[codex] Split diffusion quant CI coverage (#24599)

Co-authored-by: Codex <codex@example.com>

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

@@ -405,11 +405,14 @@ if not current_platform.is_hip():
         )
     )
 # Skip all ModelOpt tests on AMD: FP8 requires torch._scaled_mm (HIPBLAS_STATUS_NOT_SUPPORTED
-# on ROCm), NVFP4 requires flashinfer or sgl_kernel FP4 kernels (CUDA-only)
+# on ROCm), NVFP4 requires flashinfer or sgl_kernel FP4 kernels (CUDA-only).
+# Run FP8 cases on the regular H100 1-GPU CI shard and keep only B200-only
+# quantization coverage in the B200 suite.
 if current_platform.is_hip():
-    ONE_GPU_MODELOPT_CASES = []
+    ONE_GPU_MODELOPT_FP8_CASES = []
+    ONE_GPU_MODELOPT_NVFP4_CASES = []
 else:
-    ONE_GPU_MODELOPT_CASES = [
+    ONE_GPU_MODELOPT_FP8_CASES = [
         _make_modelopt_ci_case(
             "flux1_modelopt_fp8_t2i",
             model_path=DEFAULT_FLUX_1_DEV_MODEL_NAME_FOR_TEST,
@@ -457,6 +460,8 @@ else:
             sampling_params=MODELOPT_TI2I_CI_sampling_params,
             extras=["--transformer-path", MODELOPT_QWEN_IMAGE_EDIT_FP8_TRANSFORMER],
         ),
+    ]
+    ONE_GPU_MODELOPT_NVFP4_CASES = [
         _make_modelopt_ci_case(
             "flux1_modelopt_nvfp4_t2i",
             model_path=DEFAULT_FLUX_1_DEV_MODEL_NAME_FOR_TEST,
@@ -483,6 +488,8 @@ else:
         ),
     ]
 
+ONE_GPU_B200_CASES = ONE_GPU_MODELOPT_NVFP4_CASES
+
 TWO_GPU_CASES = [
     DiffusionTestCase(
         "wan2_2_i2v_a14b_2gpu",
@@ -727,5 +734,5 @@ if not current_platform.is_hip():
         )
     )
 
-ONE_GPU_CASES += ONE_GPU_MODELOPT_CASES
+ONE_GPU_CASES += ONE_GPU_MODELOPT_FP8_CASES
 TWO_GPU_CASES = _with_default_num_gpus(TWO_GPU_CASES, 2)

python/sglang/multimodal_gen/test/server/test_server_b200.py

@@ -7,7 +7,7 @@ from __future__ import annotations
 import pytest
 
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
-from sglang.multimodal_gen.test.server.gpu_cases import ONE_GPU_MODELOPT_CASES
+from sglang.multimodal_gen.test.server.gpu_cases import ONE_GPU_B200_CASES
 from sglang.multimodal_gen.test.server.test_server_common import (  # noqa: F401
     DiffusionServerBase,
     diffusion_server,
@@ -18,9 +18,9 @@ logger = init_logger(__name__)
 
 
 class TestDiffusionServerOneGpuB200(DiffusionServerBase):
-    """B200-targeted CI tests for 1-GPU ModelOpt diffusion cases."""
+    """B200-targeted CI tests for 1-GPU Blackwell-only diffusion cases."""
 
-    @pytest.fixture(params=ONE_GPU_MODELOPT_CASES, ids=lambda c: c.id)
+    @pytest.fixture(params=ONE_GPU_B200_CASES, ids=lambda c: c.id)
     def case(self, request) -> DiffusionTestCase:
         """Provide a DiffusionTestCase for each 1-GPU B200 test."""
         return request.param

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

@@ -5,12 +5,12 @@ Usage:
 
 pytest python/sglang/multimodal_gen/test/server/test_server_1_gpu.py
 # for a single testcase, look for the name of the testcase in ONE_GPU_CASES,
-# ONE_GPU_MODELOPT_CASES, or TWO_GPU_CASES
+# ONE_GPU_MODELOPT_FP8_CASES, ONE_GPU_B200_CASES, or TWO_GPU_CASES
 pytest python/sglang/multimodal_gen/test/server/test_server_1_gpu.py -k qwen_image_t2i
 
 
 To add a new testcase:
-1. add your testcase with case-id: `my_new_test_case_id` to `ONE_GPU_CASES`, `ONE_GPU_MODELOPT_CASES`, or `TWO_GPU_CASES`
+1. add your testcase with case-id: `my_new_test_case_id` to `ONE_GPU_CASES`, `ONE_GPU_MODELOPT_FP8_CASES`, `ONE_GPU_B200_CASES`, or `TWO_GPU_CASES`
 2. run `SGLANG_GEN_BASELINE=1 pytest -s python/sglang/multimodal_gen/test/server/ -k my_new_test_case_id`
 3. insert or override the corresponding scenario in `scenarios` section of perf_baselines.json with the output baseline of step-2