[diffusion] chore: refresh docs, retire stale knobs, and fix nightly attribution (#34663)

python/sglang/multimodal_gen/README.md

@@ -9,7 +9,7 @@ SGLang diffusion features an end-to-end unified pipeline for accelerating diffus
 ## Key Features
 
 SGLang Diffusion has the following features:
-  - Broad model support: Wan, FastWan, FLUX, Qwen-Image, Z-Image, Ideogram 4, Krea-2, Cosmos3, LTX-2/LTX-2.3, MiniMax-H3, LingBot World, SANA-WM, JoyEcho, MOVA, GLM-Image, ERNIE-Image, Hunyuan3D, and more
+  - Broad model support: Wan, FastWan, FLUX, Qwen-Image, Z-Image, Ideogram 4, Krea-2, Cosmos3, LTX-2/LTX-2.3, MiniMax-H3, LingBot Video MoE, LingBot World, SANA-Video/SANA-WM, JoyEcho, MOVA, GLM-Image, ERNIE-Image, Hunyuan3D, and more
   - Fast inference speed: empowered by optimized `sgl-kernel` kernels, scheduler/runtime improvements, caching acceleration, and native diffusion hot-path optimizations
   - Ease of use: OpenAI-compatible api, CLI, and python sdk support
   - Multi-platform support:

python/sglang/multimodal_gen/benchmarks/bench_serving.py

@@ -728,12 +728,6 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser(
         description="Benchmark serving for diffusion models."
     )
-    parser.add_argument(
-        "--backend",
-        type=str,
-        default=None,
-        help="DEPRECATED: --task is deprecated and will be ignored. The task will be inferred from --model.",
-    )
     parser.add_argument(
         "--base-url",
         type=str,

python/sglang/multimodal_gen/runtime/entrypoints/utils.py

@@ -639,11 +639,12 @@ def _try_save_cuda_video_direct(
                     stderr=stderr_file.read(),
                 )
         return True
-    except Exception as e:
-        logger.warning(
-            "Direct CUDA video save failed; falling back to imageio: %s",
-            str(e),
+    except Exception:
+        logger.warning_once(
+            "Direct CUDA video save failed; falling back to imageio. "
+            "Enable debug logging for exception details."
         )
+        logger.debug("Direct CUDA video save failure", exc_info=True)
         return False
     finally:
         if tmp_wav_path:
@@ -725,11 +726,12 @@ def _try_save_cuda_videos_direct(
     try:
         with ThreadPoolExecutor(max_workers=_MAX_PARALLEL_CUDA_VIDEO_SAVES) as pool:
             return list(pool.map(save_one, range(len(samples))))
-    except Exception as exc:
-        logger.warning(
-            "Parallel CUDA video save failed; falling back to serial output: %s",
-            str(exc),
+    except Exception:
+        logger.warning_once(
+            "Parallel CUDA video save failed; falling back to serial output. "
+            "Enable debug logging for exception details."
         )
+        logger.debug("Parallel CUDA video save failure", exc_info=True)
         return None
 
 

python/sglang/multimodal_gen/runtime/platforms/cuda.py

@@ -151,8 +151,7 @@ class _SageAttentionBackendResolver(_CudaAttentionBackendResolver):
     def resolve(cls, platform) -> str | AttentionBackendEnum:
         try:
             from sageattention import sageattn  # noqa: F401
-        except ImportError as e:
-            logger.info(e)
+        except ImportError:
             logger.info(
                 "Sage Attention backend is not installed (To install it, run `pip install git+https://github.com/thu-ml/SageAttention.git@d9704247a5139ab4c03bf7fc6b35cc0e2cbb5ea4 --no-build-isolation`). Falling back to Flash Attention."
             )
@@ -176,8 +175,7 @@ class _SageAttentionBackendResolver(_CudaAttentionBackendResolver):
             )
 
             return "sglang.multimodal_gen.runtime.layers.attention.backends.sage_attn.SageAttentionBackend"
-        except ImportError as e:
-            logger.info(e)
+        except ImportError:
             logger.info(
                 "Sage Attention backend failed to import. Falling back to Flash Attention."
             )
@@ -195,8 +193,7 @@ class _SageAttention3BackendResolver(_CudaAttentionBackendResolver):
             )
 
             return "sglang.multimodal_gen.runtime.layers.attention.backends.sage_attn3.SageAttention3Backend"
-        except ImportError as e:
-            logger.info(e)
+        except ImportError:
             logger.info(
                 "Sage Attention 3 backend is not installed (To install it, see https://github.com/thu-ml/SageAttention/tree/main/sageattention3_blackwell#installation). Falling back to Torch SDPA."
             )

python/sglang/multimodal_gen/runtime/server_args/server_args.py

@@ -86,7 +86,7 @@ from sglang.multimodal_gen.utils import (
 logger = init_logger(__name__)
 
