[NPU] Diffusion CI Ground Truth Generation (NPU) (#24630)

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Elizaveta Martirosian <you@example.com>
Co-authored-by: Elizaveta Martirosian <elizaveta.martirosian@gmail.com>
Co-authored-by: ronnie_zheng <zl19940307@163.com>

.github/CODEOWNERS

@@ -10,6 +10,7 @@
 /python/sglang/multimodal_gen/runtime/cache @DefTruth
 /python/sglang/multimodal_gen/runtime/layers @mickqian @yhyang201 @BBuf @yingluosanqian @ping1jing2
 /python/sglang/multimodal_gen/runtime/models/dits @mickqian @yhyang201 @BBuf @yingluosanqian @ping1jing2
+/python/sglang/multimodal_gen/test/server/ascend  @ping1jing2 @ssshinigami @Makcum888e @e-martirosian
 /python/sglang/srt/batch_invariant_ops @Fridge003 @hebiao064
 /python/sglang/srt/compilation @hebiao064 @Oasis-Git
 /python/sglang/srt/constrained @hnyls2002 @DarkSharpness

.github/workflows/diffusion-ci-gt-gen-npu.yml

@@ -0,0 +1,214 @@
+name: Diffusion CI Ground Truth Generation (NPU)
+
+on:
+  workflow_dispatch:
+    inputs:
+      ref:
+        description: 'Git ref to checkout'
+        required: false
+        default: ''
+        type: string
+      case_ids:
+        description: 'Specific case IDs to run (space-separated, optional)'
+        required: false
+        default: ''
+        type: string
+      output_name:
+        description: 'Custom local output/artifact folder name. Leave empty to use defaults.'
+        required: false
+        default: ''
+        type: string
+      publish_target_dir:
+        description: 'Remote target directory in sgl-project/ci-data. Leave empty to use diffusion-ci/consistency_gt/sglang_generated/ascend.'
+        required: false
+        default: ''
+        type: string
+
+
+concurrency:
+  group: diffusion-ci-gt-gen-${{ github.ref }}-${{ inputs.output_name || inputs.case_ids || 'default' }}
+  cancel-in-progress: true
+
+permissions:
+  contents: write
+  actions: read
+
+env:
+  SGLANG_IS_IN_CI: true
+  OUTPUT_NAME: ${{ inputs.output_name || 'diffusion-ci-outputs' }}
+  PUBLISH_TARGET_DIR: ${{ inputs.publish_target_dir || 'diffusion-ci/consistency_gt/sglang_generated/ascend' }}
+
+jobs:
+
+  compute-diffusion-partitions:
+    if: github.repository == 'sgl-project/sglang'
+    runs-on: ubuntu-latest
+    outputs:
+      matrix-1npu: ${{ steps.compute.outputs.matrix-1npu }}
+      matrix-2npu: ${{ steps.compute.outputs.matrix-2npu }}
+      partition-count-1npu: ${{ steps.compute.outputs['partition-count-1npu'] }}
+      partition-count-2npu: ${{ steps.compute.outputs['partition-count-2npu'] }}
+      plan-1npu: ${{ steps.compute.outputs.plan-1npu }}
+      plan-2npu: ${{ steps.compute.outputs.plan-2npu }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+
+      - name: Compute partitions
+        env:
+          USE_NPU_CONFIGS: true
+        id: compute
+        run: |
+          python scripts/ci/utils/diffusion/compute_diffusion_partitions.py \
+            --min-time 1200 \
+            --target-time 1800 \
+            --max-time 2400 \
+            --max-partitions 10 \
+            --parametrized-only
+
+  multimodal-diffusion-gen-1npu:
+    needs: compute-diffusion-partitions
+    if: |
+      needs.compute-diffusion-partitions.result == 'success' &&
+      needs.compute-diffusion-partitions.outputs.matrix-1npu != '{"include":[]}'
+    runs-on: linux-aarch64-a3-2
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJson(needs.compute-diffusion-partitions.outputs.matrix-1npu) }}
+    timeout-minutes: 150
+    container:
+      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:9.0.0-a3-ubuntu22.04-py3.11
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Mark repository safe
+        run: |
+          git config --system --add safe.directory ${GITHUB_WORKSPACE}
+
+      - name: Install dependencies
+        env:
+          TORCH_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/whl/cpu"
+          PYPI_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple"
+          UV_INDEX_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple"
+          GITHUB_PROXY_URL: "https://gh-proxy.test.osinfra.cn/"
+          RUSTUP_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local:8082"
+        run: |
+          # speed up by using infra cache services
+          CACHING_URL="cache-service.nginx-pypi-cache.svc.cluster.local"
+          sed -Ei "s@(ports|archive).ubuntu.com@${CACHING_URL}:8081@g" /etc/apt/sources.list
+          pip config set global.index-url http://${CACHING_URL}/pypi/simple
+          pip config set global.trusted-host "${CACHING_URL}"
+
