```shell
Loading required modules: 100%|████████████████████████████████████████████████████████████████████████████████| 6/6 [03:10<00:00, 31.74s/it]
[12-13 07:41:25] Pipelines instantiated
[12-13 07:41:25] Worker 0: Initialized device, model, and distributed environment.
[12-13 07:41:25] Worker 0: Scheduler loop started.
[12-13 07:41:25] Rank 0 scheduler listening on tcp://*:5638
Loading required modules: 100%|████████████████████████████████████████████████████████████████████████████████| 6/6 [03:10<00:00, 31.77s/it]
Loading required modules: 100%|████████████████████████████████████████████████████████████████████████████████| 6/6 [03:11<00:00, 31.84s/it]
Loading required modules: 100%|████████████████████████████████████████████████████████████████████████████████| 6/6 [03:12<00:00, 32.07s/it]
[12-13 07:41:27] Sampling params:
                       width: -1
                      height: -1
                  num_frames: 125
                      prompt: A young woman with long blonde hair, wearing a white t-shirt and blue jeans, walking through a sunny park with green trees in the background, carrying a brown shoulder bag, followed by a smooth tracking shot that moves alongside her as she walks.
                  neg_prompt: 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走
                        seed: 1024
                 infer_steps: 40
      num_outputs_per_prompt: 1
              guidance_scale: 4.0
     embedded_guidance_scale: 6.0
                    n_tokens: -1
                  flow_shift: 12.0
                  image_path: None
                 save_output: True
            output_file_path: outputs/A_young_woman_with_long_blonde_hair_wearing_a_white_t-shirt_and_blue_jeans_walking_through_a_sunny_20251213-074127_4269cbf7.mp4
        
[12-13 07:41:27] Processing prompt 1/1: A young woman with long blonde hair, wearing a white t-shirt and blue jeans, walking through a sunny
/nas/bbuf/sglang/python/sglang/multimodal_gen/runtime/utils/distributed.py:34: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:206.)
  tensor_data = torch.ByteTensor(
[12-13 07:41:27] Creating pipeline stages...
[12-13 07:41:27] Using FlashAttention (FA3 for hopper, FA4 for blackwell) backend
[12-13 07:41:27] Running pipeline stages: ['input_validation_stage', 'prompt_encoding_stage', 'conditioning_stage', 'timestep_preparation_stage', 'latent_preparation_stage', 'denoising_stage', 'decoding_stage']
[12-13 07:41:27] Profiling request: mocked_fake_id_for_offline_generate for 5 steps...
[12-13 07:41:27] Starting Profiler...
[12-13 07:41:27] [InputValidationStage] started...
[12-13 07:41:27] [InputValidationStage] finished in 0.0043 seconds
[12-13 07:41:27] [TextEncodingStage] started...
WARNING 12-13 07:41:31 [block_adapters.py:131] pipe is None, use FakeDiffusionPipeline instead.
INFO 12-13 07:41:31 [block_adapters.py:229] Auto fill blocks_name: ['blocks', 'blocks'].
INFO 12-13 07:41:31 [block_adapters.py:162] Found transformer NOT from diffusers: sglang.multimodal_gen.runtime.models.dits.wanvideo disable check_forward_pattern by default.
INFO 12-13 07:41:31 [cache_interface.py:200] cache_config is None, using default DBCacheConfig
INFO 12-13 07:41:31 [cache_adapter.py:77] Adapting Cache Acceleration using custom BlockAdapter!
WARNING 12-13 07:41:31 [block_adapters.py:478] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
WARNING 12-13 07:41:31 [block_adapters.py:478] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [cache_adapter.py:134] Use custom 'enable_separate_cfg' from BlockAdapter: True. Pipeline: FakeDiffusionPipeline.
