Revert "[codex] Update Wan2.2 ModelOpt CI checkpoints" (#25845)

docs_new/docs/sglang-diffusion/quantization.mdx

@@ -110,15 +110,14 @@ backend.
 ## Validated ModelOpt Checkpoints
 
 This section is the canonical support matrix for the nine diffusion ModelOpt
-checkpoints currently wired up in SGLang docs and validation coverage.
+checkpoints currently wired up in SGLang docs and B200 CI coverage.
 
 Published checkpoints keep the serialized quantization config as
 `quant_method=modelopt`; the FP8 vs NVFP4 split below is a documentation label
 derived from `quant_algo`.
 
-Six of the nine repos live under `lmsys/*`. The Wan2.2 entries use NVIDIA's
+Eight of the nine repos live under `lmsys/*`. The FLUX.2 NVFP4 entry keeps the
-official full Diffusers repos, and the FLUX.2 NVFP4 entry keeps the official
+official `black-forest-labs/FLUX.2-dev-NVFP4` repo.
-`black-forest-labs/FLUX.2-dev-NVFP4` repo.
 
 <table style={{width: "100%", borderCollapse: "collapse", tableLayout: "fixed"}}>
   <colgroup>
@@ -159,10 +158,10 @@ official full Diffusers repos, and the FLUX.2 NVFP4 entry keeps the official
     <tr>
       <td><code>FP8</code></td>
       <td><code>Wan-AI/Wan2.2-T2V-A14B-Diffusers</code></td>
-      <td><code>--model-path</code></td>
+      <td><code>--transformer-path</code></td>
-      <td><code>nvidia/Wan2.2-T2V-A14B-Diffusers-FP8</code></td>
+      <td><code>lmsys/wan22-t2v-a14b-modelopt-fp8-sglang-transformer</code></td>
-      <td>full Diffusers repo with ModelOpt FP8 Wan2.2 components</td>
+      <td>primary <code>transformer</code> quantized, <code>transformer_2</code> kept BF16</td>
-      <td>validated through direct <code>--model-path</code> loading</td>
+      <td>primary-transformer-only path; keep <code>transformer_2</code> on the base checkpoint, and do not describe this as dual-transformer full-model FP8 unless that path is validated separately</td>
     </tr>
     <tr>
       <td><code>FP8</code></td>
@@ -207,25 +206,24 @@ official full Diffusers repos, and the FLUX.2 NVFP4 entry keeps the official
     <tr>
       <td><code>NVFP4</code></td>
       <td><code>Wan-AI/Wan2.2-T2V-A14B-Diffusers</code></td>
-      <td><code>--model-path</code></td>
+      <td><code>--transformer-path</code></td>
-      <td><code>nvidia/Wan2.2-T2V-A14B-Diffusers-NVFP4</code></td>
+      <td><code>lmsys/wan22-t2v-a14b-modelopt-nvfp4-sglang-transformer</code></td>
-      <td>full Diffusers repo with ModelOpt NVFP4 Wan2.2 components</td>
+      <td>primary <code>transformer</code> quantized with ModelOpt NVFP4, <code>transformer_2</code> kept BF16</td>
-      <td>current B200/Blackwell bring-up uses <code>SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND=trtllm</code></td>
+      <td>primary-transformer-only path; keep <code>transformer_2</code> on the base checkpoint, and current B200/Blackwell bring-up uses <code>SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND=cudnn</code></td>
     </tr>
   </tbody>
 </table>
 
-The FP8 rows run in the regular H100 1-GPU diffusion CI shard; the NVFP4 rows
+These nine checkpoints are also the intended case set for the B200 diffusion CI
-run in the B200 diffusion CI shard (`multimodal-gen-test-1-b200`).
+job (`multimodal-gen-test-1-b200`).
 
