single FP4-capable GPU (B200 or comparable) for the
 "real-time 1080p" speed claim. Lower tiers will still run but slower.

-## Upstream surfaces the demo relies on
+## Blocking prereqs (why this draft PR cannot be merged yet)

-Everything the demo needs ships in the upstream `fastvideo` package:
+The upstream `fastvideo` package is missing three pieces that the demo
+currently depends on verbatim. Each needs its own upstreaming PR before this
+demo can actually boot:

- **`fastvideo.layers.quantization.nvfp4_config.NVFP4Config`** — the demo sets
-  `pipeline_config.dit_config.quant_config = NVFP4Config()` in `app.py` (same
-  pattern as `examples/inference/basic/basic_ltx2_distilled_fast_profile.py`).
- **LTX-2.3 refine kwargs on `VideoGenerator`** — `ltx2_refine_enabled`,
-  `ltx2_refine_upsampler_path`, `ltx2_refine_lora_path`,
-  `ltx2_refine_num_inference_steps`, `ltx2_refine_guidance_scale`,
-  `ltx2_refine_add_noise`, and `ltx2_vae_tiling` are wired in
-  `fastvideo/fastvideo_args.py`, backed by
-  `fastvideo/pipelines/basic/ltx2/stages/ltx2_refine.py`.
- **`SamplingParam`** — imported from the public `fastvideo` package root
-  (`fastvideo.api.sampling_param` under the hood).
+1. **`fastvideo.layers.quantization.fp4_config.FP4Config`** — the demo sets
+   `pipeline_config.dit_config.quant_config = FP4Config()` in `app.py`.
+   Upstream only ships `absmax_fp8.py` and `base_config.py` under
+   `fastvideo/layers/quantization/`.
+2. **LTX-2.3 refine / image-conditioning kwargs on `VideoGenerator`** —
+   `ltx2_refine_enabled`, `ltx2_refine_upsampler_path`, `ltx2_refine_lora_path`,
+   `ltx2_refine_num_inference_steps`, `ltx2_refine_guidance_scale`,
+   `ltx2_refine_add_noise`, `ltx2_images`, `ltx2_image_crf`. Upstream
+   `fastvideo/fastvideo_args.py` currently wires only `ltx2_vae_tiling`.
+   The backing stages (`ltx2_refine.py`, `ltx2_i2v_conditioning.py`) are
+   also missing from `fastvideo/pipelines/stages/`.
+3. **`fastvideo.configs.sample.base.SamplingParam`** — the import path used
+   by this demo. Upstream moved sampling params to
+   `fastvideo.api.sampling_param`. A re-export shim at the old path, or an
+   import update here once the other two prereqs land, will resolve it.

 ## Environment variables

@@ -4,10 +4,10 @@ from pathlib import Path

 import gradio as gr

-from fastvideo import SamplingParam
 from fastvideo.configs.pipelines.base import PipelineConfig
+from fastvideo.configs.sample.base import SamplingParam
 from fastvideo.entrypoints.video_generator import VideoGenerator
-from fastvideo.layers.quantization.nvfp4_config import NVFP4Config
+from fastvideo.layers.quantization.fp4_config import FP4Config
 from fastvideo.utils import maybe_download_model

 from .config import (
@@ -44,7 +44,7 @@ def main():
        resolved_model_path = Path(model_root)

        pipeline_config = PipelineConfig.from_pretrained(str(resolved_model_path))
-        pipeline_config.dit_config.quant_config = NVFP4Config()
+        pipeline_config.dit_config.quant_config = FP4Config()
        refine_upsampler_path = resolve_refine_upsampler_path(resolved_model_path)
        print(f"Using refine upsampler: {refine_upsampler_path}")

@@ -4,7 +4,7 @@ from pathlib import Path
 import torch
 import torch._inductor.config

-from fastvideo import SamplingParam
+from fastvideo.configs.sample.base import SamplingParam

 LOCAL_DEMO_DIR = Path(__file__).resolve().parent
 CLASSIFIER_DIR = Path(
@@ -5,7 +5,7 @@ from copy import deepcopy

 import gradio as gr

-from fastvideo import SamplingParam
+from fastvideo.configs.sample.base import SamplingParam
 from fastvideo.entrypoints.video_generator import VideoGenerator

