Inconsistent Handling of Zero-Initialized CLIP Context Between Training and Inference in i2v

Description

I noticed an inconsistency in the way _clip_context is zero-initialized during training versus inference when using cfg in the i2v codebase:

1. Training
   In scripts/wan2.1_fun/train.py, there is a 10% probability of zero-initializing the CLIP context:

   zero_init_clip_in = np.random.choice([True, False], p=[0.1, 0.9])
   clip_context.append(_clip_context if not zero_init_clip_in else torch.zeros_like(_clip_context))
   

2. Inference
   During inference with cfg in videox_fun/pipeline/pipeline_wan_fun_inpaint.py, both halves of the context vector are concatenated identically, regardless of zero initialization:

   clip_context_input = (
       torch.cat([clip_context] * 2) if do_classifier_free_guidance 
       else clip_context
   )
   

Question

Is this discrepancy between training and inference intentional? If not, should the inference phase also zero out the conditioning context to better match the training behavior? Thank you!

Compare the zero-initialization logic in scripts/wan2.1_fun/train.py with CFG context handling in videox_fun/pipeline/pipeline_wan_fun_inpaint.py. Determine whether the discrepancy is intentional and identify the expected inference behavior; the issue does not name a test or a specific completion criterion.