+          bash scripts/ci/npu/npu_ci_install_dependency.sh a3 diffusion
+
+      - name: Generate outputs
+        env:
+          PARTITION_PLAN_JSON: ${{ needs.compute-diffusion-partitions.outputs.plan-1npu }}
+          SGLANG_USE_MODELSCOPE: true
+          SGLANG_IS_IN_CI: true
+          HF_ENDPOINT: https://hf-mirror.com
+          TORCH_EXTENSIONS_DIR: /tmp/torch_extensions
+          PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+          STREAMS_PER_DEVICE: 32
+        run: |
+          cd python
+          python -m sglang.multimodal_gen.test.scripts.gen_diffusion_ci_outputs \
+            --suite 1-npu \
+            --partition-id ${{ matrix.part }} \
+            --total-partitions ${{ needs.compute-diffusion-partitions.outputs['partition-count-1npu'] }} \
+            --partition-plan-json "$PARTITION_PLAN_JSON" \
+            --out-dir ./${{ env.OUTPUT_NAME }} \
+            ${{ inputs.case_ids != '' && format('--case-ids {0}', inputs.case_ids) || '' }}
+
+      - name: Upload artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: ${{ env.OUTPUT_NAME }}-1npu-part${{ matrix.part }}
+          path: python/${{ env.OUTPUT_NAME }}
+          retention-days: 7
+
+      - name: Publish GT images to sgl-project/ci-data
+        env:
+          GITHUB_TOKEN: ${{ secrets.GH_PAT_FOR_NIGHTLY_CI_DATA }}
+        run: |
+          python scripts/ci/utils/diffusion/publish_diffusion_gt.py \
+            --source-dir python/${{ env.OUTPUT_NAME }} \
+            --target-dir "${{ env.PUBLISH_TARGET_DIR }}"
+
+  multimodal-diffusion-gen-2npu:
+    needs: compute-diffusion-partitions
+    if: |
+      needs.compute-diffusion-partitions.result == 'success' &&
+      needs.compute-diffusion-partitions.outputs.matrix-2npu != '{"include":[]}'
+    runs-on: linux-aarch64-a3-16
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJson(needs.compute-diffusion-partitions.outputs.matrix-2npu) }}
+    timeout-minutes: 150
+    container:
+      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:9.0.0-a3-ubuntu22.04-py3.11
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Mark repository safe
+        run: |
+          git config --system --add safe.directory ${GITHUB_WORKSPACE}
+
+      - name: Install dependencies
+        env:
+          TORCH_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/whl/cpu"
+          PYPI_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple"
+          UV_INDEX_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple"
+          GITHUB_PROXY_URL: "https://gh-proxy.test.osinfra.cn/"
+          RUSTUP_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local:8082"
+        run: |
+          # speed up by using infra cache services
+          CACHING_URL="cache-service.nginx-pypi-cache.svc.cluster.local"
+          sed -Ei "s@(ports|archive).ubuntu.com@${CACHING_URL}:8081@g" /etc/apt/sources.list
+          pip config set global.index-url http://${CACHING_URL}/pypi/simple
+          pip config set global.trusted-host "${CACHING_URL}"
+
+          bash scripts/ci/npu/npu_ci_install_dependency.sh a3 diffusion
+
+      - name: Generate outputs
+        env:
+          PARTITION_PLAN_JSON: ${{ needs.compute-diffusion-partitions.outputs.plan-2npu }}
+          SGLANG_USE_MODELSCOPE: true
+          SGLANG_IS_IN_CI: true
+          HF_ENDPOINT: https://hf-mirror.com
+          TORCH_EXTENSIONS_DIR: /tmp/torch_extensions
+          PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+          STREAMS_PER_DEVICE: 32
+        run: |
+          cd python
+          python -m sglang.multimodal_gen.test.scripts.gen_diffusion_ci_outputs \
+            --suite 2-npu \
+            --partition-id ${{ matrix.part }} \
+            --total-partitions ${{ needs.compute-diffusion-partitions.outputs['partition-count-2npu'] }} \
+            --partition-plan-json "$PARTITION_PLAN_JSON" \
+            --out-dir ./${{ env.OUTPUT_NAME }} \
+            ${{ inputs.case_ids != '' && format('--case-ids {0}', inputs.case_ids) || '' }}
+
+      - name: Upload artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: ${{ env.OUTPUT_NAME }}-2npu-part${{ matrix.part }}
+          path: python/${{ env.OUTPUT_NAME }}
+          retention-days: 7
+
+      - name: Publish GT images to sgl-project/ci-data
+        env:
+          GITHUB_TOKEN: ${{ secrets.GH_PAT_FOR_NIGHTLY_CI_DATA }}
+        run: |
+          python scripts/ci/utils/diffusion/publish_diffusion_gt.py \
+            --source-dir python/${{ env.OUTPUT_NAME }} \
+            --target-dir "${{ env.PUBLISH_TARGET_DIR }}"