 from .config import (
@@ -1,176 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""LTX-2.3 distilled text-to-video with the optimized NVFP4 inference stack.
-
-Runs `FastVideo/LTX-2.3-Distilled-Diffusers` on a single GPU with the full
-validated optimization stack:
-
-* NVFP4 block-scaled linear layers (per-16 E2M1 weights + E4M3 scale factors),
-* ATTN_QAT_INFER FP4 attention (arch-resolved kernel, receipt logged),
-* torch.compile (fullgraph) over the DiT, text encoder, and VAE,
-* single-stage 8-step distilled sampling at guidance 1.0.
-
-Quick start
-----------
-    # On GB200-class ARM hosts, unset LD_LIBRARY_PATH (see Hardware notes):
-    env -u LD_LIBRARY_PATH python examples/inference/ltx2_3/optimized_nvfp4_t2v.py
-
-    # Optional overrides:
-    #   export LTX23_MODEL_PATH=/path/to/local/snapshot
-    #   export LTX23_T2V_PROMPT="a red fox running through fresh snow"
-    #   export LTX23_OUTPUT_DIR=outputs_video/ltx2_3_nvfp4_t2v
-
-Hardware notes
--------------
- On GB200 / Blackwell, run with `env -u LD_LIBRARY_PATH ...` to avoid a
-  system-cuBLAS / torch-cuBLAS mismatch that fails every GEMM (some ARM
-  container images ship an LD_LIBRARY_PATH that breaks torch.compile's
-  toolchain discovery). The `_inductor.shape_padding = False` line below
-  also avoids a pad_mm landmine on the same generation of cards.
-"""
-from __future__ import annotations
-
-import os
-import time
-from pathlib import Path
-
-import torch._inductor.config as _inductor
-
-from fastvideo import VideoGenerator
-from fastvideo.configs.pipelines.base import PipelineConfig
-from fastvideo.layers.quantization.nvfp4_config import NVFP4Config
-from fastvideo.utils import maybe_download_model
-
-# ATTN_QAT_INFER is the FP4 attention half of the NVFP4 deploy contract. It
-# resolves per arch (CUTLASS SageAttention3-FP4 on sm_120a/sm_121a, FP4 FA4
-# on sm_100a/sm_103a) and logs a one-line receipt of what actually bound.
-os.environ.setdefault("FASTVIDEO_ATTENTION_BACKEND", "ATTN_QAT_INFER")
-os.environ.setdefault("FASTVIDEO_STAGE_LOGGING", "1")
-
-# Inductor knobs. The first one (shape_padding=False) is mandatory on
-# Blackwell to avoid a cuBLAS INVALID_VALUE crash inside pad_mm. The rest
-# are the same matmul-friendliness flags the sibling LTX-2 examples use.
-_inductor.shape_padding = False
-_inductor.conv_1x1_as_mm = True  # treat 1x1 convolutions as matrix muls
-_inductor.coordinate_descent_tuning = True
-_inductor.coordinate_descent_check_all_directions = True
-_inductor.epilogue_fusion = False  # do not fuse pointwise ops into matmuls
-
-MODEL_ID = os.path.expandvars(
-    os.path.expanduser(
-        os.getenv("LTX23_MODEL_PATH", "FastVideo/LTX-2.3-Distilled-Diffusers")
-    )
-)
-OUTPUT_DIR = Path(os.getenv("LTX23_OUTPUT_DIR", "outputs_video/ltx2_3_nvfp4_t2v"))
-DEFAULT_PROMPT = (
-    "A fashion model takes a slow step forward and shifts her weight, "
-    "the soft fabric of her clothing swaying and rippling with the "
-    "motion, her hair shifting gently, soft even studio lighting on a "
-    "clean light background, elegant slow-motion runway feel."
-)
-PROMPT = os.getenv("LTX23_T2V_PROMPT", DEFAULT_PROMPT)
-
-
-def main() -> None:
-    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
-    model_root = maybe_download_model(MODEL_ID)
-    print(f"Model:      {model_root}")
-    print(f"Output dir: {OUTPUT_DIR.resolve()}")
-
-    # Loading the pipeline config *with model_path* binds model-specific
-    # tuning (notably VAE precision/decoder defaults) into the config.
-    pipeline_config = PipelineConfig.from_pretrained(model_root)
-
-    # NVFP4 linear layers on the DiT. The default purges the original BF16
-    # weights of the always-FP4 linears right after conversion — a large
-    # peak-memory reduction. Refine-only layers (the cross-modal AV
-    # projections) always keep theirs: the base stage profile runs them
-    # dense by deployment contract. retain_original_weights=True keeps
-    # everything (debugging).
-    pipeline_config.dit_config.quant_config = NVFP4Config()
-
-    # fullgraph=True is supported on the NVFP4 path: the FP4 quantize step
-    # is a registered custom op (fastvideo::nvfp4_quantize_fa4), so dynamo
-    # traces through it without graph breaks. mode="default" — the two
-    # CUDAGraph modes ("reduce-overhead" / "max-autotune") are a separate
-    # opt-in, not part of this validated preset.
-    torch_compile_kwargs = {
-        "backend": "inductor",
-        "fullgraph": True,
-        "mode": "default",
-        "dynamic": False,
-    }
-
-    generator = VideoGenerator.from_pretrained(
-        model_root,
-        num_gpus=1,
-        pipeline_config=pipeline_config,
-        # Compile the DiT, text encoder, and VAE — all three stages benefit,
-        # and the VAE's codec submodules compile cleanly under fullgraph.
-        enable_torch_compile=True,
-        enable_torch_compile_text_encoder=True,
-        enable_torch_compile_vae=True,
-        torch_compile_kwargs=torch_compile_kwargs,
-        torch_compile_kwargs_vae=torch_compile_kwargs,
-        # Keep everything resident — no CPU offload for serving-style runs.
-        dit_cpu_offload=False,
-        text_encoder_cpu_offload=False,
-        vae_cpu_offload=False,
-        ltx2_vae_tiling=False,
-    )
-
-    common_kwargs = dict(
-        prompt=PROMPT,
-        negative_prompt="",      # distilled is CFG-free; no negative needed
-        guidance_scale=1.0,      # CFG=1 for distilled
-        height=1280, width=832,  # portrait runway aspect
-        num_frames=121, fps=24,  # ~5s clip
-        # Single-stage 8-step distilled sampling — the validated preset for
-        # this checkpoint (no two-stage refine; the NVFP4 deploy contract
-        # runs the distilled single-stage recipe).
-        num_inference_steps=8,
-        save_video=True,
-    )
-
-    try:
-        # Warmup: pays cold compile + first-shape guard work, untimed.
-        print("\n[warmup] compiling + generating…")
-        generator.generate_video(
-            output_path=str(OUTPUT_DIR / "_warmup.mp4"),
-            seed=7,
-            **common_kwargs,
-        )
-        (OUTPUT_DIR / "_warmup.mp4").unlink(missing_ok=True)
-
-        # Measured run.
-        out_path = OUTPUT_DIR / "output_ltx2_3_nvfp4_t2v.mp4"
-        print(f"\n[measured] generating: {out_path}")
-        t0 = time.perf_counter()
-        result = generator.generate_video(
-            output_path=str(out_path),
-            seed=2002,
-            **common_kwargs,
-        )
-        wall = time.perf_counter() - t0
-        e2e = (result.get("e2e_latency") if isinstance(result, dict) else None) or wall
-        print(f"[measured] e2e={e2e:.2f}s wall={wall:.2f}s")
-    finally:
-        generator.shutdown()
-
-    # Expected receipts — verify these two lines in your own run's log:
-    #
-    # 1. ATTN_QAT_INFER routing receipt (logged at backend resolution; on a
-    #    GB200-class part it reads):
-    #
-    #      ATTN_QAT_INFER resolved: arch=sm_100 kernel=flash-attention-fp4 \
-    #        qk_mode=nvfp4(per-16-e4m3-sf) pv_mode=bf16 train_sim_mismatch=measured
-    #
-    # 2. NVFP4 weight purge receipt (logged after model conversion; N/M/X
-    #    depend on the checkpoint):
-    #
-    #      NVFP4 weight purge receipt: purged N original bf16 weight tensors \
-    #        (X.XX GiB freed); retained M (refine-only dense fallback or \
-    #        retain_original_weights).
-
-
-if __name__ == "__main__":
-    main()
@@ -91,7 +91,3 @@ examples/train/
 ```

 See `configs/README.md` and `scenario/README.md` for details.
-
-The featured QAD Wan2.1 MixKit scenario runs Attn-QAT supervised fine-tuning,
-checkpoint export, and three-step DMD2. See the
-[Attn-QAT training guide](https://haoailab.com/FastVideo/training/attn_qat/).
@@ -1,70 +0,0 @@
-#!/bin/bash
-# Stage-2 QAT-aware DMD2 distillation for Kandinsky5 T2V (480p).
-#
-# Usage:
-#   bash distill_dmd_qat.sh <stage1-checkpoint> [--dotted.key value ...]
-#
-# <stage1-checkpoint> is either:
-#   - a raw stage-1 DCP checkpoint written by
-#     examples/train/configs/fine_tuning/kandinsky5/finetune_qat.sh
-#     (a checkpoint-<N>/ dir with a dcp/ subdir, or the stage-1 output_dir
-#     itself -- dcp_to_diffusers auto-picks the latest checkpoint). It is
-#     converted to a diffusers export (<checkpoint>-diffusers) with
-#     `dcp_to_diffusers --role student --verify` before launching --
-#     models.*.init_from need a diffusers model directory (model_index.json
-#     + component subfolders), not the raw DCP layout; or
-#   - an already-exported diffusers model dir (model_index.json present),
