Torch fallbacks of the linear-attention ops drop `seq_idx` / `cu_seqlens`, so packed documents leak without the optional packages

当你在没有安装 flash-linear-attention / causal-conv1d 等可选加速包的环境中,用 DataCollatorWithFlattening 做 packed(无 padding)训练或推理时,linear-attention 的 Torch fallback 会丢弃

![[Bug]: Qwen3.5-35B-A3B merged SFT BF16 intermittently produces NaN logits and exclamation-only output on H20 (v0.19.1; HF control finite)](https://www.chat-gpts.plus/wp-content/uploads/2026/10/59381-d36a4086-768x403.jpg)


![[Bug]: HunYuan Dense V1 fails CUDA graph capture on v0.29 — HF dynamic RoPE does a host-side sync during capture](https://www.chat-gpts.plus/wp-content/uploads/2026/10/57890-0b89f014-768x403.jpg)


![ValueError: The number of images in the text [2] and images [1] should be the same.](https://www.chat-gpts.plus/wp-content/uploads/2026/10/47217-a358ec78-768x403.jpg)
