[diffusion] CI: fix consistency check (#22251)

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

@@ -22,6 +22,10 @@ permissions:
   contents: write
   actions: read
 
+env:
+  SGLANG_IS_IN_CI: true
+  SGLANG_CUDA_COREDUMP: "1"
+
 jobs:
   multimodal-diffusion-gen-1gpu:
     if: github.repository == 'sgl-project/sglang'
@@ -40,6 +44,8 @@ jobs:
         run: bash scripts/ci/cuda/ci_install_dependency.sh diffusion
 
       - name: Generate outputs
+        env:
+          RUNAI_STREAMER_MEMORY_LIMIT: 0
         run: |
           cd python
           python -m sglang.multimodal_gen.test.scripts.gen_diffusion_ci_outputs \
@@ -56,6 +62,11 @@ jobs:
           path: python/diffusion-ci-outputs
           retention-days: 7
 
+      - name: Publish GT images to sglang-bot/sglang-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/diffusion-ci-outputs
+
   multimodal-diffusion-gen-2gpu:
     if: github.repository == 'sgl-project/sglang'
     runs-on: 2-gpu-h100
@@ -73,6 +84,8 @@ jobs:
         run: bash scripts/ci/cuda/ci_install_dependency.sh diffusion
 
       - name: Generate outputs
+        env:
+          RUNAI_STREAMER_MEMORY_LIMIT: 0
         run: |
           cd python
           python -m sglang.multimodal_gen.test.scripts.gen_diffusion_ci_outputs \
@@ -89,27 +102,42 @@ jobs:
           path: python/diffusion-ci-outputs
           retention-days: 7
 
-  diffusion-ci-push:
-    needs: [multimodal-diffusion-gen-1gpu, multimodal-diffusion-gen-2gpu]
+      - name: Publish GT images to sglang-bot/sglang-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/diffusion-ci-outputs
+
+  multimodal-diffusion-gen-b200:
     if: github.repository == 'sgl-project/sglang'
-    runs-on: ubuntu-latest
+    runs-on: 4-gpu-b200
+    timeout-minutes: 240
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
-
-      - name: Download artifacts
-        uses: actions/download-artifact@v4
         with:
-          pattern: diffusion-gen-*
-          path: combined
-          merge-multiple: true
+          ref: ${{ inputs.ref || github.ref }}
 
-      - name: Collect image files
+      - name: Install dependencies
+        run: bash scripts/ci/cuda/ci_install_dependency.sh diffusion
+
+      - name: Generate outputs
+        env:
+          RUNAI_STREAMER_MEMORY_LIMIT: 0
         run: |
-          mkdir -p gt_images
-          find combined \( -name "*.png" -o -name "*.jpg" -o -name "*.jpeg" -o -name "*.webp" \) -type f -exec cp -f {} gt_images/ \;
+          cd python
+          python -m sglang.multimodal_gen.test.scripts.gen_diffusion_ci_outputs \
+            --suite 1-gpu-b200 \
+            --out-dir ./diffusion-ci-outputs \
+            ${{ inputs.case_ids != '' && format('--case-ids {0}', inputs.case_ids) || '' }}
+
+      - name: Upload artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: diffusion-gen-b200
+          path: python/diffusion-ci-outputs
+          retention-days: 7
 
       - name: Publish GT images to sglang-bot/sglang-ci-data
         env:
           GITHUB_TOKEN: ${{ secrets.GH_PAT_FOR_NIGHTLY_CI_DATA }}
-        run: python scripts/ci/utils/diffusion/publish_diffusion_gt.py --source-dir gt_images
+        run: python scripts/ci/utils/diffusion/publish_diffusion_gt.py --source-dir python/diffusion-ci-outputs

python/sglang/multimodal_gen/test/scripts/gen_diffusion_ci_outputs.py

@@ -17,7 +17,12 @@ import sys
 from pathlib import Path
 
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
-from sglang.multimodal_gen.test.run_suite import SUITES, collect_test_items, run_pytest
+from sglang.multimodal_gen.test.run_suite import (
+    SUITES,
+    _maybe_pin_update_weights_model_pair,
+    collect_test_items,
+    run_pytest,
+)
 
 logger = init_logger(__name__)
 
@@ -95,6 +100,7 @@ def main():
 