-#     used as-is with no conversion.
-#
-# The export is passed to all three models.*.init_from overrides: student,
-# teacher, and critic all load the same weights. Teacher/critic are
-# automatically masked back to dense attention and full-precision weights
-# by the _loading_teacher_critic_model gate in
-# fastvideo/models/loader/component_loader.py (family-agnostic, no
-# Kandinsky5-specific handling needed) -- NOT by pointing them at a
-# different export.
-#
-# FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN keeps the student's dense/local
-# attention fake-quantized during distillation.
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "${SCRIPT_DIR}/../../../../.." && pwd)"
-
-CHECKPOINT="${1:?Usage: $0 <stage1-checkpoint (DCP dir or diffusers export)> [extra --dotted.key overrides...]}"
-shift
-
-if [[ "${CHECKPOINT}" != /* ]]; then
-    CHECKPOINT="$(pwd)/${CHECKPOINT}"
-fi
-CHECKPOINT="${CHECKPOINT%/}"
-
-if [[ -f "${CHECKPOINT}/model_index.json" ]]; then
-    # Already a diffusers export -- use as-is.
-    INIT_FROM="${CHECKPOINT}"
-else
-    # Convert the raw DCP checkpoint. Runs BEFORE ATTN_QAT_TRAIN is
-    # exported below: backend selection for that env var refuses to start
-    # (ImportError) when the training kernel isn't built, and the
-    # conversion itself must not depend on it. --overwrite keeps relaunches
-    # from silently reusing a stale export for a newer checkpoint;
-    # --verify strictly reloads the exported transformer so a key-mapping
-    # bug fails here, not deep inside the training launch below.
-    INIT_FROM="${CHECKPOINT}-diffusers"
-    echo "Converting stage-1 DCP checkpoint ${CHECKPOINT} -> ${INIT_FROM}"
-    python -m fastvideo.train.entrypoint.dcp_to_diffusers \
-        --checkpoint "${CHECKPOINT}" \
-        --output-dir "${INIT_FROM}" \
-        --role student \
-        --overwrite \
-        --verify
-fi
-
-export FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN
-
-bash "${REPO_ROOT}/examples/train/run.sh" \
-    "${SCRIPT_DIR}/dmd2_t2v_480p_qat.yaml" \
-    --models.student.init_from "${INIT_FROM}" \
-    --models.teacher.init_from "${INIT_FROM}" \
-    --models.critic.init_from "${INIT_FROM}" \
-    "$@"
@@ -1,113 +0,0 @@
-# DMD2 QAT-aware distillation: Kandinsky5 T2V Lite, 480p (teacher many-step
-# -> student 4-step), continuing from a stage-1 Attn-QAT finetune checkpoint.
-#
-# init_from below must be a diffusers model directory (model_index.json +
-# component subfolders), not the raw checkpoint-N/ stage 1 writes (dcp/ +
-# metadata/RNG state only). Launching via distill_dmd_qat.sh handles this:
-# it runs fastvideo.train.entrypoint.dcp_to_diffusers (--verify) on the
-# stage-1 checkpoint you pass it and overrides all three init_from values
-# with the export -- the placeholder paths below only apply when launching
-# the YAML directly.
-#
-# - Teacher: frozen, full-precision Kandinsky5 (loaded from the stage-1 QAT
-#   checkpoint's base weights, or the original pretrained checkpoint --
-#   the _loading_teacher_critic_model gate in component_loader.py masks
-#   quant_config and FASTVIDEO_ATTENTION_BACKEND for teacher/critic
-#   regardless, so they always run full precision / dense attention).
-# - Student: trainable, initialized from the stage-1 QAT finetune checkpoint.
-# - Critic:  trainable, initialized from the same checkpoint as the student.
-# - Validation: 4-step sampling via Kandinsky5DMDPipeline.
-#
-# Launch with FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN set (student only;
-# teacher/critic are masked to dense automatically) so the student continues
-# learning to absorb fake-quantized attention error during distillation.
-#
-# Scope: T2V, 480p only, dense/local attention only.
-
-models:
-  student:
-    _target_: fastvideo.train.models.kandinsky5.Kandinsky5Model
-    init_from: outputs/kandinsky5_t2v_qat_finetune/checkpoint-1000-diffusers
-    trainable: true
-  teacher:
-    _target_: fastvideo.train.models.kandinsky5.Kandinsky5Model
-    init_from: outputs/kandinsky5_t2v_qat_finetune/checkpoint-1000-diffusers
-    trainable: false
-    disable_custom_init_weights: true
-  critic:
-    _target_: fastvideo.train.models.kandinsky5.Kandinsky5Model
-    init_from: outputs/kandinsky5_t2v_qat_finetune/checkpoint-1000-diffusers
-    trainable: true
-    disable_custom_init_weights: true
-
-method:
-  _target_: fastvideo.train.methods.distribution_matching.dmd2.DMD2Method
-  rollout_mode: simulate
-  generator_update_interval: 5
-  real_score_guidance_scale: 4.5
-  dmd_denoising_steps: [1000, 750, 500, 250]
-
-  # Critic optimizer (required -- no fallback to training.optimizer)
-  fake_score_learning_rate: 8.0e-6
-  fake_score_betas: [0.0, 0.999]
-  fake_score_lr_scheduler: constant
-
-training:
-  distributed:
-    num_gpus: 8
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 8
-
-  data:
-    data_path: data/kandinsky5_overfit_preprocessed
-    dataloader_num_workers: 4
-    train_batch_size: 1
-    training_cfg_rate: 0.0
-    seed: 1000
-    num_latent_t: 31
-    num_height: 512
-    num_width: 768
-    num_frames: 121
-
-  optimizer:
-    learning_rate: 2.0e-6
-    betas: [0.0, 0.999]
-    weight_decay: 0.01
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    max_train_steps: 4000
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/kandinsky5_t2v_dmd2_4steps_qat
-    training_state_checkpointing_steps: 20
-    checkpoints_total_limit: 3
-
-  tracker:
-    project_name: distillation_kandinsky5
-    run_name: kandinsky5_t2v_dmd2_4steps_qat
-
-  model:
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    max_grad_norm: 1.0
-  validation:
-    pipeline_target: fastvideo.pipelines.basic.kandinsky5.kandinsky5_dmd_pipeline.Kandinsky5DMDPipeline
-    dataset_file: data/kandinsky5_overfit_preprocessed/validation_prompts.json
-    every_steps: 50
-    sampling_steps: [4]
-    sampling_timesteps: [1000, 750, 500, 250]
-    guidance_scale: 6.0
-  ema:
-    _target_: fastvideo.train.callbacks.ema.EMACallback
-    decay: 0.98
-    start_iter: 0
-
-pipeline:
-  flow_shift: 5
@@ -25,7 +25,6 @@ models:
    disable_custom_init_weights: false                # default: false
    flow_shift: 3.0                                   # default: 3.0
    enable_gradient_checkpointing_type: null           # default: null (falls back to training.model)
-    attention_backend: null                            # default: null (global/default); role-local when set

  teacher:
    _target_: fastvideo.train.models.wan.WanModel
@@ -53,8 +52,6 @@ method:
  rollout_mode: simulate              # required: "simulate" or "data_latent"
  generator_update_interval: 5        # default: 1
  dmd_denoising_steps: [1000, 750, 500, 250]  # SDE timestep schedule
-  min_timestep_ratio: 0.0            # score-model timestep lower bound
-  max_timestep_ratio: 1.0            # score-model timestep upper bound

  # Critic optimizer (all required — no fallback)
  fake_score_learning_rate: 8.0e-6
@@ -195,10 +192,6 @@ callbacks:
 # ------------------------------------------------------------------------------
 pipeline:
  flow_shift: 3                       # default: null (model-specific)
-  # LTX-2 linear QAT (SM100+ with FlashInfer); distinct from the role-local
-  # attention_backend option above, which selects attention math for Wan.
-  # dit_config:
-  #   quant_config: nvfp4_qat_train
  # flow_shift_sr: null               # default: null (super-resolution shift)
  # embedded_cfg_scale: 6.0           # default: 6.0
  # is_causal: false                  # default: false
@@ -1,198 +0,0 @@
-# Kandinsky5 QAD recipe (T2V, 480p)
-
-Quantization-aware distillation for Kandinsky-5.0 Lite T2V at
-**512×768, 121 frames, 24 fps** (the checkpoint's native `KANDINSKY5_T2V_LITE_5S`
-preset -- 121/24 ~= 5.04s, matching the "5s" in the checkpoint name),
-mirroring the Wan2.1 QAD recipe
-(`examples/training/finetune/wan_t2v_1.3B/mixkit/`) on the new
-`fastvideo/train/` stack. Scope is deliberately narrow: T2V only, 480p only,
-dense/local attention only. Kandinsky5's NABLA sparse attention path is
-never engaged at this resolution and is unsupported by this recipe --
-attempting a larger resolution will raise loudly from `Kandinsky5Model`
-rather than silently mis-scale the visual RoPE.
-
-## Data
-
-Kandinsky5 uses two text encoders (Qwen/Reason1 + CLIP), unlike Wan's single
-encoder. The CLIP pooled projection is zero-padded into a row and prepended
-to the Qwen sequence embeddings, then stored as a single `[seq+1, dim]`
-tensor in the existing `text_embedding` Parquet field -- the same trick