INFO 12-13 07:41:31 [cache_adapter.py:307] Collected Context Config: DBCache_F1B0_W4I1M0MC8_R0.24_SCM1111111111001110000011100000111000110001_dynamic, Calibrator Config: None
INFO 12-13 07:41:31 [cache_adapter.py:307] Collected Context Config: DBCache_F1B0_W2I1M0MC20_R0.24_SCM1111111111001110000011100000111000110001_dynamic, Calibrator Config: None
WARNING 12-13 07:41:31 [pattern_base.py:78] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [pattern_base.py:70] Match Blocks: CachedBlocks_Pattern_0_1_2, for blocks, cache_context: blocks_140418593588720, context_manager: FakeDiffusionPipeline_140417988320464.
WARNING 12-13 07:41:31 [pattern_base.py:78] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [pattern_base.py:70] Match Blocks: CachedBlocks_Pattern_0_1_2, for blocks, cache_context: blocks_140418588350976, context_manager: FakeDiffusionPipeline_140417988320464.
WARNING 12-13 07:41:31 [block_adapters.py:131] pipe is None, use FakeDiffusionPipeline instead.
INFO 12-13 07:41:31 [block_adapters.py:229] Auto fill blocks_name: ['blocks', 'blocks'].
INFO 12-13 07:41:31 [block_adapters.py:162] Found transformer NOT from diffusers: sglang.multimodal_gen.runtime.models.dits.wanvideo disable check_forward_pattern by default.
INFO 12-13 07:41:31 [cache_interface.py:200] cache_config is None, using default DBCacheConfig
INFO 12-13 07:41:31 [cache_adapter.py:77] Adapting Cache Acceleration using custom BlockAdapter!
WARNING 12-13 07:41:31 [block_adapters.py:478] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
WARNING 12-13 07:41:31 [block_adapters.py:478] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [cache_adapter.py:134] Use custom 'enable_separate_cfg' from BlockAdapter: True. Pipeline: FakeDiffusionPipeline.
INFO 12-13 07:41:31 [cache_adapter.py:307] Collected Context Config: DBCache_F1B0_W4I1M0MC8_R0.24_SCM1111111111001110000011100000111000110001_dynamic, Calibrator Config: None
INFO 12-13 07:41:31 [cache_adapter.py:307] Collected Context Config: DBCache_F1B0_W2I1M0MC20_R0.24_SCM1111111111001110000011100000111000110001_dynamic, Calibrator Config: None
WARNING 12-13 07:41:31 [pattern_base.py:78] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [pattern_base.py:70] Match Blocks: CachedBlocks_Pattern_0_1_2, for blocks, cache_context: blocks_140336518225488, context_manager: FakeDiffusionPipeline_140334574005872.
WARNING 12-13 07:41:31 [pattern_base.py:78] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [pattern_base.py:70] Match Blocks: CachedBlocks_Pattern_0_1_2, for blocks, cache_context: blocks_140334574005152, context_manager: FakeDiffusionPipeline_140334574005872.
WARNING 12-13 07:41:31 [block_adapters.py:131] pipe is None, use FakeDiffusionPipeline instead.
INFO 12-13 07:41:31 [block_adapters.py:229] Auto fill blocks_name: ['blocks', 'blocks'].
INFO 12-13 07:41:31 [block_adapters.py:162] Found transformer NOT from diffusers: sglang.multimodal_gen.runtime.models.dits.wanvideo disable check_forward_pattern by default.
INFO 12-13 07:41:31 [cache_interface.py:200] cache_config is None, using default DBCacheConfig
INFO 12-13 07:41:31 [cache_adapter.py:77] Adapting Cache Acceleration using custom BlockAdapter!
WARNING 12-13 07:41:31 [block_adapters.py:478] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
WARNING 12-13 07:41:31 [block_adapters.py:478] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [cache_adapter.py:134] Use custom 'enable_separate_cfg' from BlockAdapter: True. Pipeline: FakeDiffusionPipeline.
