Refactor CPU Offloading Backend Pattern (#1223)

Signed-off-by: yuanheng <jonathan.zhaoyh@gmail.com>
Signed-off-by: Yuanheng Zhao <jonathan.zhaoyh@gmail.com>
Signed-off-by: Samit <285365963@qq.com>
Co-authored-by: Samit <285365963@qq.com>

vllm_omni/diffusion/data.py

@@ -298,11 +298,8 @@ class OmniDiffusionConfig:
     # - Text encoders run on GPU while DiT is on CPU
     # - DiT runs on GPU while encoders are on CPU
     enable_cpu_offload: bool = False
-
     # Layer-wise offloading (block-level offloading) parameters
     enable_layerwise_offload: bool = False
-    # Number of transformer blocks ready for computation to keep on GPU
-    layerwise_num_gpu_layers: int = 1
 
     use_fsdp_inference: bool = False
     pin_cpu_memory: bool = True  # Use pinned memory for faster transfers when offloading