.github/workflows/pr-test-npu.yml

@@ -341,7 +341,7 @@ jobs:
           SGLANG_DIFFUSION_ARTIFACT_DIR: ${{ github.workspace }}/diffusion-failures
         run: |
           cd python
-          python3 sglang/multimodal_gen/test/run_suite_npu.py --suite 1-npu
+          python3 sglang/multimodal_gen/test/run_suite.py --suite 1-npu
 
       - name: Upload diffusion failure artifacts
         if: always()
@@ -398,7 +398,7 @@ jobs:
           SGLANG_DIFFUSION_ARTIFACT_DIR: ${{ github.workspace }}/diffusion-failures
         run: |
           cd python
-          python3 sglang/multimodal_gen/test/run_suite_npu.py --suite 2-npu
+          python3 sglang/multimodal_gen/test/run_suite.py --suite 2-npu
 
       - name: Upload diffusion failure artifacts
         if: always()
@@ -409,63 +409,6 @@ jobs:
           if-no-files-found: ignore
           retention-days: 7
 
-  multimodal-gen-test-8-npu-a3:
-    needs: [check-changes, pr-gate, set-image-config]
-    if: needs.check-changes.outputs.multimodal_gen == 'true'
-    runs-on: linux-aarch64-a3-8
-    container:
-      image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-
-      - name: Mark repository safe
-        run: |
-          git config --system --add safe.directory ${GITHUB_WORKSPACE}
-
-      - name: Install dependencies
-        env:
-          TORCH_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/whl/cpu"
-          PYPI_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple"
-          UV_INDEX_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple"
-          GITHUB_PROXY_URL: "https://gh-proxy.test.osinfra.cn/"
-          RUSTUP_CACHE_URL: "http://cache-service.nginx-pypi-cache.svc.cluster.local:8082"
-        run: |
-          # speed up by using infra cache services
-          CACHING_URL="cache-service.nginx-pypi-cache.svc.cluster.local"
-          sed -Ei "s@(ports|archive).ubuntu.com@${CACHING_URL}:8081@g" /etc/apt/sources.list
-          pip config set global.index-url http://${CACHING_URL}/pypi/simple
-          pip config set global.trusted-host "${CACHING_URL}"
-
-          bash scripts/ci/npu/npu_ci_install_dependency.sh a3
-          # copy required file from our daily cache
-          cp ~/.cache/modelscope/hub/datasets/otavia/ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json /tmp
-          # copy gsm8k dataset
-          cp ~/.cache/modelscope/hub/datasets/tmp/test.jsonl /tmp
-
-      - name: Run test
-        timeout-minutes: 60
-        env:
-          SGLANG_USE_MODELSCOPE: true
-          SGLANG_IS_IN_CI: true
-          HF_ENDPOINT: https://hf-mirror.com
-          TORCH_EXTENSIONS_DIR: /tmp/torch_extensions
-          PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
-          STREAMS_PER_DEVICE: 32
-          SGLANG_DIFFUSION_ARTIFACT_DIR: ${{ github.workspace }}/diffusion-failures
-        run: |
-          cd python
-          python3 sglang/multimodal_gen/test/run_suite_npu.py --suite 8-npu
-
-      - name: Upload diffusion failure artifacts
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: diffusion-failures-npu-8-${{ github.run_attempt }}
-          path: diffusion-failures/
-          if-no-files-found: ignore
-          retention-days: 7
-
   pr-test-npu-finish:
     needs:
       [
@@ -478,7 +421,6 @@ jobs:
 