 ## ModelOpt FP8
 
 ### Usage Examples
 
-Converted ModelOpt FP8 transformer repos should be loaded as transformer
+Converted ModelOpt FP8 checkpoints should be loaded as transformer component
-component overrides. If the repo or local directory already contains
+overrides. If the repo or local directory already contains `config.json`, use
-`config.json`, use `--transformer-path`. Full Diffusers repos such as the
+`--transformer-path`.
-NVIDIA Wan2.2 FP8 checkpoint can be passed directly with `--model-path`.
 
 ```bash
 sglang generate \
@@ -237,7 +235,8 @@ sglang generate \
 
 ```bash
 sglang generate \
-  --model-path nvidia/Wan2.2-T2V-A14B-Diffusers-FP8 \
+  --model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --transformer-path lmsys/wan22-t2v-a14b-modelopt-fp8-sglang-transformer \
   --prompt "a fox walking through neon rain" \
   --save-output
 ```
@@ -324,12 +323,14 @@ sglang generate \
   --save-output
 ```
 
-For Wan2.2 NVFP4:
+For a dual-transformer Wan2.2 export where only the primary `transformer`
+was quantized:
 
 ```bash
-SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND=trtllm \
+SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND=cudnn \
 sglang generate \
-  --model-path nvidia/Wan2.2-T2V-A14B-Diffusers-NVFP4 \
+  --model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --transformer-path lmsys/wan22-t2v-a14b-modelopt-nvfp4-sglang-transformer \
   --prompt "a fox walking through neon rain" \
   --save-output
 ```
@@ -340,16 +341,17 @@ sglang generate \
   directories that already include `config.json`.
 - Use `--transformer-weights-path` for raw NVFP4 exports, individual
   safetensors files, or repo layouts that should be treated as weights first.
-- For legacy mixed Wan2.2 transformer overrides, the primary
+- For dual-transformer pipelines such as `Wan2.2-T2V-A14B-Diffusers`, the
-  `--transformer-path` override targets only `transformer`. Use a per-component
+  primary `--transformer-path` override targets only `transformer`. Use a
-  override such as `--transformer-2-path` only when you intentionally want a
+  per-component override such as `--transformer-2-path` only when you
-  non-default `transformer_2`.
+  intentionally want a non-default `transformer_2`.
 - On Blackwell, the validated Wan2.2 ModelOpt NVFP4 path currently prefers
   FlashInfer FP4 GEMM via
-  `SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND=trtllm`.
+  `SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND=cudnn`.
-- This environment-variable override selects the validated Wan2.2 NVFP4
+- This environment-variable override is a current workaround for NVFP4 cases
-  full-repo path on Blackwell while the other NVFP4 CI cases continue to use
+  where the default sglang JIT/CUTLASS `sm100` path rejects a large-M shape at
-  the generic `cudnn` backend.
+  `can_implement()`. The intended long-term fix is to add a validated CUTLASS
+  fallback for those shapes rather than rely on the override.
 - Direct `--model-path` loading is a compatibility path for FLUX.2 NVFP4-style
   repos or local directories.
 - If `--transformer-weights-path` is provided explicitly, it takes precedence

python/sglang/multimodal_gen/registry.py

@@ -739,10 +739,7 @@ def _register_configs():
     register_configs(
         sampling_param_cls=Wan2_2_T2V_A14B_SamplingParam,
         pipeline_config_cls=Wan2_2_T2V_A14B_Config,
-        hf_model_paths=[
+        hf_model_paths=["Wan-AI/Wan2.2-T2V-A14B-Diffusers"],
-            "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
-            "nvidia/Wan2.2-T2V-A14B-Diffusers-NVFP4",
-        ],
     )
     register_configs(
         sampling_param_cls=Wan2_2_I2V_A14B_SamplingParam,

python/sglang/multimodal_gen/runtime/layers/quantization/modelopt_quant.py

@@ -202,7 +202,7 @@ class ModelOptFp4Config(ModelOptQuantConfig):
         exclude_modules: List[str] = None,
         packed_modules_mapping: Optional[Dict[str, List[str]]] = None,
         checkpoint_uses_packed_qkv: bool = False,
-        swap_weight_nibbles: bool = False,
+        swap_weight_nibbles: bool = True,
     ) -> None:
         super().__init__(exclude_modules, packed_modules_mapping)
         self.is_checkpoint_nvfp4_serialized = is_checkpoint_nvfp4_serialized
@@ -261,7 +261,7 @@ class ModelOptFp4Config(ModelOptQuantConfig):
     def from_config(cls, config: Dict[str, Any]) -> ModelOptFp4Config:
         group_size = None
         exclude_modules = []
-        swap_weight_nibbles = False
+        swap_weight_nibbles = True
 
