[diffusion] chore: auto-enable best parallel setting if unspecified (#22763)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

python/sglang/multimodal_gen/runtime/server_args.py

@@ -130,8 +130,8 @@ class ServerArgs:
     dp_size: int = 1
     # number of gpu in a dp group
     dp_degree: int = 1
-    # cfg parallel
-    enable_cfg_parallel: bool = False
+    # cfg parallel (None = auto-decide based on num_gpus)
+    enable_cfg_parallel: Optional[bool] = None
 
     hsdp_replicate_dim: int = 1
     hsdp_shard_dim: Optional[int] = None
@@ -428,6 +428,7 @@ class ServerArgs:
         sp_unspecified = self.sp_degree is None
         ulysses_unspecified = self.ulysses_degree is None
         ring_unspecified = self.ring_degree is None
+        cfg_unspecified = self.enable_cfg_parallel is None
 
         if self.hsdp_shard_dim is None:
             self.hsdp_shard_dim = self.num_gpus
@@ -435,6 +436,30 @@ class ServerArgs:
         if self.tp_size is None:
             self.tp_size = 1
 
+        # Auto-enable CFG parallel when user hasn't set any parallelism flags
+        # and there are enough GPUs.  Only auto-enable for models whose default
+        # SamplingParams use classifier-free guidance (negative_prompt is not None),
+        # because non-CFG models (e.g. FLUX) crash when CFG parallel splits ranks.
+        if cfg_unspecified:
+            cfg_group_size = self.dp_size * self.tp_size * 2
+            if (
+                self.num_gpus >= 2
+                and self.num_gpus % cfg_group_size == 0
+                and sp_unspecified
+                and ulysses_unspecified
+                and ring_unspecified
+                and self._model_default_uses_cfg()
+            ):
+                self.enable_cfg_parallel = True
+                logger.info(
+                    "Automatically enabled CFG parallel for %d GPUs. "
+                    "Use --sp-degree / --ulysses-degree to use sequence "
+                    "parallelism instead.",
+                    self.num_gpus,
+                )
+            else:
+                self.enable_cfg_parallel = False
+
         # adjust sp_degree: allocate all remaining GPUs after TP and DP
         if self.sp_degree is None:
             num_gpus_per_group = self.dp_size * self.tp_size
@@ -466,6 +491,28 @@ class ServerArgs:
             self.ring_degree = 1
             logger.debug(f"Ring degree not set, using default value {self.ring_degree}")
 
+    def _model_default_uses_cfg(self) -> bool:
+        """
+        Check whether the model uses classifier-free guidance by default.
+
+        CFG is active when *both* ``negative_prompt is not None`` and ``guidance_scale > 1``.
+        """
+        from sglang.multimodal_gen.registry import get_model_info
+
+        model_info = get_model_info(self.model_path, self.backend, self.model_id)
+        if model_info is None:
+            return False
+        default_params = model_info.sampling_param_cls()
+
+        # for ltx2.3, cfg-parallel performs worse than ulysses-sp
+        is_ltx = "ltx" in type(default_params).__name__.lower()
+        if is_ltx:
+            return False
+        return (
+            getattr(default_params, "negative_prompt", None) is not None
+            and getattr(default_params, "guidance_scale", 0) > 1.0
+        )
+
     @staticmethod
     def _is_ltx23_model_path(model_path: str | None) -> bool:
         if not model_path:
@@ -679,8 +726,8 @@ class ServerArgs:
         parser.add_argument(
             "--enable-cfg-parallel",
             action="store_true",
-            default=ServerArgs.enable_cfg_parallel,
-            help="Enable cfg parallel.",
+            default=None,
+            help="Enable cfg parallel. Auto-enabled when num_gpus >= 2 and no SP flags are set.",
         )
         parser.add_argument(
             "--data-parallel-size",

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

@@ -1655,7 +1655,8 @@
             },
             "expected_e2e_ms": 149864.99,
             "expected_avg_denoise_ms": 3608.89,
-            "expected_median_denoise_ms": 2400.38
+            "expected_median_denoise_ms": 2400.38,
+            "estimated_full_test_time_s": 253.9
         },
         "wan2_1_t2v_14b_2gpu": {
             "stages_ms": {

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

@@ -154,7 +154,9 @@ def diffusion_server(case: DiffusionTestCase) -> ServerContext:
         # pipeline class.  This avoids hard failures when a model needs a
         # newer diffusers release than what is currently installed in CI.
         msg = str(exc)
-        if "not found in diffusers" in msg or "has no attribute" in msg:
+        if "not found in diffusers" in msg or (
+            "has no attribute" in msg and "diffusers" in msg.lower()
+        ):
             pytest.skip(
                 f"Skipping {case.id}: required diffusers pipeline class "
                 f"is not available in the installed version. "

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

@@ -808,6 +808,7 @@ TWO_GPU_CASES_A = [
             modality="video",
             custom_validator="video",
             num_gpus=2,
+            extras=["--ulysses-degree=2"],
         ),
         T2V_sampling_params,
     ),
@@ -821,6 +822,7 @@ TWO_GPU_CASES_A = [
             modality="video",
             custom_validator="video",
             num_gpus=2,
+            extras=["--ulysses-degree=2"],
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -912,7 +914,7 @@ TWO_GPU_CASES_A = [
             model_path="Lightricks/LTX-2",
             modality="video",
             num_gpus=2,
-            extras=["--pipeline-class-name LTX2TwoStagePipeline"],
+            extras=["--pipeline-class-name LTX2TwoStagePipeline", "--ulysses-degree=2"],
         ),
         T2V_sampling_params,
     ),
@@ -936,6 +938,7 @@ TWO_GPU_CASES_B = [
             modality="video",
             custom_validator="video",
             num_gpus=2,
+            extras=["--ulysses-degree=2"],
         ),
         TI2V_sampling_params,
     ),
@@ -958,6 +961,7 @@ TWO_GPU_CASES_B = [
             custom_validator="video",
             num_gpus=2,
             lora_path="starsfriday/Wan2.1-Divine-Power-LoRA",
+            extras=["--ulysses-degree=2"],
         ),
         TI2V_sampling_params,
         run_lora_basic_api_check=True,
@@ -969,6 +973,7 @@ TWO_GPU_CASES_B = [
             modality="video",
             custom_validator="video",
             num_gpus=2,
+            extras=["--ulysses-degree=2"],
         ),
         TI2V_sampling_params,
     ),