         multimodal-gen-test-1-npu-a3,
         multimodal-gen-test-2-npu-a3,
-        multimodal-gen-test-8-npu-a3,
       ]
     if: always()
     runs-on: ubuntu-latest

python/sglang/multimodal_gen/test/run_suite.py

@@ -19,117 +19,52 @@ from pathlib import Path
 
 import tabulate
 
+from sglang.multimodal_gen.runtime.platforms import current_platform
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
 from sglang.multimodal_gen.test.partitioning import (
     PartitionItem,
     partition_items_by_lpt,
 )
-from sglang.multimodal_gen.test.server.gpu_cases import (
-    ONE_GPU_CASES,
-    TWO_GPU_CASES,
-)
 from sglang.multimodal_gen.test.server.testcase_configs import (
     BASELINE_CONFIG,
     DiffusionTestCase,
 )
 
-logger = init_logger(__name__)
+# TODO: remove duplicated code
-
+if current_platform.is_npu():
-DEFAULT_EST_TIME_SECONDS = 300.0
+    from sglang.multimodal_gen.test.server.ascend.testcase_configs_npu import (
-STARTUP_OVERHEAD_SECONDS = 120.0
+        _UPDATE_WEIGHTS_FROM_DISK_TEST_FILE,
-DEFAULT_STANDALONE_EST_TIME_SECONDS = 300.0
+        COMPONENT_ACCURACY_SUITES,
-
+        DEFAULT_EST_TIME_SECONDS,
-_UPDATE_WEIGHTS_FROM_DISK_TEST_FILE = "test_update_weights_from_disk.py"
+        DEFAULT_STANDALONE_EST_TIME_SECONDS,
-_UPDATE_WEIGHTS_MODEL_PAIR_ENV = "SGLANG_MMGEN_UPDATE_WEIGHTS_PAIR"
+        FILE_SUITES,
-_UPDATE_WEIGHTS_MODEL_PAIR_IDS = (
+        PARAMETRIZED_CASE_GROUPS,
-    "FLUX.2-klein-base-4B",
+        STANDALONE_FILES,
-    "Qwen-Image",
+        STARTUP_OVERHEAD_SECONDS,
-)
+        SUITES,
-
+    )
-
+else:
-def _discover_unit_tests() -> list[str]:
+    from sglang.multimodal_gen.test.server.gpu_cases import (  # noqa: F401 It is used by ci scripts
-    unit_dir = Path(__file__).resolve().parent / "unit"
+        _UPDATE_WEIGHTS_FROM_DISK_TEST_FILE,
-    if not unit_dir.is_dir():
+        _UPDATE_WEIGHTS_MODEL_PAIR_ENV,
-        return []
+        _UPDATE_WEIGHTS_MODEL_PAIR_IDS,
-    return sorted(
+        COMPONENT_ACCURACY_FILE_NUM_GPUS,
-        f"../unit/{f.name}" for f in unit_dir.glob("test_*.py") if f.is_file()
+        COMPONENT_ACCURACY_SUITES,
+        DEFAULT_EST_TIME_SECONDS,
+        DEFAULT_STANDALONE_EST_TIME_SECONDS,
+        FILE_SUITES,
+        ONE_GPU_CASES,
+        PARAMETRIZED_CASE_GROUPS,
+        STANDALONE_FILE_EST_TIMES,
+        STANDALONE_FILES,
+        STARTUP_OVERHEAD_SECONDS,
+        STRICT_SUITES,
+        SUITES,
+        TWO_GPU_CASES,
     )
 
 
-FILE_SUITES = {
+logger = init_logger(__name__)
-    "unit": _discover_unit_tests(),
-    "component-accuracy": [
-        "test_component_accuracy_1_gpu.py",
-        "test_component_accuracy_2_gpu.py",
-    ],
-    "component-accuracy-1-gpu": [
-        "test_component_accuracy_1_gpu.py",
-    ],
-    "component-accuracy-2-gpu": [
-        "test_component_accuracy_2_gpu.py",
-    ],
-    "1-gpu-b200": [
-        "test_server_b200.py",
-    ],
-}
-
-PARAMETRIZED_CASE_GROUPS = {
-    "1-gpu": [
-        ("test_server_1_gpu.py", ONE_GPU_CASES),
-    ],
-    "2-gpu": [
-        ("test_server_2_gpu.py", TWO_GPU_CASES),
-    ],
-}
-
-STANDALONE_FILES = {
-    "1-gpu": [
-        "../cli/test_generate_t2i_perf.py",
-        # Temporarily disabled: 24 timeout failures since 2026-04-09 across
-        # multimodal-gen-test-1-gpu. Re-enable after the flakiness is fixed.
-        # "test_update_weights_from_disk.py",
-    ],
-    "2-gpu": [
-        "test_disagg_server.py",
-    ],
-}
-
-# New standalone files may omit an estimate once to learn the real CI runtime.
-# CI will use a fallback estimate for sharding, run the test, then print a
-# measured value that must be copied into STANDALONE_FILE_EST_TIMES.
-STANDALONE_FILE_EST_TIMES = {
-    "1-gpu": {
-        "../cli/test_generate_t2i_perf.py": 240.0,
-        # See STANDALONE_FILES note above — temporarily disabled.
-        # "test_update_weights_from_disk.py": 480.0,
-    },
-    "2-gpu": {
-        # Two disagg clusters × (~3 min startup + ~1 min generate) ≈ 8 min.
-        # Raise if CI reports a higher measured time.
-        "test_disagg_server.py": 600.0,
-    },
-}
-
-# Backward-compatible suite view for scripts that still operate on file lists.
-SUITES = {
-    **FILE_SUITES,
-    **{
-        suite: [filename for filename, _ in case_groups]
-        + STANDALONE_FILES.get(suite, [])
-        for suite, case_groups in PARAMETRIZED_CASE_GROUPS.items()
-    },
-}
-
-STRICT_SUITES = {"unit"}
-COMPONENT_ACCURACY_SUITES = {
-    "component-accuracy",
-    "component-accuracy-1-gpu",
-    "component-accuracy-2-gpu",
-}
-COMPONENT_ACCURACY_FILE_NUM_GPUS = {
-    "test_component_accuracy_1_gpu.py": 1,
-    "test_component_accuracy_2_gpu.py": 2,
-}
 