INFO 12-13 07:41:31 [cache_adapter.py:307] Collected Context Config: DBCache_F1B0_W4I1M0MC8_R0.24_SCM1111111111001110000011100000111000110001_dynamic, Calibrator Config: None
INFO 12-13 07:41:31 [cache_adapter.py:307] Collected Context Config: DBCache_F1B0_W2I1M0MC20_R0.24_SCM1111111111001110000011100000111000110001_dynamic, Calibrator Config: None
WARNING 12-13 07:41:31 [pattern_base.py:78] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [pattern_base.py:70] Match Blocks: CachedBlocks_Pattern_0_1_2, for blocks, cache_context: blocks_139832392396256, context_manager: FakeDiffusionPipeline_139831119788416.
WARNING 12-13 07:41:31 [pattern_base.py:78] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [pattern_base.py:70] Match Blocks: CachedBlocks_Pattern_0_1_2, for blocks, cache_context: blocks_139831119936208, context_manager: FakeDiffusionPipeline_139831119788416.
[12-13 07:41:31] [TextEncodingStage] finished in 3.5739 seconds
[12-13 07:41:31] [ConditioningStage] started...
[12-13 07:41:31] [ConditioningStage] finished in 0.0001 seconds
[12-13 07:41:31] [TimestepPreparationStage] started...
[12-13 07:41:31] [TimestepPreparationStage] finished in 0.0036 seconds
[12-13 07:41:31] [LatentPreparationStage] started...
[12-13 07:41:31] [LatentPreparationStage] finished in 0.0021 seconds
[12-13 07:41:31] [DenoisingStage] started...
[12-13 07:41:31] cache-dit is running in distributed environment (world_size=4). Using local caching strategy: each GPU caches its own activation shards. This may be less accurate than single-GPU caching but provides speedup.
[12-13 07:41:31] SCM: generated mask with 22 compute steps, 18 cache steps (preset=medium)
[12-13 07:41:31] Enabling cache-dit on wan2.2 dual transformers with BlockAdapter
[12-13 07:41:31]   Primary (transformer): Fn=1, Bn=0, W=4, R=0.24, MC=8, TaylorSeer=False
[12-13 07:41:31]   Secondary (transformer_2): Fn=1, Bn=0, W=2, R=0.24, MC=20, TaylorSeer=False
[12-13 07:41:31]   SCM enabled: 22 compute steps, 18 cache steps, policy=dynamic
WARNING 12-13 07:41:31 [block_adapters.py:131] pipe is None, use FakeDiffusionPipeline instead.
INFO 12-13 07:41:31 [block_adapters.py:229] Auto fill blocks_name: ['blocks', 'blocks'].
INFO 12-13 07:41:31 [block_adapters.py:162] Found transformer NOT from diffusers: sglang.multimodal_gen.runtime.models.dits.wanvideo disable check_forward_pattern by default.
INFO 12-13 07:41:31 [cache_interface.py:200] cache_config is None, using default DBCacheConfig
INFO 12-13 07:41:31 [cache_adapter.py:77] Adapting Cache Acceleration using custom BlockAdapter!
WARNING 12-13 07:41:31 [block_adapters.py:478] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
WARNING 12-13 07:41:31 [block_adapters.py:478] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [cache_adapter.py:134] Use custom 'enable_separate_cfg' from BlockAdapter: True. Pipeline: FakeDiffusionPipeline.
INFO 12-13 07:41:31 [cache_adapter.py:307] Collected Context Config: DBCache_F1B0_W4I1M0MC8_R0.24_SCM1111111111001110000011100000111000110001_dynamic, Calibrator Config: None
INFO 12-13 07:41:31 [cache_adapter.py:307] Collected Context Config: DBCache_F1B0_W2I1M0MC20_R0.24_SCM1111111111001110000011100000111000110001_dynamic, Calibrator Config: None
WARNING 12-13 07:41:31 [pattern_base.py:78] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [pattern_base.py:70] Match Blocks: CachedBlocks_Pattern_0_1_2, for blocks, cache_context: blocks_140168208559296, context_manager: FakeDiffusionPipeline_140164652767888.
WARNING 12-13 07:41:31 [pattern_base.py:78] Skipped Forward Pattern Check: ForwardPattern.Pattern_2
INFO 12-13 07:41:31 [pattern_base.py:70] Match Blocks: CachedBlocks_Pattern_0_1_2, for blocks, cache_context: blocks_140165252596704, context_manager: FakeDiffusionPipeline_140164652767888.