-`preprocess_hunyuan_overfit.py` uses for LLaMA+CLIP. No new Parquet schema
-is needed.
-
-Build a small Parquet dataset from raw videos + captions
-(`videos2caption.json` + `videos/*.mp4`, matching the Hunyuan overfit script's
-layout) with:
-
-```bash
-# from the repo root; reads data/kandinsky5_overfit and writes
-# data/kandinsky5_overfit_preprocessed by default -- override with
-# KANDINSKY5_OVERFIT_DATA_DIR / KANDINSKY5_OVERFIT_OUTPUT_DIR env vars
-python -m fastvideo.pipelines.preprocess.preprocess_kandinsky5_overfit
-```
-
-This loads Kandinsky5's VAE (shared with HunyuanVideo) and both text
-encoders via FastVideo's own loaders (`VAELoader`/`TextEncoderLoader`), and
-writes `data_00000.parquet` + `validation_prompts.json` into the configured
-output directory. Point `training.data.data_path` at that output directory
-in the YAML configs below.
-
-There is currently no scalable/production preprocessing pipeline (the kind
-built on `BasePreprocessPipeline` that Wan/LTX2 use) for Kandinsky5 -- like
-Hunyuan, its dual text encoders don't fit that base class's single-encoder
-assumption, and `BasePreprocessPipeline.preprocess_video_and_text` isn't
-cleanly overridable in pieces. If you need to scale past a handful of
-overfit samples, extend `preprocess_kandinsky5_overfit.py`'s loop rather
-than forking the production pipeline base class.
-
-## Train stage 1 (Attn-QAT finetune)
-
-```bash
-bash examples/train/configs/fine_tuning/kandinsky5/finetune_qat.sh \
-    --training.data.data_path data/kandinsky5_overfit_preprocessed
-```
-
-`FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN` (set by the script) routes
-Kandinsky5's dense/local attention through the fake-quantized
-straight-through-estimator Triton kernel, so the DiT learns to absorb
-quantization error instead of fighting it. This is purely env-var driven,
-model-agnostic, and does not touch weights during training -- weight-level
-FP4/FP8 quantization is applied post-hoc at inference time (see below), not
-during either training stage.
-
-## Train stage 2 (QAT-aware DMD distillation)
-
-Stage 1 writes a raw DCP checkpoint (`checkpoint-<N>/dcp` + metadata/RNG
-state) under `training.checkpoint.output_dir`
-(`outputs/kandinsky5_t2v_qat_finetune/` by default) -- `models.*.init_from`
-needs a diffusers model directory instead (`model_index.json` + component
-subfolders). Pass the stage-1 checkpoint to the launcher, which performs
-the conversion and injects the export into all three `init_from` overrides:
-
-```bash
-bash examples/train/configs/distribution_matching/kandinsky5/distill_dmd_qat.sh \
-    outputs/kandinsky5_t2v_qat_finetune/checkpoint-<N>
-```
-
-Under the hood it runs exactly:
-
-```bash
-python -m fastvideo.train.entrypoint.dcp_to_diffusers \
-    --checkpoint outputs/kandinsky5_t2v_qat_finetune/checkpoint-<N> \
-    --output-dir outputs/kandinsky5_t2v_qat_finetune/checkpoint-<N>-diffusers \
-    --role student --overwrite --verify
-```
-
-then launches `dmd2_t2v_480p_qat.yaml` with
-`--models.student/teacher/critic.init_from` all pointing at the export.
-(Passing an already-exported diffusers directory to the launcher skips the
-conversion and uses it directly.)
-
-`--role student` exports the trained transformer weights (the only role
-that matters here: student/teacher/critic in stage 2 all load from the same
-checkpoint, and `_loading_teacher_critic_model` handles the full-precision
-masking for teacher/critic at load time, not at export time). `--verify`
-strictly reloads the exported transformer immediately, so a key-mapping bug
-fails here instead of deep inside the stage-2 launch.
-
-Only the student is quantized (Attn-QAT); the teacher and critic stay full
-precision / dense attention. This is enforced in the loader
-(`fastvideo/models/loader/component_loader.py`, via the
-`_loading_teacher_critic_model` flag), which masks `quant_config` and clears
-`FASTVIDEO_ATTENTION_BACKEND` for teacher/critic -- the same global env var
-reaches only the student, with no per-model flags or Kandinsky5-specific
-handling. Validation runs the distilled student through
-`Kandinsky5DMDPipeline` at the step counts configured in
-`callbacks.validation.sampling_steps`.
-
-## Export the stage-2 student for inference
-
-Stage 2 writes raw DCP checkpoints too, so the distilled student must be
-exported the same way stage 1 was before anything can load it -- the
-`path/to/kandinsky5_dmd_checkpoint` used below is exactly this export:
-
-```bash
-python -m fastvideo.train.entrypoint.dcp_to_diffusers \
-    --checkpoint outputs/kandinsky5_t2v_dmd2_4steps_qat/checkpoint-<N> \
-    --output-dir outputs/kandinsky5_t2v_dmd2_4steps_qat/checkpoint-<N>-diffusers \
-    --role student --verify
-```
-
-## Inference (NVFP4 / FP8 weight quantization)
-
-Kandinsky5's attention out-projection (`self_attention.out_layer` /
-`cross_attention.out_layer`) and FFN (`feed_forward.mlp.fc_in` /
-`feed_forward.mlp.fc_out`) names differ from Wan's (`to_out`, `ffn.fc_in` /
-`ffn.fc_out`); `to_query`/`to_key`/`to_value` already substring-match the
-`to_q`/`to_k`/`to_v` entries the quant configs use for Wan. Both
-`nvfp4_qat` and `FP8` now include Kandinsky5's out-projection/FFN names in
-their default layer lists (`fastvideo/layers/quantization/nvfp4_qat_config.py`,
-`fastvideo/layers/quantization/fp8_config.py`), so no `target_layers`
-override is needed:
-
-`dcp_to_diffusers` copies the base T2V checkpoint's `model_index.json`
-unchanged into every export (see "Train stage 2" above), so `_class_name`
-still says the base T2V pipeline and the registry
-(`fastvideo/registry.py`) has no way to auto-detect that a given directory
-is actually a DMD (four-step re-noise sampler) export -- it will resolve to
-`Kandinsky5T2VPipeline` and run the full-length sampler on DMD-distilled
-weights. Pass `override_pipeline_cls_name` and a `Kandinsky5DMDConfig`
-explicitly to select the right pipeline/sampler and get
-`dmd_denoising_steps` set (`Kandinsky5T2VConfig`'s default of `None` makes
-`Kandinsky5DmdDenoisingStage` raise immediately):
-
-```python
-from fastvideo import VideoGenerator
-from fastvideo.configs.pipelines.kandinsky5 import Kandinsky5DMDConfig
-from fastvideo.layers.quantization import get_quantization_config
-
-gen = VideoGenerator.from_pretrained(
-    "path/to/kandinsky5_dmd_checkpoint", num_gpus=1,
-    override_pipeline_cls_name="Kandinsky5DMDPipeline",
-    pipeline_config=Kandinsky5DMDConfig(),
-    transformer_quant=get_quantization_config("nvfp4_qat")(),
-    use_fsdp_inference=False,
-)
-# Kandinsky5DmdDenoisingStage ignores request.sampling.num_inference_steps
-# and guidance_scale -- it's a fixed 4-step, no-CFG sampler driven entirely
-# by pipeline_config.dmd_denoising_steps (set above). Adjust
-# Kandinsky5DMDConfig(dmd_denoising_steps=[...]) instead if the checkpoint
-# was distilled/validated with a different schedule than the default
-# [1000, 750, 500, 250].
-gen.generate(request={"prompt": "...", "output": {"save_video": True}})
-```
-
-As with Wan, combine with `FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_INFER` on an
-sm_120 GPU for the full 4-bit path; on other GPUs the attention falls back
-to a supported dense backend while the FP4 linear layers still run.
-`flashinfer` is required for the FP4 path.
-
-## Tests
-
- `fastvideo/tests/train/models/test_load_kandinsky5.py` -- loads the real
-  checkpoint and runs one transformer forward pass.
- `fastvideo/tests/train/models/test_kandinsky5_qat_attention_engages.py` --
-  confirms `ATTN_QAT_TRAIN` actually engages (not a silent SDPA fallback)
-  and that gradients flow through a forward+backward pass.
- `fastvideo/tests/api/test_kandinsky5_dmd_pipeline_resolution.py` --
-  confirms an unmodified export resolves to `Kandinsky5T2VPipeline` (the bug)
-  and that `override_pipeline_cls_name="Kandinsky5DMDPipeline"` fixes it (the
-  documented workaround above), plus `Kandinsky5DMDConfig`'s
-  `dmd_denoising_steps` default.
- `fastvideo/tests/nightly/test_e2e_kandinsky5_dmd_t2v_overfit.py` -- a few
-  steps of both training stages on a single synthetic sample, exercising the
-  whole recipe end to end: the `dcp_to_diffusers --verify` conversion
-  between the stages AND of the final stage-2 student, then reloading that
-  export through the documented `Kandinsky5DMDPipeline` /
-  `Kandinsky5DMDConfig` override, generating deterministically (fixed
-  seed), and comparing MS-SSIM against a committed reference video. The
-  test fails (does not skip) if the reference is missing; record it once on
-  a sanctioned GPU box with `KANDINSKY5_E2E_WRITE_REFERENCE=1`, review the
-  written video, and commit it (see the test's module docstring). Nightly