 
 @dataclass(frozen=True)

python/sglang/multimodal_gen/test/run_suite_npu.py

@@ -1,299 +0,0 @@
-"""
-Test runner for multimodal_gen that manages test suites and parallel execution.
-
-Usage:
-    python3 run_suite_npu.py --suite <suite_name> --partition-id <id> --total-partitions <num>
-
-Example:
-    python3 run_suite_npu.py --suite 1-npu --partition-id 0 --total-partitions 4
-"""
-
-import argparse
-import os
-import subprocess
-import sys
-from pathlib import Path
-
-import tabulate
-
-from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
-
-logger = init_logger(__name__)
-
-SUITES = {
-    "1-npu": [
-        "ascend/test_server_1_npu.py",
-        # add new 1-npu test files here
-    ],
-    "2-npu": [
-        "ascend/test_server_2_npu.py",
-        # add new 2-npu test files here
-    ],
-    "8-npu": [
-        "ascend/test_server_8_npu.py",
-        # add new 8-npu test files here
-    ],
-}
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Run multimodal_gen test suite")
-    parser.add_argument(
-        "--suite",
-        type=str,
-        required=True,
-        choices=list(SUITES.keys()),
-        help="The test suite to run (valid names are defined in SUITES)",
-    )
-    parser.add_argument(
-        "--partition-id",
-        type=int,
-        default=0,
-        help="Index of the current partition (for parallel execution)",
-    )
-    parser.add_argument(
-        "--total-partitions",
-        type=int,
-        default=1,
-        help="Total number of partitions",
-    )
-    parser.add_argument(
-        "--base-dir",
-        type=str,
-        default="server",
-        help="Base directory for tests relative to this script's parent",
-    )
-    parser.add_argument(
-        "-k",
-        "--filter",
-        type=str,
-        default=None,
-        help="Pytest filter expression (passed to pytest -k)",
-    )
-    parser.add_argument(
-        "--continue-on-error",
-        action="store_true",
-        default=False,
-        help="Continue running remaining tests even if one fails (for CI consistency; pytest already continues by default)",
-    )
-    return parser.parse_args()
-
-
-def collect_test_items(files, filter_expr=None):
-    """Collect test item node IDs from the given files using pytest --collect-only."""
-    cmd = [sys.executable, "-m", "pytest", "--collect-only", "-q"]
-    if filter_expr:
-        cmd.extend(["-k", filter_expr])
-    cmd.extend(files)
-
-    print(f"Collecting tests with command: {' '.join(cmd)}")
-    result = subprocess.run(cmd, capture_output=True, text=True)
-
-    # Check for collection errors
-    # pytest exit codes:
-    #   0: success
-    #   1: tests collected but some had errors during collection
-    #   2: test execution interrupted
-    #   3: internal error
-    #   4: command line usage error
-    #   5: no tests collected (may be expected with filters)
-    if result.returncode not in (0, 5):
-        error_msg = (
-            f"pytest --collect-only failed with exit code {result.returncode}\n"
-            f"Command: {' '.join(cmd)}\n"
-        )
-        if result.stderr:
-            error_msg += f"stderr:\n{result.stderr}\n"
-        if result.stdout:
-            error_msg += f"stdout:\n{result.stdout}\n"
-        logger.error(error_msg)
-        raise RuntimeError(error_msg)
-
-    if result.returncode == 5:
-        print(
-            "No tests were collected (exit code 5). This may be expected with filters."
-        )
-
-    # Parse the output to extract test node IDs
-    # pytest -q outputs lines like: test_file.py::TestClass::test_method[param]
-    test_items = []
-    for line in result.stdout.strip().split("\n"):
-        line = line.strip()
-        # Skip empty lines and summary lines
-        if line and "::" in line and not line.startswith(("=", "-", " ")):
-            # Handle lines that might have extra info after the test ID
-            test_id = line.split()[0] if " " in line else line
-            if "::" in test_id:
-                test_items.append(test_id)
-
-    print(f"Collected {len(test_items)} test items")
-    return test_items
-
-
-def run_pytest(files, filter_expr=None, exitfirst=False):
-    if not files:
-        print("No files to run.")
-        return 0
-
-    base_cmd = [sys.executable, "-m", "pytest", "-s", "-v"]
-    if exitfirst:
-        base_cmd.append("-x")
-
-    # Add pytest -k filter if provided
-    if filter_expr:
-        base_cmd.extend(["-k", filter_expr])
-
-    max_retries = 6
-    # retry if the perf assertion failed, for {max_retries} times
-    for i in range(max_retries + 1):
-        cmd = list(base_cmd)
-        if i > 0:
-            cmd.append("--last-failed")
-        # Always include files to constrain test discovery scope
-        # This prevents pytest from scanning the entire rootdir and
-        # discovering unrelated tests that may have missing dependencies
-        cmd.extend(files)
-
-        if i > 0:
-            print(
-                f"Performance assertion failed. Retrying ({i}/{max_retries}) with --last-failed..."
-            )
-
-        print(f"Running command: {' '.join(cmd)}")
-
-        process = subprocess.Popen(
-            cmd,
-            stdout=subprocess.PIPE,
-            stderr=subprocess.STDOUT,
-            bufsize=0,
-        )
-
-        output_bytes = bytearray()
-        while True:
-            chunk = process.stdout.read(4096)
-            if not chunk:
-                break
-            sys.stdout.buffer.write(chunk)
-            sys.stdout.buffer.flush()
-            output_bytes.extend(chunk)
-
-        process.wait()
-        returncode = process.returncode
-
-        if returncode == 0:
-            return 0
-
-        # Exit code 5 means no tests were collected/selected - treat as success
-        # when using filters, since some partitions may have all tests filtered out
-        if returncode == 5:
-            print(
-                "No tests collected (exit code 5). This is expected when filters "
-                "deselect all tests in a partition. Treating as success."
-            )
-            return 0
-
-        # check if the failure is due to an assertion in test_server_utils.py
-        full_output = output_bytes.decode("utf-8", errors="replace")
-        is_perf_assertion = (
-            "multimodal_gen/test/server/test_server_utils.py" in full_output
-            and "AssertionError" in full_output
-        )
-
-        is_flaky_ci_assertion = (
-            "SafetensorError" in full_output
-            or "FileNotFoundError" in full_output
-            or "TimeoutError" in full_output
-        )
-
-        is_oom_error = (
-            "out of memory" in full_output.lower()
-            or "oom killer" in full_output.lower()
-        )
-
-        if not (is_perf_assertion or is_flaky_ci_assertion or is_oom_error):
-            return returncode
-
-    print(f"Max retry exceeded")
-    return returncode
-
-
-def main():
-    args = parse_args()
-
-    # 1. resolve base path
-    current_file_path = Path(__file__).resolve()
-    test_root_dir = current_file_path.parent
-    target_dir = test_root_dir / args.base_dir
-
-    if not target_dir.exists():
-        print(f"Error: Target directory {target_dir} does not exist.")
-        sys.exit(1)
-
-    # 2. get files from suite
-    suite_files_rel = SUITES[args.suite]
-
-    suite_files_abs = []
-    for f_rel in suite_files_rel:
-        f_abs = target_dir / f_rel
-        if not f_abs.exists():
-            msg = f"Test file {f_rel} not found in {target_dir}."
-            print(f"Warning: {msg} Skipping.")
-            continue
-        suite_files_abs.append(str(f_abs))
-
-    if not suite_files_abs:
-        print(f"No valid test files found for suite '{args.suite}'.")
-        sys.exit(0)
-
-    # 3. collect all test items and partition by items (not files)
-    all_test_items = collect_test_items(suite_files_abs, filter_expr=args.filter)
-
-    if not all_test_items:
-        print(f"No test items found for suite '{args.suite}'.")
-        sys.exit(0)
-
-    # Partition by test items
-    my_items = [
-        item
-        for i, item in enumerate(all_test_items)
-        if i % args.total_partitions == args.partition_id
-    ]
-
-    # Print test info at beginning (similar to test/run_suite.py pretty_print_tests)
-    partition_info = f"{args.partition_id + 1}/{args.total_partitions} (0-based id={args.partition_id})"
-    headers = ["Suite", "Partition"]
-    rows = [[args.suite, partition_info]]
-    msg = tabulate.tabulate(rows, headers=headers, tablefmt="psql") + "\n"
-    msg += f"✅ Enabled {len(my_items)} test(s):\n"
-    for item in my_items:
-        msg += f"  - {item}\n"
-    print(msg, flush=True)
-    print(
-        f"Suite: {args.suite} | Partition: {args.partition_id}/{args.total_partitions}"
-    )
-    print(f"Selected {len(suite_files_abs)} files:")
-    for f in suite_files_abs:
-        print(f"  - {os.path.basename(f)}")
-
-    if not my_items:
-        print("No items assigned to this partition. Exiting success.")
-        sys.exit(0)
-
-    print(f"Running {len(my_items)} items in this shard: {', '.join(my_items)}")
-
-    # 4. execute with the specific test items
-    # Fast-fail: stop on first failure unless --continue-on-error is set
-    exit_code = run_pytest(my_items, exitfirst=not args.continue_on_error)
-
-    # Print tests again at the end for visibility
-    msg = "\n" + tabulate.tabulate(rows, headers=headers, tablefmt="psql") + "\n"
-    msg += f"✅ Executed {len(my_items)} test(s):\n"
-    for item in my_items:
-        msg += f"  - {item}\n"
-    print(msg, flush=True)
-
-    sys.exit(exit_code)
-
-
-if __name__ == "__main__":
-    main()

