[bugfix] Fix Qwen3.5 non-packing causal conv routing

[taking-lying-flat]

What changed

• Treat cu_seq_lens_q=None as a non-packing Qwen3.5 forward instead of selecting the FLA packing/sequence-parallel implementation merely because the optional kwarg is present.

Root cause

In transformers==5.12.1, the failing line reported in #9810 is after the causal-convolution call, not immediately after in_proj_qkv. The native causal_conv1d_fn returns a tensor, while FLA's causal_conv1d returns an (output, state) tuple.

Older ms-swift code could install the FLA callable into mod.causal_conv1d_fn when the packed path was selected. Although #9643 isolated that callable on current main, the wrapper still selected the packed path whenever the cu_seq_lens_q key existed, even when its value was None. Optional FlashAttention kwargs can be propagated with None in a regular non-packing forward.

This change uses the value of cu_seq_lens_q to select the packed path, so a normal Qwen3.5 SFT forward stays on the Transformers implementation.

Impact

Qwen3.5 non-packing SFT, including LoRA, gradient checkpointing, DeepSpeed, and resume flows, no longer enters the tuple-returning FLA causal-convolution path solely because an empty optional kwarg was propagated.

Fixes #9810.

Validation

• flake8 swift/model/models/qwen.py
• isort --check-only swift/model/models/qwen.py
• yapf --diff swift/model/models/qwen.py
• git diff --check

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.