-  tests are not collected per-PR (`fastvideo/tests/contract/
-  test_ci_test_collection.py` allowlists the directory), so run it
-  explicitly:
-
-  ```bash
-  pytest fastvideo/tests/nightly/test_e2e_kandinsky5_dmd_t2v_overfit.py -vs
-  ```
@@ -1,33 +0,0 @@
-#!/bin/bash
-# QAD recipe -- quantization-aware finetune of Kandinsky5 T2V (480p) with
-# fake-quant (Attn-QAT) attention.
-#
-# The fake-quantized attention path is selected purely by env var
-# (config-driven, no monkey-patching): FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN
-# routes Kandinsky5's dense/local attention through the fake-quantized
-# straight-through-estimator kernel, so the DiT learns to absorb the
-# quantization error instead of fighting it. No weight quantization happens
-# during this stage; that's applied post-hoc at inference time (see the
-# stage-2 config and README).
-#
-# Scope: T2V, 480p only, dense/local attention only -- NABLA sparse attention
-# is never engaged by Kandinsky5Model at this resolution.
-#
-# Data: preprocess with fastvideo/pipelines/preprocess/preprocess_kandinsky5_overfit.py
-# (or an equivalent parquet dataset matching pyarrow_schema_t2v) first.
-#
-# Next: this writes a raw DCP checkpoint (checkpoint-N/dcp + metadata/RNG
-# state), not something stage 2 can load directly. Pass that checkpoint to
-# ../../distribution_matching/kandinsky5/distill_dmd_qat.sh, which converts
-# it with fastvideo.train.entrypoint.dcp_to_diffusers (--verify) and feeds
-# the export to all three models.*.init_from overrides automatically.
-set -euo pipefail
-
-export FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN   # <-- enables Attn-QAT training
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "${SCRIPT_DIR}/../../../../.." && pwd)"
-
-bash "${REPO_ROOT}/examples/train/run.sh" \
-    "${SCRIPT_DIR}/t2v_480p_qat.yaml" \
-    "$@"
@@ -1,85 +0,0 @@
-# Kandinsky5 T2V 480p stage-1 QAT finetune config.
-#
-# QAT here means attention-only fake-quant during training: launch with
-# FASTVIDEO_ATTENTION_BACKEND=ATTN_QAT_TRAIN set (see finetune_qat.sh), which
-# routes Kandinsky5's dense/local attention through the fake-quantized
-# straight-through-estimator kernel. No weight quantization happens during
-# training; that's applied post-hoc at inference time via the nvfp4_qat/FP8
-# quant configs (see the stage-2 config and README for that step).
-#
-# Scope: T2V, 480p only, dense/local attention only -- NABLA sparse attention
-# is never engaged by Kandinsky5Model at this resolution.
-#
-# Data must be preprocessed with Kandinsky5's VAE (shared with HunyuanVideo)
-# + dual text encoders (Qwen/Reason1 + CLIP) into parquet format before
-# training. See fastvideo/pipelines/preprocess/preprocess_kandinsky5_overfit.py.
-
-models:
-  student:
-    _target_: fastvideo.train.models.kandinsky5.Kandinsky5Model
-    init_from: kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers
-    trainable: true
-
-method:
-  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
-
-training:
-  distributed:
-    num_gpus: 8
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 8
-
-  data:
-    data_path: data/kandinsky5_overfit_preprocessed
-    dataloader_num_workers: 4
-    train_batch_size: 1
-    training_cfg_rate: 0.1
-    seed: 1000
-    num_latent_t: 31
-    num_height: 512
-    num_width: 768
-    num_frames: 121
-
-  optimizer:
-    learning_rate: 5.0e-5
-    betas: [0.9, 0.999]
-    weight_decay: 1.0e-4
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    max_train_steps: 2000
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/kandinsky5_t2v_qat_finetune
-    training_state_checkpointing_steps: 500
-    checkpoints_total_limit: 3
-    resume_from_checkpoint: latest
-
-  tracker:
-    project_name: fastvideo_kandinsky5
-    run_name: kandinsky5_t2v_qat_finetune
-
-  model:
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
-    max_grad_norm: 1.0
-  validation:
-    _target_: fastvideo.train.callbacks.validation.ValidationCallback
-    pipeline_target: fastvideo.pipelines.basic.kandinsky5.kandinsky5_pipeline.Kandinsky5T2VPipeline
-    dataset_file: data/kandinsky5_overfit_preprocessed/validation_prompts.json
-    every_steps: 250
-    # Cut way down from 50 while shaking out validation-path bugs -- at
-    # 512x768x121 each step is ~30s, so 50 steps/sample x 6 samples/rank
-    # was ~2.5h per validation round. Bump back to 50 for real runs.
-    sampling_steps: [25]
-    guidance_scale: 6.0
-
-pipeline:
-  flow_shift: 5
@@ -1,140 +0,0 @@
-# LTX-2 T2V NVFP4 quantization-aware fine-tune (QAT).
-#
-# Same recipe as fine_tuning/ltx2/t2v.yaml, with the deployment-matched
-# NVFP4 quantization from the validated
-# examples/train/configs/overfit_ltx2_t2v_nvfp4_qat.yaml run: the curated
-# attention/FFN linear set used by LTX-2 NVFP4 deployment runs its
-# forward through real NVFP4 GEMMs (flashinfer cutlass) with a
-# straight-through-estimator backward into the dense fp32 master weights.
-# No extra parameters or buffers are added, so FSDP sharding,
-# checkpointing, and export are identical to dense training. Video
-# attention uses ATTN_QAT_TRAIN for its quantized forward and STE
-# backward, then ATTN_QAT_INFER during validation. Head-dim-64 audio
-# attention and masked text attention remain dense.
-#
-# Validation-time ATTN_QAT_INFER is arch-aware:
-#   * sm_120 / sm_121: the bundled fastvideo-kernel CUTLASS extension.
-#   * sm_100 (GB200-class) / sm_103 (GB300-class): the FP4 FA4 kernel —
-#     requires the flash-attention-fp4 fork and its dependency pins; see
-#     docs/inference/optimizations.md.
-#   * any other GPU: falls back to flash attention with a logged receipt
-#     (validation still runs, just not with the quantized kernel).
-#
-# Prerequisites: preprocess your dataset into the trainer's latent format
-# (see docs/training/data_preprocess.md;
-# fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py is a minimal
-# LTX-2 reference to adapt). Data paths, step counts, and learning rate
-# below are placeholders you must adapt to your workload.
-#
-# Run (multi-GPU; see the validation note above for kernel availability):
-#   NUM_GPUS=4 bash examples/train/run.sh \
-#     examples/train/configs/fine_tuning/ltx2/nvfp4_qat_t2v.yaml
-#
-# `attn_qat_infer: true` below is the deployment-representative choice on
-# every arch with a quantized kernel available (see the validation note
-# above). Flip it off only when your GPU has no ATTN_QAT_INFER kernel and
-# you prefer validating with the train-time backend instead of the flash
-# fallback:
-#   NUM_GPUS=4 bash examples/train/run.sh \
-#     examples/train/configs/fine_tuning/ltx2/nvfp4_qat_t2v.yaml \
-#     --callbacks.validation.attn_qat_infer false
-
-models:
-  student:
-    _target_: fastvideo.train.models.ltx2.LTX2Model
-    # Validation settings below (8 steps, guidance 1.0) assume this
-    # distilled checkpoint.
-    init_from: FastVideo/LTX2-Distilled-Diffusers
-    trainable: true
-    enable_gradient_checkpointing_type: full
-    attention_backend: ATTN_QAT_TRAIN
-
-method:
-  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
-
-training:
-  distributed:
-    # The 13B video branch needs sharding; 4-way HSDP shard matches the
-    # validated LTX-2 runs. Scale num_gpus/hsdp_shard_dim together.
-    num_gpus: 4
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 4
-
-  data:
-    # REQUIRED: path to your preprocessed dataset.
-    data_path: data/your_dataset_preprocessed
-    dataloader_num_workers: 4
-    train_batch_size: 1
-    # LTX2Model requires 0.0: CFG dropout would zero post-connector
-    # embeddings, which is not the model's unconditional input.
-    training_cfg_rate: 0.0
-    seed: 42
-    # Must match your preprocessed resolution/length.
-    # num_latent_t = (num_frames - 1) / 8 + 1.
-    num_latent_t: 11
-    num_height: 480
-    num_width: 832
-    num_frames: 81
-
-  optimizer:
-    # Carried from the validated LTX-2 QAT overfit run as a starting
-    # point; tune for your dataset size and batch configuration.
-    learning_rate: 5.0e-5
-    betas: [0.9, 0.999]
-    weight_decay: 0.0
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    # Workload-dependent: set from your dataset size and target epochs.
-    max_train_steps: 2000
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/ltx2_t2v_nvfp4_qat_finetune
-    # A full training-state checkpoint is ~150GB for the 13B trainable
-    # video branch — size the interval and total limit to your storage.
-    training_state_checkpointing_steps: 500
-    checkpoints_total_limit: 2
-    resume_from_checkpoint: latest