python/sglang/multimodal_gen/test/server/ascend/perf_baselines_npu.json

@@ -1,7 +1,7 @@
 {
     "metadata": {
         "model": "Diffusion Server",
-        "hardware": "CI A2 64GB pool",
+        "hardware": "Ascend A3",
         "description": "Reference numbers captured from the CI diffusion server baseline run"
     },
     "scenarios": {
@@ -72,69 +72,69 @@
         },
         "flux_2_image_t2i_2npu": {
             "stages_ms": {
-                "InputValidationStage": 0.06,
+                "InputValidationStage": 0.08,
-                "TextEncodingStage": 5628.31,
+                "TextEncodingStage": 192.4,
                 "ImageVAEEncodingStage": 0.01,
-                "LatentPreparationStage": 0.75,
+                "LatentPreparationStage": 0.97,
-                "TimestepPreparationStage": 30.68,
+                "TimestepPreparationStage": 34.65,
-                "DenoisingStage": 55002.26,
+                "DenoisingStage": 45390.41,
-                "DecodingStage": 43.73
+                "DecodingStage": 5.32
             },
             "denoise_step_ms": {
-                "0": 110.35,
+                "0": 84.23,
-                "1": 301.82,
+                "1": 80.0,
-                "2": 1139.81,
+                "2": 874.6,
-                "3": 1114.17,
+                "3": 918.81,
-                "4": 1099.34,
+                "4": 900.4,
-                "5": 1099.12,
+                "5": 910.67,
-                "6": 1100.16,
+                "6": 903.36,
-                "7": 1099.67,
+                "7": 904.97,
-                "8": 1099.09,
+                "8": 906.84,
-                "9": 1089.81,
+                "9": 906.39,
-                "10": 1109.73,
+                "10": 904.99,
-                "11": 1099.97,
+                "11": 909.96,
-                "12": 1100.26,
+                "12": 901.67,
-                "13": 1099.67,
+                "13": 908.8,
-                "14": 1099.79,
+                "14": 902.93,
-                "15": 1099.6,
+                "15": 906.17,
-                "16": 1100.16,
+                "16": 906.67,
-                "17": 1099.87,
+                "17": 905.89,
-                "18": 1100.02,
+                "18": 906.9,
-                "19": 1099.34,
+                "19": 907.1,
-                "20": 1099.6,
+                "20": 905.31,
-                "21": 1099.45,
+                "21": 907.93,
-                "22": 1100.2,
+                "22": 903.68,
-                "23": 1099.29,
+                "23": 904.49,
-                "24": 1098.86,
+                "24": 905.73,
-                "25": 1090.38,
+                "25": 907.66,
-                "26": 1109.19,
+                "26": 906.71,
-                "27": 1099.67,
+                "27": 912.69,
-                "28": 1100.06,
+                "28": 901.4,
-                "29": 1099.22,
+                "29": 909.9,
-                "30": 1100.08,
+                "30": 901.72,
-                "31": 1098.86,
+                "31": 904.35,
-                "32": 1099.73,
+                "32": 905.61,
-                "33": 1099.11,
+                "33": 905.97,
-                "34": 1100.13,
+                "34": 906.5,
-                "35": 1103.97,
+                "35": 921.45,
-                "36": 1095.26,
+                "36": 892.58,
-                "37": 1099.38,
+                "37": 909.03,
-                "38": 1099.34,
+                "38": 903.82,
-                "39": 1099.17,
+                "39": 906.25,
-                "40": 1100.08,
+                "40": 905.18,
-                "41": 1089.89,
+                "41": 905.06,
-                "42": 1106.69,
+                "42": 906.21,
-                "43": 1102.57,
+                "43": 911.62,
-                "44": 1100.17,
+                "44": 901.04,
-                "45": 1099.21,
+                "45": 909.09,
-                "46": 1100.42,
+                "46": 904.29,
-                "47": 1099.38,
+                "47": 905.16,
-                "48": 1099.59,
+                "48": 907.35,
-                "49": 1099.47
+                "49": 905.65
             },
-            "expected_e2e_ms": 64195.08,
+            "expected_e2e_ms": 46557.7,
-            "expected_avg_denoise_ms": 1065.0,
+            "expected_avg_denoise_ms": 872.7,
-            "expected_median_denoise_ms": 1099.63
+            "expected_median_denoise_ms": 905.81
         },
         "wan2_1_t2v_1.3b_1_npu": {
             "stages_ms": {
@@ -203,61 +203,61 @@
             "expected_median_denoise_ms": 537.54,
             "estimated_full_test_time_s": 157.8
         },
-       "wan2_2_t2v_14b_w8a8_8npu": {
+        "wan2_2_t2v_14b_w8a8_2npu": {
             "stages_ms": {
-                "InputValidationStage": 0.14,
+                "InputValidationStage": 0.09,
-                "TextEncodingStage": 3020.73,
+                "TextEncodingStage": 2789.3,
-                "LatentPreparationStage": 0.19,
+                "LatentPreparationStage": 0.28,
-                "TimestepPreparationStage": 5.01,
+                "TimestepPreparationStage": 3.19,
-                "DenoisingStage": 82744.33,
+                "DenoisingStage": 187650.19,
-                "DecodingStage": 932.41,
+                "DecodingStage": 3491.88,
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 1232.32,
+                "0": 1415.84,
-                "1": 2091.77,
+                "1": 4801.54,
-                "2": 2097.62,
+                "2": 4781.57,
-                "3": 2087.53,
+                "3": 4770.19,
-                "4": 2088.54,
+                "4": 4808.86,
-                "5": 2087.96,
+                "5": 4737.55,
-                "6": 2088.28,
+                "6": 4774.02,
-                "7": 2089.77,
+                "7": 4774.85,
-                "8": 2101.9,
+                "8": 4773.25,
-                "9": 2088.73,
+                "9": 4775.33,
-                "10": 2088.04,
+                "10": 4771.31,
-                "11": 2087.53,
+                "11": 4773.07,
-                "12": 2088.89,
+                "12": 4773.18,
-                "13": 2087.09,
+                "13": 4772.77,
-                "14": 2088.25,
+                "14": 4773.68,
-                "15": 2087.96,
+                "15": 4771.91,
-                "16": 2088.24,
+                "16": 4776.16,
-                "17": 2088.45,
+                "17": 4773.88,
-                "18": 2104.7,
+                "18": 4769.87,
-                "19": 2088.44,
+                "19": 4772.1,
-                "20": 2087.19,
+                "20": 4774.55,
-                "21": 2088.19,
+                "21": 4772.54,
-                "22": 2088.37,
+                "22": 4772.39,
-                "23": 2087.6,
+                "23": 4776.14,
-                "24": 2088.13,
+                "24": 4772.84,
-                "25": 2088.06,
+                "25": 4772.82,
-                "26": 2126.23,
+                "26": 4789.34,
-                "27": 2089.92,
+                "27": 4792.58,
-                "28": 2087.37,
+                "28": 4807.52,
-                "29": 2089.21,
+                "29": 4740.66,
-                "30": 2088.29,
+                "30": 4773.46,
-                "31": 2087.89,
+                "31": 4774.19,
-                "32": 2073.1,
+                "32": 4772.8,
-                "33": 2086.71,
+                "33": 4774.84,
-                "34": 2087.88,
+                "34": 4772.52,
-                "35": 2088.64,
+                "35": 4773.77,
-                "36": 2088.1,
+                "36": 4775.43,
-                "37": 2089.14,
+                "37": 4771.58,
-                "38": 2087.5,
+                "38": 4772.68,
-                "39": 2087.86
+                "39": 4769.1
             },
-            "expected_e2e_ms": 86719.57,
+            "expected_e2e_ms": 193947.19,
-            "expected_avg_denoise_ms": 2068.43,
+            "expected_avg_denoise_ms": 4691.07,
-            "expected_median_denoise_ms": 2088.21
+            "expected_median_denoise_ms": 4773.22
         },
         "qwen_image_t2i_2npu": {
             "stages_ms": {

python/sglang/multimodal_gen/test/server/ascend/test_server_8_npu.py

@@ -1,31 +0,0 @@
-"""
-Config-driven diffusion performance test with pytest parametrization.
-
-
-If the actual run is significantly better than the baseline, the improved cases with their updated baseline will be printed
-"""
-
-from __future__ import annotations
-
-import pytest
-
-from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
-from sglang.multimodal_gen.test.server.ascend.testcase_configs_npu import (
-    EIGHT_NPU_CASES,
-)
-from sglang.multimodal_gen.test.server.test_server_common import (  # noqa: F401
-    DiffusionServerBase,
-    diffusion_server,
-)
-from sglang.multimodal_gen.test.server.testcase_configs import DiffusionTestCase
-
-logger = init_logger(__name__)
-
-
-class TestDiffusionServerEightNpu(DiffusionServerBase):
-    """Performance tests for 8-NPU diffusion cases."""
-
-    @pytest.fixture(params=EIGHT_NPU_CASES, ids=lambda c: c.id)
-    def case(self, request) -> DiffusionTestCase:
-        """Provide a DiffusionTestCase for each 8-NPU test."""
-        return request.param

python/sglang/multimodal_gen/test/server/ascend/testcase_configs_npu.py

@@ -1,3 +1,5 @@
+import os
+
 from sglang.multimodal_gen.test.server.testcase_configs import (
     T2V_PROMPT,
     DiffusionSamplingParams,
@@ -6,12 +8,31 @@ from sglang.multimodal_gen.test.server.testcase_configs import (
     T2I_sampling_params,
 )
 