         # Flat format (config.json quantization_config)
         quant_method = config.get("quant_algo")
@@ -273,10 +273,7 @@ class ModelOptFp4Config(ModelOptQuantConfig):
                     first_group = next(iter(config_groups.values()), {})
                     group_size = first_group.get("weights", {}).get("group_size")
             exclude_modules = config.get("ignore", [])
-            swap_weight_nibbles = config.get(
+            swap_weight_nibbles = config.get("swap_weight_nibbles", True)
-                "swap_weight_nibbles",
-                config.get("checkpoint_uses_packed_qkv", False),
-            )
         else:
             # Nested format (hf_quant_config.json)
             try:
@@ -286,10 +283,7 @@ class ModelOptFp4Config(ModelOptQuantConfig):
                 exclude_modules = quant_config.get("exclude_modules", [])
                 swap_weight_nibbles = quant_config.get(
                     "swap_weight_nibbles",
-                    config.get(
+                    config.get("swap_weight_nibbles", True),
-                        "swap_weight_nibbles",
-                        config.get("checkpoint_uses_packed_qkv", False),
-                    ),
                 )
             except (ValueError, KeyError):
                 raise ValueError("Cannot find 'quant_algo' in quantization config.")
@@ -500,9 +494,7 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
         w = layer.weight.data
         w_swapped = _prepare_nvfp4_weight_bytes(
             w,
-            swap_weight_nibbles=getattr(
+            swap_weight_nibbles=getattr(self.quant_config, "swap_weight_nibbles", True),
-                self.quant_config, "swap_weight_nibbles", False
-            ),
         )
 
         _, flashinfer_backend = _get_fp4_gemm_op()
@@ -562,13 +554,8 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
         padded_scales[:B, :M, :K] = scales
 
         _, flashinfer_backend = _get_fp4_gemm_op()
-        uses_flux1_scale_layout = not getattr(
+        if flashinfer_backend is None:
-            self.quant_config, "checkpoint_uses_packed_qkv", False
+            # CUTLASS (sgl_kernel) path: blockwise interleave to TMA layout
-        ) and getattr(layer, "prefix", "").startswith(
-            ("transformer_blocks.", "single_transformer_blocks.")
-        )
-        if flashinfer_backend is None or uses_flux1_scale_layout:
-            # CUTLASS and FLUX.1 CUDNN paths need the TMA scale layout.
             padded_scales = padded_scales.reshape(
                 B, M_padded // 128, 4, 32, K_padded // 4, 4
             )

python/sglang/multimodal_gen/runtime/loader/transformer_load_utils.py

@@ -88,12 +88,12 @@ def _merge_modelopt_fp4_configs(
     inferred_config.packed_modules_mapping = getattr(
         existing_config, "packed_modules_mapping", {}
     )
+    inferred_config.swap_weight_nibbles = getattr(
+        existing_config, "swap_weight_nibbles", True
+    )
     inferred_config.checkpoint_uses_packed_qkv = getattr(
         inferred_config, "checkpoint_uses_packed_qkv", False
     ) or getattr(existing_config, "checkpoint_uses_packed_qkv", False)
-    inferred_config.swap_weight_nibbles = getattr(
-        inferred_config, "swap_weight_nibbles", False
-    ) or getattr(existing_config, "swap_weight_nibbles", False)
     if getattr(inferred_config, "group_size", None) is None:
         inferred_config.group_size = getattr(existing_config, "group_size", None)
 

