[diffusion] Keep LongLive2 components resident on large GPUs (#35993)

docs/cookbook/diffusion/LongLive/LongLive-2.0.mdx

@@ -25,6 +25,11 @@ Please refer to the [official SGLang-diffusion installation guide](/docs/sglang-
 sglang serve --model-path Rabinovich/LongLive-2.0-5B-Diffusers
 ```
 
+In `auto` mode, GPUs with at least 60 GiB available keep the DiT, text encoder,
+and VAE resident. This uses about 44 GiB on H200 for the 832x480 preset and
+avoids moving encoder and decoder layers from host memory on every request.
+Smaller GPUs retain the layerwise-offload defaults.
+
 If the GPU runs out of memory, move the text encoder, VAE, and DiT to CPU between stages:
 
 ```bash Command

python/sglang/multimodal_gen/configs/pipeline_configs/longlive2.py

@@ -6,6 +6,9 @@ from dataclasses import dataclass, field
 from sglang.multimodal_gen.configs.models import DiTConfig
 from sglang.multimodal_gen.configs.models.dits.longlive2 import LongLive2VideoConfig
 from sglang.multimodal_gen.configs.pipeline_configs.base import ModelTaskType
+from sglang.multimodal_gen.configs.pipeline_configs.model_deployment_config import (
+    ModelDeploymentConfig,
+)
 from sglang.multimodal_gen.configs.pipeline_configs.wan import Wan2_2_TI2V_5B_Config
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
 
@@ -28,6 +31,13 @@ class LongLive2T2VConfig(Wan2_2_TI2V_5B_Config):
 
     dit_config: DiTConfig = field(default_factory=LongLive2VideoConfig)
 
+    def get_model_deployment_config(self) -> ModelDeploymentConfig:
+        return ModelDeploymentConfig(
+            dit_layerwise_offload_modes=("memory",),
+            keep_resident_min_available_gb=60,
+            keep_resident_components=("dit", "text_encoder", "vae"),
+        )
+
     def adjust_num_frames(self, num_frames: int) -> int:
         num_frames = super().adjust_num_frames(num_frames)
         vae_scale_factor_temporal = self.vae_config.arch_config.scale_factor_temporal

python/sglang/multimodal_gen/test/server/perf_baselines/h100.json

@@ -2713,8 +2713,8 @@
             "expected_e2e_ms": 5648.49,
             "expected_avg_denoise_ms": 477.49,
             "expected_median_denoise_ms": 56.74,
-            "load_peak_vram_mb": 16110.0,
-            "runtime_peak_vram_mb": 45756.0,
+            "load_peak_vram_mb": 32892.0,
+            "runtime_peak_vram_mb": 61190.0,
             "estimated_full_test_time_s": 153.1
         },
         "longlive2_i2v": {
@@ -2731,8 +2731,8 @@
             "expected_e2e_ms": 9086.81,
             "expected_avg_denoise_ms": 492.02,
             "expected_median_denoise_ms": 151.9,
-            "load_peak_vram_mb": 16110.0,
-            "runtime_peak_vram_mb": 45756.0,
+            "load_peak_vram_mb": 32892.0,
+            "runtime_peak_vram_mb": 62510.0,
             "estimated_full_test_time_s": 149.4
         },
         "lingbot_video_moe_t2v": {

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

@@ -24,6 +24,9 @@ from sglang.multimodal_gen.configs.pipeline_configs.lingbot_world import (
 from sglang.multimodal_gen.configs.pipeline_configs.longcat_image import (
     LongCatImagePipelineConfig,
 )
+from sglang.multimodal_gen.configs.pipeline_configs.longlive2 import (
+    LongLive2T2VConfig,
+)
 from sglang.multimodal_gen.configs.pipeline_configs.ltx_2 import (
     LTX2PipelineConfig,
     LTX23PipelineConfig,
@@ -1417,6 +1420,7 @@ class TestOffloadDefaults(unittest.TestCase):
         lingbot_deployment = LingBotWorldCausalDMDConfig().get_model_deployment_config()
         ltx_deployment = LTX2PipelineConfig().get_model_deployment_config()
         ltx23_config = LTX23PipelineConfig()
+        longlive_deployment = LongLive2T2VConfig().get_model_deployment_config()
         sana_wm_deployment = SanaWMPipelineConfig().get_model_deployment_config()
 
         self.assertIsNone(qwen_deployment.fsdp_auto_min_available_memory_gb)
@@ -1454,6 +1458,11 @@ class TestOffloadDefaults(unittest.TestCase):
         self.assertEqual(ltx_deployment.get_auto_cfg_parallel_degree(4), 1)
         self.assertEqual(ltx_deployment.get_auto_cfg_parallel_degree(8), 1)
         self.assertEqual(ltx_deployment.get_auto_cfg_parallel_degree(2), 2)
+        self.assertEqual(longlive_deployment.keep_resident_min_available_gb, 60)
+        self.assertEqual(
+            longlive_deployment.keep_resident_components,
+            ("dit", "text_encoder", "vae"),
+        )
         self.assertFalse(
             LTX2PipelineConfig().dit_config.arch_config.enable_packed_qkv_input_a2a
         )
@@ -1486,6 +1495,25 @@ class TestOffloadDefaults(unittest.TestCase):
         self.assertEqual(qwen_deployment.keep_resident_components, ("vae",))
         self.assertIsNone(qwen_deployment.keep_resident_min_available_gb)
 
+    def test_longlive_residency_scales_with_available_memory(self):
+        high_memory_args = self._from_dict_with_pipeline_config(
+            LongLive2T2VConfig(),
+            memory_gb=80,
+            kwargs={"performance_mode": "auto"},
+        )
+        high_memory_offload = high_memory_args.layerwise_offload_components or []
+        self.assertNotIn("text_encoder", high_memory_offload)
+        self.assertNotIn("vae", high_memory_offload)
+
+        constrained_args = self._from_dict_with_pipeline_config(
+            LongLive2T2VConfig(),
+            memory_gb=50,
+            kwargs={"performance_mode": "auto"},
+        )
+        constrained_offload = constrained_args.layerwise_offload_components or []
+        self.assertIn("text_encoder", constrained_offload)
+        self.assertIn("vae", constrained_offload)
+
     def test_qwen_ar_generation_residency_scales_with_available_memory(self):
         pipeline_configs = (
             QwenImageLayeredPipelineConfig(),