+MODEL_WEIGHTS_DIR = "/root/.cache/modelscope/hub/models/"
+
+FLUX_1_DEV_WEIGHTS_PATH = os.path.join(
+    MODEL_WEIGHTS_DIR, "black-forest-labs/FLUX.1-dev"
+)
+FLUX_2_DEV_WEIGHTS_PATH = os.path.join(
+    MODEL_WEIGHTS_DIR, "black-forest-labs/FLUX.2-dev"
+)
+QWEN_IMAGE_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "Qwen/Qwen-Image")
+WAN2_1_T2V_1_3B_DIFFUSERS_WEIGHTS_PATH = os.path.join(
+    MODEL_WEIGHTS_DIR, "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+)
+WAN2_2_T2V_A14B_DIFFUSERS_W8A8_WEIGHTS_PATH = os.path.join(
+    MODEL_WEIGHTS_DIR, "Eco-Tech/Wan2.2-T2V-A14B-Diffusers-w8a8"
+)
+
+EXTRAS_DISABLE_WARMUP = ["--server-warmup", "false"]
+
 ONE_NPU_CASES: list[DiffusionTestCase] = [
     # === Text to Image (T2I) ===
     DiffusionTestCase(
         "flux_image_t2i_npu",
         DiffusionServerArgs(
-            model_path="/root/.cache/modelscope/hub/models/black-forest-labs/FLUX.1-dev",
+            model_path=FLUX_1_DEV_WEIGHTS_PATH,
+            extras=EXTRAS_DISABLE_WARMUP,
         ),
         T2I_sampling_params,
         run_consistency_check=False,
@@ -20,7 +41,8 @@ ONE_NPU_CASES: list[DiffusionTestCase] = [
     DiffusionTestCase(
         "wan2_1_t2v_1.3b_1_npu",
         DiffusionServerArgs(
-            model_path="/root/.cache/modelscope/hub/models/Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+            model_path=WAN2_1_T2V_1_3B_DIFFUSERS_WEIGHTS_PATH,
+            extras=EXTRAS_DISABLE_WARMUP,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -34,9 +56,10 @@ TWO_NPU_CASES: list[DiffusionTestCase] = [
     DiffusionTestCase(
         "flux_2_image_t2i_2npu",
         DiffusionServerArgs(
-            model_path="/root/.cache/modelscope/hub/models/black-forest-labs/FLUX.2-dev",
+            model_path=FLUX_2_DEV_WEIGHTS_PATH,
             num_gpus=2,
             tp_size=2,
+            extras=EXTRAS_DISABLE_WARMUP,
         ),
         T2I_sampling_params,
         run_consistency_check=False,
@@ -44,26 +67,25 @@ TWO_NPU_CASES: list[DiffusionTestCase] = [
     DiffusionTestCase(
         "qwen_image_t2i_2npu",
         DiffusionServerArgs(
-            model_path="/root/.cache/modelscope/hub/models/Qwen/Qwen-Image",
+            model_path=QWEN_IMAGE_WEIGHTS_PATH,
             num_gpus=2,
             # test ring attn
             ulysses_degree=1,
             ring_degree=2,
+            extras=EXTRAS_DISABLE_WARMUP,
         ),
         T2I_sampling_params,
         run_consistency_check=False,
     ),
-]
-
-EIGHT_NPU_CASES: list[DiffusionTestCase] = [
     # === Text to Video (T2V) ===
     DiffusionTestCase(
-        "wan2_2_t2v_14b_w8a8_8npu",
+        "wan2_2_t2v_14b_w8a8_2npu",
         DiffusionServerArgs(
-            model_path="/root/.cache/modelscope/hub/models/Eco-Tech/Wan2.2-T2V-A14B-Diffusers-w8a8",
+            model_path=WAN2_2_T2V_A14B_DIFFUSERS_W8A8_WEIGHTS_PATH,
-            num_gpus=8,
+            num_gpus=2,
-            tp_size=4,
+            tp_size=1,
             ulysses_degree=2,
+            extras=EXTRAS_DISABLE_WARMUP,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -71,3 +93,32 @@ EIGHT_NPU_CASES: list[DiffusionTestCase] = [
         run_consistency_check=False,
     ),
 ]
+
+DEFAULT_EST_TIME_SECONDS = 300.0
+STARTUP_OVERHEAD_SECONDS = 120.0
+DEFAULT_STANDALONE_EST_TIME_SECONDS = 300.0
+
+SUITES = {
+    "1-npu": [
+        "ascend/test_server_1_npu.py",
+        # add new 1-npu test files here
+    ],
+    "2-npu": [
+        "ascend/test_server_2_npu.py",
+        # add new 2-npu test files here
+    ],
+}
+
+PARAMETRIZED_CASE_GROUPS = {
+    "1-npu": [
+        ("ascend/test_server_1_npu.py", ONE_NPU_CASES),
+    ],
+    "2-npu": [
+        ("ascend/test_server_2_npu.py", TWO_NPU_CASES),
+    ],
+}
+
+FILE_SUITES = {}
+STANDALONE_FILES = {}
+COMPONENT_ACCURACY_SUITES = {}
+_UPDATE_WEIGHTS_FROM_DISK_TEST_FILE = None

python/sglang/multimodal_gen/test/server/gpu_cases.py

@@ -787,3 +787,99 @@ if not current_platform.is_hip():
 
