v1.0.2 — Qwen structured-prompt normalization

Fixes plaintext T2V, TI2V, and FLF prompts that Qwen emits with world_knowledge as a string, object, null, or missing field. Deterministic normalization now produces the required non-empty string array without an unnecessary second Qwen pass. The FP8 checkpoint is byte-for-byte unchanged from v1.0.1 and is republished under this tag to keep the version-pinned downloader coherent.

v1.0.1 — Release hardening

This hardening release makes the public Dense 1.3B FP8 package safer to install, reproduce, and troubleshoot.

• Corrected all workflows to conservative 640×352, 20 FPS, 3-second, 28-step defaults with fixed seed 42.
• Disabled FlashVSR and RIFE by default and repaired lazy switch wiring.
• Added an early CUDA FP8 capability probe with an actionable BF16 fallback.
• Added mandatory dependency declarations, portable/source install instructions, CI, release digest verification, and public support/security metadata.
• Preserved the exact verified FP8 checkpoint bytes from v1.0.0; SHA-256 hashes are unchanged.

See CHANGELOG.md and BENCHMARKS.md for details and measurement limitations.

LingBot-Video Dense 1.3B FP8 v1.0.0

Community FP8 checkpoint and ComfyUI runtime for LingBot-Video Dense 1.3B. The selective E4M3 W8A8 path accelerates fused QKV and MLP gate/up while retaining other paths in BF16. Matched RTX 5080 smoke: 3.651 s FP8 vs 4.651 s BF16 (1.27x). Download both model files into transformer_fp8_dense, or use tools/download_fp8.py. SHA-256 checksums are attached and tracked in the repository. This is an independent community release, not an official Robbyant release.