qwen3.5-4b模型对于padding_free的支持问题

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

在ms-swift源码中：

swift 的 sp_linear_forward 内部（简化示意）

def sp_linear_forward(self, hidden_states, **kwargs):
if not sequence_parallel.enabled() and 'cu_seq_lens_q' not in kwargs:
# 第一分支：走 transformers 原始 forward（fla CUDA 内核，正常）
return original_forward(self, hidden_states, **kwargs)
else:
# 第二分支：走 swift varlen 路径 → 尝试 import flash_linear_attention → 崩溃！
return _run_sequence_parallel_forward(self, hidden_states, **kwargs)

此处的条件判断由于伪packing机制导致所有训练无论batch_size的大小都会走varlen实现去跑，但是varlen中实际没有写回退原生torch实现的方法。而Windows中目前虽然有fla和triton实现，但是没有合现成的causal-conv1d whl，因此出现ExceptImport。

How to Reproduce / 如何复现

1

Additional Information / 补充信息

No response

Start by tracing the described sp_linear_forward condition and the _run_sequence_parallel_forward path in the ms-swift source, comparing them with original_forward for qwen3.5-4b and padding_free training. Reproduce the Windows failure involving flash_linear_attention and the missing causal-conv1d wheel; done means training no longer crashes when the varlen dependencies are unavailable.