 ONE_GPU_CASES += ONE_GPU_MODELOPT_FP8_CASES
 TWO_GPU_CASES = _with_default_num_gpus(TWO_GPU_CASES, 2)
+
+
+def _discover_unit_tests() -> list[str]:
+    unit_dir = Path(__file__).resolve().parent.parent / "unit"
+    if not unit_dir.is_dir():
+        return []
+    return sorted(
+        f"../unit/{f.name}" for f in unit_dir.glob("test_*.py") if f.is_file()
+    )
+
+
+FILE_SUITES = {
+    "unit": _discover_unit_tests(),
+    "component-accuracy": [
+        "test_component_accuracy_1_gpu.py",
+        "test_component_accuracy_2_gpu.py",
+    ],
+    "component-accuracy-1-gpu": [
+        "test_component_accuracy_1_gpu.py",
+    ],
+    "component-accuracy-2-gpu": [
+        "test_component_accuracy_2_gpu.py",
+    ],
+    "1-gpu-b200": [
+        "test_server_b200.py",
+    ],
+}
+
+PARAMETRIZED_CASE_GROUPS = {
+    "1-gpu": [
+        ("test_server_1_gpu.py", ONE_GPU_CASES),
+    ],
+    "2-gpu": [
+        ("test_server_2_gpu.py", TWO_GPU_CASES),
+    ],
+}
+
+STANDALONE_FILES = {
+    "1-gpu": [
+        "../cli/test_generate_t2i_perf.py",
+        # Temporarily disabled: 24 timeout failures since 2026-04-09 across
+        # multimodal-gen-test-1-gpu. Re-enable after the flakiness is fixed.
+        # "test_update_weights_from_disk.py",
+    ],
+    "2-gpu": [
+        "test_disagg_server.py",
+    ],
+}
+
+# New standalone files may omit an estimate once to learn the real CI runtime.
+# CI will use a fallback estimate for sharding, run the test, then print a
+# measured value that must be copied into STANDALONE_FILE_EST_TIMES.
+STANDALONE_FILE_EST_TIMES = {
+    "1-gpu": {
+        "../cli/test_generate_t2i_perf.py": 240.0,
+        # See STANDALONE_FILES note above — temporarily disabled.
+        # "test_update_weights_from_disk.py": 480.0,
+    },
+    "2-gpu": {
+        # Two disagg clusters × (~3 min startup + ~1 min generate) ≈ 8 min.
+        # Raise if CI reports a higher measured time.
+        "test_disagg_server.py": 600.0,
+    },
+}
+
+# Backward-compatible suite view for scripts that still operate on file lists.
+SUITES = {
+    **FILE_SUITES,
+    **{
+        suite: [filename for filename, _ in case_groups]
+        + STANDALONE_FILES.get(suite, [])
+        for suite, case_groups in PARAMETRIZED_CASE_GROUPS.items()
+    },
+}
+
+STRICT_SUITES = {"unit"}
+COMPONENT_ACCURACY_SUITES = {
+    "component-accuracy",
+    "component-accuracy-1-gpu",
+    "component-accuracy-2-gpu",
+}
+COMPONENT_ACCURACY_FILE_NUM_GPUS = {
+    "test_component_accuracy_1_gpu.py": 1,
+    "test_component_accuracy_2_gpu.py": 2,
+}
+
+DEFAULT_EST_TIME_SECONDS = 300.0
+STARTUP_OVERHEAD_SECONDS = 120.0
+DEFAULT_STANDALONE_EST_TIME_SECONDS = 300.0
+
+_UPDATE_WEIGHTS_FROM_DISK_TEST_FILE = "test_update_weights_from_disk.py"
+_UPDATE_WEIGHTS_MODEL_PAIR_ENV = "SGLANG_MMGEN_UPDATE_WEIGHTS_PAIR"
+_UPDATE_WEIGHTS_MODEL_PAIR_IDS = (
+    "FLUX.2-klein-base-4B",
+    "Qwen-Image",
+)

python/sglang/multimodal_gen/test/test_utils.py

@@ -45,10 +45,18 @@ SGL_TEST_FILES_OFFICIAL_CONSISTENCY_GT_BASE = (
 SGL_TEST_FILES_SGLANG_CONSISTENCY_GT_BASE = (
     f"{SGL_TEST_FILES_CONSISTENCY_GT_ROOT}/sglang_generated"
 )
+SGL_TEST_FILES_OFFICIAL_CONSISTENCY_GT_BASE_ASCEND = (
+    f"{SGL_TEST_FILES_CONSISTENCY_GT_ROOT}/official_generated/ascend"
+)
+SGL_TEST_FILES_SGLANG_CONSISTENCY_GT_BASE_ASCEND = (
+    f"{SGL_TEST_FILES_CONSISTENCY_GT_ROOT}/sglang_generated/ascend"
+)
 SGL_TEST_FILES_CONSISTENCY_GT_BASE = SGL_TEST_FILES_SGLANG_CONSISTENCY_GT_BASE
 SGL_TEST_FILES_CONSISTENCY_GT_BASES = (
     SGL_TEST_FILES_OFFICIAL_CONSISTENCY_GT_BASE,
     SGL_TEST_FILES_SGLANG_CONSISTENCY_GT_BASE,
+    SGL_TEST_FILES_OFFICIAL_CONSISTENCY_GT_BASE_ASCEND,
+    SGL_TEST_FILES_SGLANG_CONSISTENCY_GT_BASE_ASCEND,
 )
 # LTX cases listed here compare against official-generated GT.
 SGL_TEST_FILES_OFFICIAL_CONSISTENCY_GT_CASES = frozenset(

scripts/ci/utils/diffusion/compute_diffusion_partitions.py

@@ -38,6 +38,12 @@ def _load_partitioning_helpers():
 PartitionItem, partition_items_by_lpt = _load_partitioning_helpers()
 
 SUITE_OUTPUT_NAMES = {"1-gpu": "1gpu", "2-gpu": "2gpu", "1-gpu-b200": "b200"}
+
+USE_NPU_CONFIGS = os.getenv("USE_NPU_CONFIGS", "0").lower() in ("1", "true")
+
+if USE_NPU_CONFIGS:
+    SUITE_OUTPUT_NAMES = {"1-npu": "1npu", "2-npu": "2npu"}
+
 DEFAULT_STANDALONE_EST_TIME_SECONDS = 300.0
 
 
@@ -261,6 +267,12 @@ def main():
         print(f"Error: Run suite not found: {run_suite_path}")
         sys.exit(1)
     try:
+        if USE_NPU_CONFIGS:
+            case_config_path = (
+                repo_root
+                / "python/sglang/multimodal_gen/test/server/ascend/testcase_configs_npu.py"
+            )
+        else:
             case_config_path = resolve_case_config_path(repo_root, run_suite_path)
     except (RuntimeError, FileNotFoundError) as exc:
         print(f"Error: {exc}")

scripts/ci/utils/diffusion/diffusion_case_parser.py

@@ -16,6 +16,7 @@ Usage:
 
 import ast
 import json
+import os
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Dict, List, Optional
@@ -44,6 +45,17 @@ STARTUP_OVERHEAD_SECONDS = 120.0
 BASELINE_REL_PATH = "python/sglang/multimodal_gen/test/server/perf_baselines.json"
 RUN_SUITE_REL_PATH = "python/sglang/multimodal_gen/test/run_suite.py"
 
+USE_NPU_CONFIGS = os.getenv("USE_NPU_CONFIGS", "0").lower() in ("1", "true")
+
+if USE_NPU_CONFIGS:
+    BASELINE_REL_PATH = (
+        "python/sglang/multimodal_gen/test/server/perf_baselines_npu.json"
+    )
+    CASE_LIST_TO_SUITE = {
+        "ONE_NPU_CASES": "1-npu",
+        "TWO_NPU_CASES": "2-npu",
+    }
+
 
 @dataclass
 class DiffusionCaseInfo: