Fix preserving output_router_logits in multimodal template forward hooks

[hkxxxxx]

Summary

Preserve output_router_logits when Template.pre_forward_hook restores model-forward kwargs after multimodal _post_encode.

Fixes #10189.

Background

Native Swift DPO/SFT can enable the MoE router auxiliary loss with:

output_router_logits = True

The Qwen3.5 multimodal template reuses a Qwen2VL post-encoding path that returns a new kwargs dictionary. The generic template hook restores only fields in a manually maintained whitelist. Because output_router_logits was missing from that whitelist, the flag was silently removed before the Hugging Face model forward.

With router_aux_loss_coef > 0, Qwen3.5 MoE then returns aux_loss=None; native DPO fails when it evaluates:

self.aux_loss_coef * model_output['aux_loss']

Changes

• Add output_router_logits to the forward-kwargs restoration whitelist in Template.pre_forward_hook.
• Add a regression test covering a post-encode path that rebuilds kwargs while preserving output_router_logits=True.

Validation

• Reproduced the failure with Qwen3.5/Qwen3.6 MoE native DPO, LoRA, DeepSpeed, and router_aux_loss_coef=1e-3.
• Confirmed the first Python failure is aux_loss=None, not NCCL initialization.
• Confirmed the one-line fix produces a numeric aux_loss and allows the training step to continue.
• conda run -n llm python -m pytest tests/general/test_template_forward_hook.py -q — 1 passed.
• python -m py_compile swift/template/base.py tests/general/test_template_forward_hook.py.
• git diff --check.

Compatibility and risk

This change only preserves a caller-provided forward kwarg. It does not change the default model behavior when the kwarg is absent, and it does not alter the auxiliary-loss coefficient or loss formula.

The field is already used by the native DPO/SFT trainers for supported MoE models. The regression test targets the parameter-loss mechanism directly and does not require model weights or private data.

Checklist

• Bug Fix
• Add/update regression test.
• Run the relevant focused test.
• Confirm no private paths, logs, credentials, or local environment files are included.
• Run a full Qwen3.5 MoE smoke test when the target hardware/environment is available.

[tastelikefeet]

Conflicting