-
-  tracker:
-    trackers: [wandb]
-    project_name: fastvideo_ltx2
-    run_name: ltx2_t2v_nvfp4_qat_finetune
-
-  model:
-    # LTX2Model.predict_noise converts the DiT's x0 output to velocity,
-    # so the default noise-minus-clean target reproduces the official
-    # unweighted masked-MSE (mask is all-ones for plain T2V).
-    precondition_outputs: false
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
-    max_grad_norm: 1.0
-  validation:
-    _target_: fastvideo.train.callbacks.validation.ValidationCallback
-    pipeline_target: fastvideo.pipelines.basic.ltx2.ltx2_pipeline.LTX2Pipeline
-    # REQUIRED: prompts to sample during training.
-    dataset_file: data/your_dataset_preprocessed/validation_prompts.json
-    every_steps: 250
-    # 8-step single-pass sampling matches the distilled checkpoint
-    # (validated in the LTX-2 overfit runs).
-    sampling_steps: [8]
-    guidance_scale: 1.0
-    num_frames: 81
-    # Validation reuses the live transformer and temporarily replaces its
-    # ATTN_QAT_TRAIN implementations with the sm120 inference kernel.
-    # Set false on GB200 (see header).
-    attn_qat_infer: true
-
-# quant_config selects NVFP4 QAT for the same curated linear prefixes as
-# LTX-2 NVFP4 deployment. They fake-quantize through real FP4 GEMMs with
-# an STE backward. The string resolves to NVFP4QATTrainConfig at parse.
-pipeline:
-  dit_config:
-    quant_config: nvfp4_qat_train
@@ -1,110 +0,0 @@
-# LTX-2 T2V full fine-tune.
-#
-# Starting point for fine-tuning LTX-2 on your own dataset with the
-# modular trainer. Model wiring (checkpoint, CFG rate, output
-# preconditioning, sharding) follows the validated
-# examples/train/configs/overfit_ltx2_t2v.yaml run; data paths, step
-# counts, and learning rate are placeholders you must adapt — the right
-# values depend on your dataset and are marked below.
-#
-# Prerequisites: preprocess your dataset into the trainer's latent format
-# (see docs/training/data_preprocess.md;
-# fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py is a minimal
-# LTX-2 reference to adapt).
-#
-# Run:
-#   NUM_GPUS=4 bash examples/train/run.sh \
-#     examples/train/configs/fine_tuning/ltx2/t2v.yaml
-
-models:
-  student:
-    _target_: fastvideo.train.models.ltx2.LTX2Model
-    # Validation settings below (8 steps, guidance 1.0) assume this
-    # distilled checkpoint.
-    init_from: FastVideo/LTX2-Distilled-Diffusers
-    trainable: true
-    enable_gradient_checkpointing_type: full
-
-method:
-  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
-
-training:
-  distributed:
-    # The 13B video branch needs sharding; 4-way HSDP shard matches the
-    # validated LTX-2 runs. Scale num_gpus/hsdp_shard_dim together.
-    num_gpus: 4
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 4
-
-  data:
-    # REQUIRED: path to your preprocessed dataset.
-    data_path: data/your_dataset_preprocessed
-    dataloader_num_workers: 4
-    train_batch_size: 1
-    # LTX2Model requires 0.0: CFG dropout would zero post-connector
-    # embeddings, which is not the model's unconditional input.
-    training_cfg_rate: 0.0
-    seed: 42
-    # Must match your preprocessed resolution/length.
-    # num_latent_t = (num_frames - 1) / 8 + 1.
-    num_latent_t: 11
-    num_height: 480
-    num_width: 832
-    num_frames: 81
-
-  optimizer:
-    # Carried from the validated LTX-2 overfit run as a starting point;
-    # tune for your dataset size and batch configuration.
-    learning_rate: 5.0e-5
-    betas: [0.9, 0.999]
-    weight_decay: 0.0
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    # Workload-dependent: set from your dataset size and target epochs.
-    max_train_steps: 2000
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/ltx2_t2v_finetune
-    # A full training-state checkpoint is ~150GB for the 13B trainable
-    # video branch — size the interval and total limit to your storage.
-    training_state_checkpointing_steps: 500
-    checkpoints_total_limit: 2
-    resume_from_checkpoint: latest
-
-  tracker:
-    trackers: [wandb]
-    project_name: fastvideo_ltx2
-    run_name: ltx2_t2v_finetune
-
-  model:
-    # LTX2Model.predict_noise converts the DiT's x0 output to velocity,
-    # so the default noise-minus-clean target reproduces the official
-    # unweighted masked-MSE (mask is all-ones for plain T2V).
-    precondition_outputs: false
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
-    max_grad_norm: 1.0
-  validation:
-    _target_: fastvideo.train.callbacks.validation.ValidationCallback
-    pipeline_target: fastvideo.pipelines.basic.ltx2.ltx2_pipeline.LTX2Pipeline
-    # REQUIRED: prompts to sample during training.
-    dataset_file: data/your_dataset_preprocessed/validation_prompts.json
-    every_steps: 250
-    # 8-step single-pass sampling matches the distilled checkpoint
-    # (validated in the LTX-2 overfit runs).
-    sampling_steps: [8]
-    guidance_scale: 1.0
-    num_frames: 81
-
-# Required so the LTX2T2VConfig pipeline config is resolved from
-# init_from (without a `pipeline:` key the loader falls back to a
-# generic PipelineConfig and the LTX-2 DiT cannot be constructed).
-pipeline: {}
@@ -1,114 +0,0 @@
-# LTX-2 T2V LoRA fine-tune.
-#
-# Same recipe as fine_tuning/ltx2/t2v.yaml, but only LoRA adapters on the
-# attention projections train — the base weights stay frozen, so
-# training-state checkpoints are adapter-sized instead of ~150GB.
-# See that file's header for prerequisites and placeholder notes.
-#
-# Run:
-#   NUM_GPUS=4 bash examples/train/run.sh \
-#     examples/train/configs/fine_tuning/ltx2/t2v_lora.yaml
-
-models:
-  student:
-    _target_: fastvideo.train.models.ltx2.LTX2Model
-    # Validation settings below (8 steps, guidance 1.0) assume this
-    # distilled checkpoint.
-    init_from: FastVideo/LTX2-Distilled-Diffusers
-    trainable: true
-    enable_gradient_checkpointing_type: full
-    lora:
-      enable: true
-      # rank/alpha and the attention-projection target set follow the
-      # Wan/Hunyuan LoRA recipes; targets match by substring against
-      # module names (LTX-2 blocks name theirs to_q/to_k/to_v/to_out).
-      rank: 16
-      alpha: 32
-      target_modules:
-        - to_q
-        - to_k
-        - to_v
-        - to_out
-
-method:
-  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
-
-training:
-  distributed:
-    # The 13B video branch needs sharding even with frozen base weights;
-    # 4-way HSDP shard matches the validated LTX-2 runs.
-    num_gpus: 4
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 4
-
-  data:
-    # REQUIRED: path to your preprocessed dataset.
-    data_path: data/your_dataset_preprocessed
-    dataloader_num_workers: 4
-    train_batch_size: 1
-    # LTX2Model requires 0.0: CFG dropout would zero post-connector
-    # embeddings, which is not the model's unconditional input.
-    training_cfg_rate: 0.0
-    seed: 42
-    # Must match your preprocessed resolution/length.
-    # num_latent_t = (num_frames - 1) / 8 + 1.
-    num_latent_t: 11
-    num_height: 480
-    num_width: 832
-    num_frames: 81
-
-  optimizer:
-    # LoRA-only training usually tolerates a higher LR than full
-    # fine-tuning (the Wan/Hunyuan LoRA recipes use 1.0e-4); tune for
-    # your dataset.
-    learning_rate: 1.0e-4
-    betas: [0.9, 0.999]
-    weight_decay: 0.0
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    # Workload-dependent: set from your dataset size and target epochs.
-    max_train_steps: 2000
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/ltx2_t2v_lora_finetune
-    training_state_checkpointing_steps: 500
-    checkpoints_total_limit: 2
-    resume_from_checkpoint: latest
-
-  tracker:
-    trackers: [wandb]
-    project_name: fastvideo_ltx2
-    run_name: ltx2_t2v_lora_finetune
-
-  model:
-    # LTX2Model.predict_noise converts the DiT's x0 output to velocity,
-    # so the default noise-minus-clean target reproduces the official
-    # unweighted masked-MSE (mask is all-ones for plain T2V).
-    precondition_outputs: false
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
-    max_grad_norm: 1.0
-  validation:
-    _target_: fastvideo.train.callbacks.validation.ValidationCallback
-    pipeline_target: fastvideo.pipelines.basic.ltx2.ltx2_pipeline.LTX2Pipeline
-    # REQUIRED: prompts to sample during training.
-    dataset_file: data/your_dataset_preprocessed/validation_prompts.json
-    every_steps: 250
-    # 8-step single-pass sampling matches the distilled checkpoint
-    # (validated in the LTX-2 overfit runs).
-    sampling_steps: [8]
-    guidance_scale: 1.0
-    num_frames: 81