 LTX2_TWO_STAGE_DEVICE_MODES = ("original", "resident")
-LTX2_TWO_STAGE_DEVICE_MODE_CHOICES = (*LTX2_TWO_STAGE_DEVICE_MODES, "snapshot")
+LTX2_TWO_STAGE_DEVICE_MODE_CHOICES = LTX2_TWO_STAGE_DEVICE_MODES
 LTX2_TWO_STAGE_PIPELINE_NAMES = ("LTX2TwoStagePipeline", "LTX2TwoStageHQPipeline")
 # H200-class GPUs (>=130 GiB total) can usually keep both LTX2 DiTs resident.
 LTX2_RESIDENT_AUTO_ENABLE_MEM_GB = 130
@@ -100,15 +100,7 @@ RING_CAPABLE_ATTENTION_BACKENDS = ("fa", "sage_attn")
 def _normalize_ltx2_two_stage_device_mode(mode: str | None) -> str | None:
     if mode is None:
         return None
-    mode = mode.lower()
-    if mode == "snapshot":
-        logger.warning(
-            "ltx2_two_stage_device_mode=snapshot is deprecated and is treated "
-            "as original. Please use ltx2_two_stage_device_mode=original or "
-            "resident instead. This alias may be removed after two release cycles."
-        )
-        return "original"
-    return mode
+    return mode.lower()
 
 
 def is_ltx2_two_stage_pipeline_name(pipeline_class_name: str | None) -> bool:
@@ -2211,8 +2203,6 @@ class ServerArgs(DisaggServerArgsMixin):
                 "LTX-2.3 two-stage device residency mode: "
                 "'original' keeps official two-stage semantics without premerged stage2, "
                 "'resident' keeps both transformers resident on GPU. "
-                "'snapshot' is deprecated, treated as 'original', and may be "
-                "removed after two release cycles. "
                 "Default is auto: resident on H200/high-memory CUDA GPUs, otherwise original."
             ),
         )

python/sglang/multimodal_gen/test/unit/test_server_args.py

@@ -2001,24 +2001,19 @@ class TestOffloadDefaults(unittest.TestCase):
             ["text_encoder", "image_encoder", "vae"],
         )
 
-    def test_ltx23_snapshot_device_mode_is_deprecated_alias_for_original(self):
-        args = self._from_dict_with_pipeline_config(
-            LTX2PipelineConfig(),
-            memory_gb=140,
-            available_memory_gb=134,
-            kwargs={
-                "model_path": "Lightricks/LTX-2.3",
-                "num_gpus": 2,
-                "pipeline_class_name": "LTX2TwoStagePipeline",
-                "ltx2_two_stage_device_mode": "snapshot",
-            },
-        )
-
-        self.assertEqual(args.ltx2_two_stage_device_mode, "original")
-        self.assertEqual(
-            args.layerwise_offload_components,
-            ["text_encoder", "image_encoder", "vae"],
-        )
+    def test_ltx23_snapshot_device_mode_is_rejected(self):
+        with self.assertRaisesRegex(ValueError, "Expected one of"):
+            self._from_dict_with_pipeline_config(
+                LTX2PipelineConfig(),
+                memory_gb=140,
+                available_memory_gb=134,
+                kwargs={
+                    "model_path": "Lightricks/LTX-2.3",
+                    "num_gpus": 2,
+                    "pipeline_class_name": "LTX2TwoStagePipeline",
+                    "ltx2_two_stage_device_mode": "snapshot",
+                },
+            )
 
     def test_explicit_layerwise_components_preserved_in_ltx23_resident(self):
         args = self._from_dict_with_pipeline_config(
@@ -2425,7 +2420,7 @@ class TestOffloadDefaults(unittest.TestCase):
         self.assertFalse(server_args.use_fsdp_inference)
         self.assertFalse(server_args.enable_cfg_parallel)
 
-    def test_ltx23_snapshot_device_mode_cli_alias_is_accepted(self):
+    def test_ltx23_snapshot_device_mode_cli_is_rejected(self):
         parser = FlexibleArgumentParser()
         ServerArgs.add_cli_args(parser)
         argv = [
@@ -2437,36 +2432,8 @@ class TestOffloadDefaults(unittest.TestCase):
             "snapshot",
         ]
 
-        with (
-            patch.object(sys, "argv", ["sglang"] + argv),
-            patch.object(
-                PipelineConfig, "from_kwargs", return_value=LTX2PipelineConfig()
-            ),
-            patch(
-                "sglang.multimodal_gen.runtime.platforms.current_platform.is_cpu",
-                return_value=False,
-            ),
-            patch(
-                "sglang.multimodal_gen.runtime.platforms.current_platform.is_mps",
-                return_value=False,
-            ),
-            patch(
-                "sglang.multimodal_gen.runtime.platforms.current_platform.is_cuda",
-                return_value=True,
-            ),
-            patch(
-                "sglang.multimodal_gen.runtime.platforms.current_platform.get_device_total_memory",
-                return_value=140 * 1024**3,
-            ),
-            patch(
-                "sglang.multimodal_gen.runtime.platforms.current_platform.get_available_gpu_memory",
-                return_value=134,
-            ),
-        ):
-            args, unknown_args = parser.parse_known_args(argv)
-            server_args = ServerArgs.from_cli_args(args, unknown_args)
-
-        self.assertEqual(server_args.ltx2_two_stage_device_mode, "original")
+        with self.assertRaises(SystemExit):
+            parser.parse_known_args(argv)
 
 
 class TestKVGatherDegree(unittest.TestCase):