Wired up FastVideo/LingBot-World-Fast-Diffusers, the old LingBot-World
repo's distilled causal I2V checkpoint that was never ported. Reused
the existing LingBot World 2 causal-fast DiT unchanged (verified all
1421 parameter names and shapes against the released checkpoint) and
added the arch/pipeline configs, registry entries, preset, pipeline
subclass, example, and SSIM test needed to run it.
Fixed a model-family detector collision where the LingBot World (Cam)
detector would have wrongly claimed this checkpoint's path and
model_index name, since it only excluded the "causal-fast" spelling.
Fixed WanCrossAttention in the shared LingBot World 2 causal-fast DiT
to route through the attention() dispatcher instead of calling
flash_attention() directly, so the already-advertised TORCH_SDPA
backend actually works when flash-attn isn't installed. Self-attention
already used this dispatcher; cross-attention did not.
Verified end to end: weights load (18.54B params, 16 shards) and
generation produces a coherent video with correct prompt-following
camera motion. Not yet verified against the original implementation's
numerics (activation-trace parity) or via the SSIM regression test,
both of which need hardware unavailable in this environment.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