-
-# Required so the LTX2T2VConfig pipeline config is resolved from
-# init_from (without a `pipeline:` key the loader falls back to a
-# generic PipelineConfig and the LTX-2 DiT cannot be constructed).
-pipeline: {}
@@ -1,151 +0,0 @@
-# LTX-2.3 T2V NVFP4 quantization-aware fine-tune (QAT).
-#
-# Same recipe as fine_tuning/ltx2_3/t2v.yaml, with the deployment-matched
-# NVFP4 quantization wiring carried from the validated LTX-2.0 QAT run
-# (examples/train/configs/overfit_ltx2_t2v_nvfp4_qat.yaml): the curated
-# attention/FFN linear set runs its forward through real NVFP4 GEMMs
-# (flashinfer cutlass) with a straight-through-estimator backward into
-# the dense fp32 master weights. No extra parameters or buffers are
-# added, so FSDP sharding, checkpointing, and export are identical to
-# dense training. Video attention uses ATTN_QAT_TRAIN for its quantized
-# forward and STE backward, then ATTN_QAT_INFER during validation.
-# Head-dim-64 audio attention and masked text attention remain dense.
-#
-# LTX-2.3 applicability: the curated linear prefix set
-# (fastvideo/layers/quantization/nvfp4_config.py) targets
-# ltx2.blocks.{0..47}.* — LTX-2.3 uses the same LTX-2 transformer class,
-# module naming, and 48-block depth, so the same 576 linears attach.
-# CAVEAT: the set was curated against LTX-2.0 NVFP4 deployment, and
-# LTX-2.3 QAT training has no validated overfit run yet. The 2.3
-# gated-attention to_gate_logits projections are not in the set and
-# remain dense.
-#
-# Validation-time ATTN_QAT_INFER is arch-aware:
-#   * sm_120 / sm_121: the bundled fastvideo-kernel CUTLASS extension.
-#   * sm_100 (GB200-class) / sm_103 (GB300-class): the FP4 FA4 kernel —
-#     requires the flash-attention-fp4 fork and its dependency pins; see
-#     docs/inference/optimizations.md.
-#   * any other GPU: falls back to flash attention with a logged receipt
-#     (validation still runs, just not with the quantized kernel).
-#
-# Prerequisites: preprocess your dataset into the trainer's latent format
-# (see docs/training/data_preprocess.md;
-# fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py is a minimal
-# LTX-2 reference to adapt — set
-# LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers). Data paths,
-# step counts, and learning rate below are placeholders you must adapt
-# to your workload.
-#
-# Run (multi-GPU; see the validation note above for kernel availability):
-#   NUM_GPUS=4 bash examples/train/run.sh \
-#     examples/train/configs/fine_tuning/ltx2_3/nvfp4_qat_t2v.yaml
-#
-# `attn_qat_infer: true` below is the deployment-representative choice on
-# every arch with a quantized kernel available (see the validation note
-# above). Flip it off only when your GPU has no ATTN_QAT_INFER kernel and
-# you prefer validating with the train-time backend instead of the flash
-# fallback:
-#   NUM_GPUS=4 bash examples/train/run.sh \
-#     examples/train/configs/fine_tuning/ltx2_3/nvfp4_qat_t2v.yaml \
-#     --callbacks.validation.attn_qat_infer false
-
-models:
-  student:
-    _target_: fastvideo.train.models.ltx2.LTX2Model
-    # Validation settings below (8 steps, guidance 1.0) assume this
-    # distilled checkpoint.
-    init_from: FastVideo/LTX-2.3-Distilled-Diffusers
-    trainable: true
-    enable_gradient_checkpointing_type: full
-    attention_backend: ATTN_QAT_TRAIN
-
-method:
-  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
-
-training:
-  distributed:
-    # The 13B video branch needs sharding; 4-way HSDP shard matches the
-    # validated LTX-2.3 runs. Scale num_gpus/hsdp_shard_dim together.
-    num_gpus: 4
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 4
-
-  data:
-    # REQUIRED: path to your preprocessed dataset.
-    data_path: data/your_dataset_preprocessed
-    dataloader_num_workers: 4
-    train_batch_size: 1
-    # LTX2Model requires 0.0: CFG dropout would zero post-connector
-    # embeddings, which is not the model's unconditional input.
-    training_cfg_rate: 0.0
-    seed: 42
-    # Must match your preprocessed resolution/length.
-    # num_latent_t = (num_frames - 1) / 8 + 1.
-    num_latent_t: 11
-    num_height: 480
-    num_width: 832
-    num_frames: 81
-
-  optimizer:
-    # Carried from the validated LTX-2.0 QAT overfit run as a starting
-    # point; tune for your dataset size and batch configuration.
-    learning_rate: 5.0e-5
-    betas: [0.9, 0.999]
-    weight_decay: 0.0
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    # Workload-dependent: set from your dataset size and target epochs.
-    max_train_steps: 2000
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/ltx2_3_t2v_nvfp4_qat_finetune
-    # A full training-state checkpoint is ~150GB for the 13B trainable
-    # video branch — size the interval and total limit to your storage.
-    training_state_checkpointing_steps: 500
-    checkpoints_total_limit: 2
-    resume_from_checkpoint: latest
-
-  tracker:
-    trackers: [wandb]
-    project_name: fastvideo_ltx2
-    run_name: ltx2_3_t2v_nvfp4_qat_finetune
-
-  model:
-    # LTX2Model.predict_noise converts the DiT's x0 output to velocity,
-    # so the default noise-minus-clean target reproduces the official
-    # unweighted masked-MSE (mask is all-ones for plain T2V).
-    precondition_outputs: false
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
-    max_grad_norm: 1.0
-  validation:
-    _target_: fastvideo.train.callbacks.validation.ValidationCallback
-    pipeline_target: fastvideo.pipelines.basic.ltx2.ltx2_pipeline.LTX2Pipeline
-    # REQUIRED: prompts to sample during training.
-    dataset_file: data/your_dataset_preprocessed/validation_prompts.json
-    every_steps: 250
-    # 8-step single-pass sampling matches the distilled checkpoint
-    # (validated in the LTX-2.3 overfit runs).
-    sampling_steps: [8]
-    guidance_scale: 1.0
-    num_frames: 81
-    # Validation reuses the live transformer and temporarily replaces its
-    # ATTN_QAT_TRAIN implementations with the sm120 inference kernel.
-    # Set false on GB200 (see header).
-    attn_qat_infer: true
-
-# quant_config selects NVFP4 QAT for the same curated linear prefixes as
-# LTX-2 NVFP4 deployment (see the LTX-2.3 applicability note in the
-# header). They fake-quantize through real FP4 GEMMs with an STE
-# backward. The string resolves to NVFP4QATTrainConfig at parse.
-pipeline:
-  dit_config:
-    quant_config: nvfp4_qat_train
@@ -1,112 +0,0 @@
-# LTX-2.3 T2V full fine-tune.
-#
-# Starting point for fine-tuning LTX-2.3 (gated attention, cross-attn
-# AdaLN, 4096-d post-connector text embeddings, no in-DiT caption
-# projection) on your own dataset with the modular trainer. Model wiring
-# (checkpoint, CFG rate, output preconditioning, sharding) follows the
-# validated examples/train/configs/overfit_ltx2_3_t2v.yaml run; data
-# paths, step counts, and learning rate are placeholders you must
-# adapt — the right values depend on your dataset and are marked below.
-#
-# Prerequisites: preprocess your dataset into the trainer's latent format
-# (see docs/training/data_preprocess.md;
-# fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py is a minimal
-# LTX-2 reference to adapt — set
-# LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers).
-#
-# Run:
-#   NUM_GPUS=4 bash examples/train/run.sh \
-#     examples/train/configs/fine_tuning/ltx2_3/t2v.yaml
-
-models:
-  student:
-    _target_: fastvideo.train.models.ltx2.LTX2Model
-    # Validation settings below (8 steps, guidance 1.0) assume this
-    # distilled checkpoint.
-    init_from: FastVideo/LTX-2.3-Distilled-Diffusers
-    trainable: true
-    enable_gradient_checkpointing_type: full
-
-method:
-  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
-
-training:
-  distributed:
-    # The 13B video branch needs sharding; 4-way HSDP shard matches the
-    # validated LTX-2.3 runs. Scale num_gpus/hsdp_shard_dim together.
-    num_gpus: 4
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 4
-
-  data:
-    # REQUIRED: path to your preprocessed dataset.
-    data_path: data/your_dataset_preprocessed
-    dataloader_num_workers: 4
-    train_batch_size: 1
-    # LTX2Model requires 0.0: CFG dropout would zero post-connector
-    # embeddings, which is not the model's unconditional input.
-    training_cfg_rate: 0.0
-    seed: 42
-    # Must match your preprocessed resolution/length.
-    # num_latent_t = (num_frames - 1) / 8 + 1.
-    num_latent_t: 11
-    num_height: 480
-    num_width: 832
-    num_frames: 81
-
-  optimizer:
-    # Carried from the validated LTX-2.3 overfit run as a starting point;
-    # tune for your dataset size and batch configuration.