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

@@ -13,12 +13,6 @@
             "psnr_threshold": 24.0,
             "mean_abs_diff_threshold": 8.0
         },
-        "flux1_modelopt_fp8_t2i": {
-            "clip_threshold": 0.92,
-            "ssim_threshold": 0.94,
-            "psnr_threshold": 28.0,
-            "mean_abs_diff_threshold": 8.0
-        },
         "flux_2_klein_image_t2i": {
             "clip_threshold": 0.94,
             "ssim_threshold": 0.78,

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

@@ -11,9 +11,8 @@ from sglang.multimodal_gen.test.server.testcase_configs import (
     MODELOPT_NVFP4_B200_ENV_VARS,
     MODELOPT_QWEN_IMAGE_EDIT_FP8_TRANSFORMER,
     MODELOPT_QWEN_IMAGE_FP8_TRANSFORMER,
-    MODELOPT_WAN22_FP8_MODEL,
+    MODELOPT_WAN22_FP8_TRANSFORMER,
-    MODELOPT_WAN22_NVFP4_B200_ENV_VARS,
+    MODELOPT_WAN22_NVFP4_TRANSFORMER,
-    MODELOPT_WAN22_NVFP4_MODEL,
     T2V_PROMPT,
     DiffusionSamplingParams,
     DiffusionServerArgs,
@@ -420,7 +419,6 @@ else:
             modality="image",
             sampling_params=MODELOPT_T2I_CI_sampling_params,
             extras=["--transformer-path", MODELOPT_FLUX1_FP8_TRANSFORMER],
-            run_consistency_check=True,
         ),
         _make_modelopt_ci_case(
             "flux2_modelopt_fp8_t2i",
@@ -428,15 +426,13 @@ else:
             modality="image",
             sampling_params=MODELOPT_T2I_CI_sampling_params,
             extras=["--transformer-path", MODELOPT_FLUX2_FP8_TRANSFORMER],
-            run_consistency_check=True,
         ),
         _make_modelopt_ci_case(
             "wan22_modelopt_fp8_t2v",
-            model_path=MODELOPT_WAN22_FP8_MODEL,
+            model_path=DEFAULT_WAN_2_2_T2V_A14B_MODEL_NAME_FOR_TEST,
             modality="video",
             sampling_params=MODELOPT_T2V_CI_sampling_params,
-            extras=[],
+            extras=["--transformer-path", MODELOPT_WAN22_FP8_TRANSFORMER],
-            run_consistency_check=True,
         ),
         _make_modelopt_ci_case(
             "hunyuanvideo_modelopt_fp8_t2v",
@@ -473,7 +469,6 @@ else:
             sampling_params=MODELOPT_T2I_CI_sampling_params,
             extras=["--transformer-path", MODELOPT_FLUX1_NVFP4_TRANSFORMER],
             env_vars=MODELOPT_NVFP4_B200_ENV_VARS,
-            run_consistency_check=True,
         ),
         _make_modelopt_ci_case(
             "flux2_modelopt_nvfp4_t2i",
@@ -482,16 +477,14 @@ else:
             sampling_params=MODELOPT_T2I_CI_sampling_params,
             extras=["--transformer-weights-path", MODELOPT_FLUX2_NVFP4_WEIGHTS],
             env_vars=MODELOPT_NVFP4_B200_ENV_VARS,
-            run_consistency_check=True,
         ),
         _make_modelopt_ci_case(
             "wan22_modelopt_nvfp4_t2v",
-            model_path=MODELOPT_WAN22_NVFP4_MODEL,
+            model_path=DEFAULT_WAN_2_2_T2V_A14B_MODEL_NAME_FOR_TEST,
             modality="video",
             sampling_params=MODELOPT_T2V_CI_sampling_params,
-            extras=[],
+            extras=["--transformer-path", MODELOPT_WAN22_NVFP4_TRANSFORMER],
-            env_vars=MODELOPT_WAN22_NVFP4_B200_ENV_VARS,
+            env_vars=MODELOPT_NVFP4_B200_ENV_VARS,
-            run_consistency_check=True,
         ),
     ]
 

