Arch config for Wan2.2-S2V-14B (speech-to-video).
Every architecture value here is transcribed verbatim from the official checkpoint manifest at Wan-AI/Wan2.2-S2V-14B/config.json -- do not "tidy" them. dim: 5120 and num_heads: 40 in that file become num_attention_heads=40 / attention_head_dim=128 here (5120 / 40).
Unlike every other Wan variant in this package, the S2V checkpoint ships in native Wan format (blocks.0.self_attn.q), not diffusers format (blocks.0.attn1.to_q). param_names_mapping therefore translates from the native layout; see lora_param_names_mapping in wanvideo.py for the same naming scheme applied to Wan LoRA adapters.