-    learning_rate: 5.0e-5
-    betas: [0.9, 0.999]
-    weight_decay: 0.0
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    # Workload-dependent: set from your dataset size and target epochs.
-    max_train_steps: 2000
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/ltx2_3_t2v_finetune
-    # A full training-state checkpoint is ~150GB for the 13B trainable
-    # video branch — size the interval and total limit to your storage.
-    training_state_checkpointing_steps: 500
-    checkpoints_total_limit: 2
-    resume_from_checkpoint: latest
-
-  tracker:
-    trackers: [wandb]
-    project_name: fastvideo_ltx2
-    run_name: ltx2_3_t2v_finetune
-
-  model:
-    # LTX2Model.predict_noise converts the DiT's x0 output to velocity,
-    # so the default noise-minus-clean target reproduces the official
-    # unweighted masked-MSE (mask is all-ones for plain T2V).
-    precondition_outputs: false
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
-    max_grad_norm: 1.0
-  validation:
-    _target_: fastvideo.train.callbacks.validation.ValidationCallback
-    pipeline_target: fastvideo.pipelines.basic.ltx2.ltx2_pipeline.LTX2Pipeline
-    # REQUIRED: prompts to sample during training.
-    dataset_file: data/your_dataset_preprocessed/validation_prompts.json
-    every_steps: 250
-    # 8-step single-pass sampling matches the distilled checkpoint
-    # (validated in the LTX-2.3 overfit runs).
-    sampling_steps: [8]
-    guidance_scale: 1.0
-    num_frames: 81
-
-# Required so the LTX2T2VConfig pipeline config is resolved from
-# init_from (without a `pipeline:` key the loader falls back to a
-# generic PipelineConfig and the LTX-2 DiT cannot be constructed).
-pipeline: {}
@@ -1,116 +0,0 @@
-# LTX-2.3 T2V LoRA fine-tune.
-#
-# Same recipe as fine_tuning/ltx2_3/t2v.yaml, but only LoRA adapters on
-# the attention projections train — the base weights stay frozen, so
-# training-state checkpoints are adapter-sized instead of ~150GB.
-# See that file's header for prerequisites and placeholder notes.
-#
-# Run:
-#   NUM_GPUS=4 bash examples/train/run.sh \
-#     examples/train/configs/fine_tuning/ltx2_3/t2v_lora.yaml
-
-models:
-  student:
-    _target_: fastvideo.train.models.ltx2.LTX2Model
-    # Validation settings below (8 steps, guidance 1.0) assume this
-    # distilled checkpoint.
-    init_from: FastVideo/LTX-2.3-Distilled-Diffusers
-    trainable: true
-    enable_gradient_checkpointing_type: full
-    lora:
-      enable: true
-      # rank/alpha and the attention-projection target set follow the
-      # Wan/Hunyuan LoRA recipes; targets match by substring against
-      # module names (LTX-2.3 blocks name theirs to_q/to_k/to_v/to_out;
-      # the 2.3 gated-attention to_gate_logits projection is not
-      # targeted).
-      rank: 16
-      alpha: 32
-      target_modules:
-        - to_q
-        - to_k
-        - to_v
-        - to_out
-
-method:
-  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
-
-training:
-  distributed:
-    # The 13B video branch needs sharding even with frozen base weights;
-    # 4-way HSDP shard matches the validated LTX-2.3 runs.
-    num_gpus: 4
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 4
-
-  data:
-    # REQUIRED: path to your preprocessed dataset.
-    data_path: data/your_dataset_preprocessed
-    dataloader_num_workers: 4
-    train_batch_size: 1
-    # LTX2Model requires 0.0: CFG dropout would zero post-connector
-    # embeddings, which is not the model's unconditional input.
-    training_cfg_rate: 0.0
-    seed: 42
-    # Must match your preprocessed resolution/length.
-    # num_latent_t = (num_frames - 1) / 8 + 1.
-    num_latent_t: 11
-    num_height: 480
-    num_width: 832
-    num_frames: 81
-
-  optimizer:
-    # LoRA-only training usually tolerates a higher LR than full
-    # fine-tuning (the Wan/Hunyuan LoRA recipes use 1.0e-4); tune for
-    # your dataset.
-    learning_rate: 1.0e-4
-    betas: [0.9, 0.999]
-    weight_decay: 0.0
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    # Workload-dependent: set from your dataset size and target epochs.
-    max_train_steps: 2000
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/ltx2_3_t2v_lora_finetune
-    training_state_checkpointing_steps: 500
-    checkpoints_total_limit: 2
-    resume_from_checkpoint: latest
-
-  tracker:
-    trackers: [wandb]
-    project_name: fastvideo_ltx2
-    run_name: ltx2_3_t2v_lora_finetune
-
-  model:
-    # LTX2Model.predict_noise converts the DiT's x0 output to velocity,
-    # so the default noise-minus-clean target reproduces the official
-    # unweighted masked-MSE (mask is all-ones for plain T2V).
-    precondition_outputs: false
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
-    max_grad_norm: 1.0
-  validation:
-    _target_: fastvideo.train.callbacks.validation.ValidationCallback
-    pipeline_target: fastvideo.pipelines.basic.ltx2.ltx2_pipeline.LTX2Pipeline
-    # REQUIRED: prompts to sample during training.
-    dataset_file: data/your_dataset_preprocessed/validation_prompts.json
-    every_steps: 250
-    # 8-step single-pass sampling matches the distilled checkpoint
-    # (validated in the LTX-2.3 overfit runs).
-    sampling_steps: [8]
-    guidance_scale: 1.0
-    num_frames: 81
-
-# Required so the LTX2T2VConfig pipeline config is resolved from
-# init_from (without a `pipeline:` key the loader falls back to a
-# generic PipelineConfig and the LTX-2 DiT cannot be constructed).
-pipeline: {}
@@ -1,95 +0,0 @@
-# LTX-2.3 T2V overfitting test config.
-#
-# Overfits on a single short video (480x832, 81 frames @ 24fps) to
-# verify the LTX-2 training plugin works end-to-end on an LTX-2.3
-# checkpoint (gated attention, cross-attention AdaLN, 4096-d
-# post-connector text embeddings, no in-DiT caption projection).
-# Uses the distilled checkpoint (validation is 8 sampling steps).
-#
-# Preprocess data first (writes data/ltx2_3_overfit_preprocessed):
-#   CUDA_VISIBLE_DEVICES=0 \
-#   LTX2_OVERFIT_MODEL=FastVideo/LTX-2.3-Distilled-Diffusers \
-#   LTX2_OVERFIT_OUTPUT_DIR=data/ltx2_3_overfit_preprocessed \
-#   python fastvideo/pipelines/preprocess/preprocess_ltx2_overfit.py
-#
-# Run:
-#   NUM_GPUS=4 bash examples/train/run.sh examples/train/configs/overfit_ltx2_3_t2v.yaml
-
-models:
-  student:
-    _target_: fastvideo.train.models.ltx2.LTX2Model
-    init_from: FastVideo/LTX-2.3-Distilled-Diffusers
-    trainable: true
-    enable_gradient_checkpointing_type: full
-
-method:
-  _target_: fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod
-
-training:
-  distributed:
-    num_gpus: 4
-    sp_size: 1
-    tp_size: 1
-    hsdp_replicate_dim: 1
-    hsdp_shard_dim: 4
-
-  data:
-    data_path: data/ltx2_3_overfit_preprocessed
-    dataloader_num_workers: 0
-    train_batch_size: 1
-    # LTX2Model requires 0.0: CFG dropout would zero post-connector
-    # embeddings, which is not the model's unconditional input.
-    training_cfg_rate: 0.0
-    seed: 42
-    num_latent_t: 11  # (81 - 1) / 8 + 1
-    num_height: 480
-    num_width: 832
-    num_frames: 81
-
-  optimizer:
-    learning_rate: 5.0e-5
-    betas: [0.9, 0.999]
-    weight_decay: 0.0
-    lr_scheduler: constant
-    lr_warmup_steps: 0
-
-  loop:
-    max_train_steps: 300
-    gradient_accumulation_steps: 1
-
-  checkpoint:
-    output_dir: outputs/ltx2_3_overfit
-    # A full training-state checkpoint is ~150GB for the 13B trainable
-    # video branch; disable saves for the overfit smoke run.
-    training_state_checkpointing_steps: 0
-    checkpoints_total_limit: 1
-
-  tracker:
-    trackers: [wandb]
-    project_name: fastvideo_ltx2
-    run_name: ltx2_3_overfit
-
-  model:
-    # LTX2Model.predict_noise converts the DiT's x0 output to velocity,
-    # so the default noise-minus-clean target reproduces the official
-    # unweighted masked-MSE (mask is all-ones for plain T2V).
-    precondition_outputs: false
-    enable_gradient_checkpointing_type: full
-
-callbacks:
-  grad_clip:
-    _target_: fastvideo.train.callbacks.grad_clip.GradNormClipCallback
-    max_grad_norm: 1.0
-  validation:
-    _target_: fastvideo.train.callbacks.validation.ValidationCallback
-    pipeline_target: fastvideo.pipelines.basic.ltx2.ltx2_pipeline.LTX2Pipeline
-    dataset_file: data/ltx2_3_overfit_preprocessed/validation_prompts.json
-    every_steps: 50
-    sampling_steps: [8]
-    guidance_scale: 1.0
-    num_frames: 81
-
-# Required so the LTX2T2VConfig pipeline config is resolved from
-# init_from (without a `pipeline:` key the loader falls back to a
-# generic PipelineConfig and the LTX-2 DiT cannot be constructed).
-pipeline: {}
--- a/Show More
+++ b/Show More