     # Get files from suite (same as run_suite.py)
     suite_files_rel = SUITES[args.suite]
+    _maybe_pin_update_weights_model_pair(suite_files_rel)
     suite_files_abs = []
     for f_rel in suite_files_rel:
         f_abs = target_dir / f_rel

python/sglang/multimodal_gen/test/server/consistency_threshold.json

@@ -49,12 +49,6 @@
             "psnr_threshold": 19.0,
             "mean_abs_diff_threshold": 10.0
         },
-        "sana_image_t2i": {
-            "clip_threshold": 0.91,
-            "ssim_threshold": 0.88,
-            "psnr_threshold": 21.0,
-            "mean_abs_diff_threshold": 8.4
-        },
         "qwen_image_edit_2509_ti2i": {
             "clip_threshold": 0.92,
             "ssim_threshold": 0.65,

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

@@ -51,14 +51,6 @@ from sglang.multimodal_gen.test.test_utils import (
 logger = init_logger(__name__)
 
 
-def _is_lora_case(case: DiffusionTestCase) -> bool:
-    return bool(
-        case.server_args.lora_path
-        or case.server_args.dynamic_lora_path
-        or case.server_args.second_lora_path
-    )
-
-
 @pytest.fixture
 def diffusion_server(case: DiffusionTestCase) -> ServerContext:
     """Start a diffusion server for a single case and tear it down afterwards."""
@@ -81,11 +73,6 @@ def diffusion_server(case: DiffusionTestCase) -> ServerContext:
     sampling_params = case.sampling_params
     extra_args = os.environ.get("SGLANG_TEST_SERVE_ARGS", "")
 
-    # Keep LoRA GT on the normal backend path so adapter state matches CI.
-    if os.environ.get("SGLANG_GEN_GT", "0") == "1":
-        if not _is_lora_case(case) and "--backend" not in extra_args:
-            extra_args = "--backend diffusers " + extra_args.strip()
-
     extra_args += f" --num-gpus {server_args.num_gpus}"
 
     if server_args.tp_size is not None:
@@ -235,18 +222,21 @@ Consider updating perf_baselines.json with the snippets below:
         ctx: ServerContext,
         case_id: str,
         generate_fn: Callable[[str, openai.Client], tuple[str, bytes]],
-    ) -> tuple[RequestPerfRecord, bytes]:
-        """Run generation and collect performance records.
+        collect_perf: bool = True,
+    ) -> tuple[RequestPerfRecord | None, bytes]:
+        """Run generation and optionally collect performance records.
 
         Returns:
             Tuple of (performance_record, content_bytes)
         """
-        log_path = ctx.perf_log_path
-        log_wait_timeout = 30
-
         client = self._client(ctx)
         rid, content = generate_fn(case_id, client)
 
+        if not collect_perf:
+            return None, content
+
+        log_path = ctx.perf_log_path
+        log_wait_timeout = 30
         req_perf_record = wait_for_req_perf_record(
             rid,
             log_path,
@@ -1024,6 +1014,7 @@ Repository: https://github.com/sglang-bot/sglang-ci-data (path: diffusion-ci/con
             diffusion_server,
             case.id,
             generate_fn,
+            collect_perf=not is_gt_gen_mode,
         )
 
         if is_gt_gen_mode:

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

@@ -500,15 +500,6 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
         run_lora_dynamic_switch_check=True,
         run_multi_lora_api_check=True,
     ),
-    DiffusionTestCase(
-        "sana_image_t2i",
-        DiffusionServerArgs(
-            model_path="Efficient-Large-Model/Sana_600M_1024px_diffusers",
-            modality="image",
-        ),
-        T2I_sampling_params,
-        run_perf_check=False,
-    ),
     # === Text and Image to Image (TI2I) ===
     DiffusionTestCase(
         "qwen_image_edit_ti2i",
@@ -804,7 +795,6 @@ ONE_GPU_CASES_C = [
             modality="image",
         ),
         T2I_sampling_params,
-        run_consistency_check=False,
     )
 ]
 
@@ -945,7 +935,6 @@ TWO_GPU_CASES_A = [
             extras=["--pipeline-class-name LTX2TwoStagePipeline"],
         ),
         T2V_sampling_params,
-        run_consistency_check=False,
     ),
 ]