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

@@ -441,7 +441,7 @@ HUNYUAN3D_SHAPE_sampling_params = DiffusionSamplingParams(
 
 MODELOPT_FLUX1_FP8_TRANSFORMER = "lmsys/flux1-dev-modelopt-fp8-sglang-transformer"
 MODELOPT_FLUX2_FP8_TRANSFORMER = "lmsys/flux2-dev-modelopt-fp8-sglang-transformer"
-MODELOPT_WAN22_FP8_MODEL = "nvidia/Wan2.2-T2V-A14B-Diffusers-FP8"
+MODELOPT_WAN22_FP8_TRANSFORMER = "lmsys/wan22-t2v-a14b-modelopt-fp8-sglang-transformer"
 MODELOPT_HUNYUANVIDEO_FP8_TRANSFORMER = (
     "lmsys/hunyuanvideo-modelopt-fp8-sglang-transformer"
 )
@@ -451,11 +451,10 @@ MODELOPT_QWEN_IMAGE_EDIT_FP8_TRANSFORMER = (
 )
 MODELOPT_FLUX1_NVFP4_TRANSFORMER = "lmsys/flux1-dev-modelopt-nvfp4-sglang-transformer"
 MODELOPT_FLUX2_NVFP4_WEIGHTS = "black-forest-labs/FLUX.2-dev-NVFP4"
-MODELOPT_WAN22_NVFP4_MODEL = "nvidia/Wan2.2-T2V-A14B-Diffusers-NVFP4"
+MODELOPT_WAN22_NVFP4_TRANSFORMER = (
+    "lmsys/wan22-t2v-a14b-modelopt-nvfp4-sglang-transformer"
+)
 MODELOPT_NVFP4_B200_ENV_VARS = {"SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND": "cudnn"}
-MODELOPT_WAN22_NVFP4_B200_ENV_VARS = {
-    "SGLANG_DIFFUSION_FLASHINFER_FP4_GEMM_BACKEND": "trtllm"
-}
 
 
 def _make_modelopt_ci_case(
@@ -466,7 +465,6 @@ def _make_modelopt_ci_case(
     sampling_params: DiffusionSamplingParams,
     extras: list[str],
     env_vars: dict[str, str] | None = None,
-    run_consistency_check: bool = False,
 ) -> DiffusionTestCase:
     return DiffusionTestCase(
         case_id,
@@ -479,7 +477,7 @@ def _make_modelopt_ci_case(
         ),
         sampling_params,
         run_perf_check=False,
-        run_consistency_check=run_consistency_check,
+        run_consistency_check=False,
         run_component_accuracy_check=False,
     )
 

python/sglang/multimodal_gen/test/test_utils.py

@@ -33,7 +33,7 @@ if TYPE_CHECKING:
 
 logger = init_logger(__name__)
 
-SGL_TEST_FILES_CI_DATA_REVISION = "94eab4fcca6d4ddc77cdb3622f13033b61e81002"
+SGL_TEST_FILES_CI_DATA_REVISION = "c8305f1dd8cc82197f36c17d0f503adc94016cc7"
 SGL_TEST_FILES_CONSISTENCY_GT_ROOT = (
     "https://raw.githubusercontent.com/"
     f"sgl-project/ci-data/{SGL_TEST_FILES_CI_DATA_REVISION}/"

python/sglang/multimodal_gen/tools/build_modelopt_nvfp4_transformer.py

@@ -218,7 +218,9 @@ def build_modelopt_nvfp4_transformer(
         patterns.extend(keep_bf16_patterns)
 
     resolved_swap_weight_nibbles = (
-        swap_weight_nibbles if swap_weight_nibbles is not None else False
+        swap_weight_nibbles
+        if swap_weight_nibbles is not None
+        else (False if pattern_preset == "flux1-nvfp4" else True)
     )
     output_config = _updated_quant_config(
         _load_config(source_dir),
@@ -371,7 +373,7 @@ def _parse_args() -> argparse.Namespace:
         default=None,
         help=(
             "Whether the runtime should swap packed FP4 nibbles before padding. "
-            "Defaults to false."
+            "Defaults to false for --pattern-preset flux1-nvfp4 and true otherwise."
         ),
     )
     parser.add_argument(