[12-13 07:41:31] cache-dit enabled on dual transformers (steps=40)
  0%|                                                                                                                 | 0/40 [00:29<?, ?it/s]
[12-13 07:42:01] [DenoisingStage] Error during execution after 29571.3200 ms: The size of tensor a (135) must match the size of tensor b (134) at non-singleton dimension 4
Traceback (most recent call last):
  File "/nas/bbuf/sglang/python/sglang/multimodal_gen/runtime/pipelines_core/executors/parallel_executor.py", line 90, in _execute
    batch = stage(batch, server_args)
            ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/nas/bbuf/sglang/python/sglang/multimodal_gen/runtime/pipelines_core/stages/base.py", line 192, in __call__
    result = self.forward(batch, server_args)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/nas/bbuf/sglang/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py", line 1010, in forward
    latents = self.scheduler.step(
              ^^^^^^^^^^^^^^^^^^^^
  File "/nas/bbuf/sglang/python/sglang/multimodal_gen/runtime/models/schedulers/scheduling_flow_unipc_multistep.py", line 745, in step
    model_output_convert = self.convert_model_output(model_output, sample=sample)
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/nas/bbuf/sglang/python/sglang/multimodal_gen/runtime/models/schedulers/scheduling_flow_unipc_multistep.py", line 333, in convert_model_output
    x0_pred = sample - sigma_t * model_output
              ~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~
RuntimeError: The size of tensor a (135) must match the size of tensor b (134) at non-singleton dimension 4
[12-13 07:42:01] Stopping Profiler...
[12-13 07:42:01] Saving profiler traces to: /nas/bbuf/sglang/logs/mocked_fake_id_for_offline_generate-5_steps-global-rank0.trace.json.gz
[12-13 07:42:01] Failed to generate output for prompt 1: Error executing request mocked_fake_id_for_offline_generate: The size of tensor a (135) must match the size of tensor b (134) at non-singleton dimension 4
Traceback (most recent call last):
  File "/nas/bbuf/sglang/python/sglang/multimodal_gen/runtime/utils/logging_utils.py", line 495, in log_generation_timer
    yield timer
  File "/nas/bbuf/sglang/python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py", line 273, in generate
    raise Exception(f"{output_batch.error}")
Exception: Error executing request mocked_fake_id_for_offline_generate: The size of tensor a (135) must match the size of tensor b (134) at non-singleton dimension 4
[12-13 07:42:01] Completed batch processing. Generated 0 outputs in 33.90 seconds.
[12-13 07:42:01] Generator was garbage collected without being shut down. Attempting to shut down the local server and client.
/usr/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 4 leaked semaphore objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '

### Reproduction

```shell
SGLANG_CACHE_DIT_ENABLED=true SGLANG_CACHE_DIT_WARMUP=4 SGLANG_CACHE_DIT_MC=8 SGLANG_CACHE_DIT_RDT=0.24 SGLANG_CACHE_DIT_FN=1 SGLANG_CACHE_DIT_BN=0 SGLANG_CACHE_DIT_SECONDARY_WARMUP=2 SGLANG_CACHE_DIT_SECONDARY_MC=20 SGLANG_CACHE_DIT_SECONDARY_RDT=0.24 SGLANG_CACHE_DIT_SECONDARY_FN=1 SGLANG_CACHE_DIT_SECONDARY_BN=0 SGLANG_CACHE_DIT_SCM_PRESET=medium SGLANG_CACHE_DIT_SCM_POLICY=dynamic sglang generate   --model-path /nas/bbuf/Wan2.2-T2V-A14B-Diffusers   --text-encoder-cpu-offload   --pin-cpu-memory   --num-gpus 4   --ulysses-degree 4 --profile --prompt "A young woman with long blonde hair, wearing a white t-shirt and blue jeans, walking through a sunny park with green trees in the background, carrying a brown shoulder bag, followed by a smooth tracking shot that moves alongside her as she walks."