[Bug] Qwen3.5-35B-A3B LoRA SFT: loss=0, token_acc=0

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

Environment

• ms-swift: v4.0.0.dev0 (main branch, pulled 2026-02-28)
• transformers: 5.2.0
• deepspeed: ZeRO-3
• GPU: 4 nodes × 4 L40s (16 GPUs)
• Model: Qwen3.5-35B-A3B

Problem

When running LoRA SFT on Qwen3.5-35B-A3B with swift sft, loss is always 0.0, token_acc is always 0.0, and grad_norm is stuck at 1.0 (likely only from router_aux_loss).

The same training data works perfectly fine with Qwen3-VL-4B and Qwen3-VL-8B.

Training log (logging.jsonl):

{"current_steps": 1, "loss": 0.0, "grad_norm": 1.0, "learning_rate": 0.0, "epoch": 0.0, "token_acc": 0.0}
{"current_steps": 10, "loss": 0.0, "grad_norm": 1.0, "learning_rate": 5.1e-05, "epoch": 0.03, "token_acc": 0.0}
{"current_steps": 20, "loss": 0.0, "grad_norm": 1.0, "learning_rate": 0.0001, "epoch": 0.06, "token_acc": 0.0}
{"current_steps": 30, "loss": 0.0, "grad_norm": 1.0, "learning_rate": 9.9e-05, "epoch": 0.09, "token_acc": 0.0}

Training Command

swift sft
--model /path/to/Qwen3.5-35B-A3B
--dataset train_data.jsonl
--train_type lora
--lora_rank 16 --lora_alpha 32
--deepspeed ds_zero3.json
--max_length 12288
--padding_free false --packing false
--attn_impl flash_attn
（loss_scale and add_non_thinking_prefix not specified (using defaults)）

args.json confirms correct model detection:
model_type: qwen3_5_moe
template: qwen3_5
loss_scale: default

How to Reproduce

1. Prepare any SFT dataset in standard messages format (with or without images)
2. Run training with default settings:

swift sft \
  --model Qwen/Qwen3.5-35B-A3B \
  --dataset your_data.jsonl \
  --train_type lora \
  --max_length 12288
3. Observe loss: 0.0, token_acc: 0.0, grad_norm: 1.0 in logging.jsonl

Start by reproducing the report with swift sft using Qwen/Qwen3.5-35B-A3B, LoRA, and the default settings, then inspect logging.jsonl and args.json for the zero loss and token accuracy. Compare the run with the working Qwen3-VL cases mentioned in the report; done means SFT produces nonzero loss and token accuracy with gradients reflecting the training loss.