feat(npu): Add FSDP2 LoRA example for Qwen3.8-27B for Ascend NPU

[zheliuyu]

Add Qwen3.8-27B Ascend NPU FSDP2 LoRA Training Example

ms-swift + flash-linear-attention NPU adaptation

• @sunyi0505
• @addsubmuldiv
• @ChengQianqian
• @hazelduan
• @xvxuopop
• @OsirisDuan

Summary

Qwen3.8 and Qwen3.5 share a similar architecture—both use GDN—so this PR is primarily about demonstrating that ms-swift + flash-linear-attention works on Ascend NPU.

The main advantage of this setup is that no extra NPU training stack is required: both ms-swift and flash-linear-attention provide native NPU support.

Timeline for this PR:

• Aug 14 — Qwen3.8 open-sourced; swap in the Qwen3.8 model name, verify training, and open the PR
• Aug 13 — Environment setup; validate end-to-end training with Qwen3.5 first
• Early Aug — ms-swift + flash-linear-attention fully working on Qwen3.5
• Early Jul — flash-linear-attention NPU adaptation
• ms-swift NPU adaptation

PR Type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

Description

Adds an FSDP2 + LoRA SFT training example for Qwen3.8-27B on Ascend NPU, ready for post-release NPU adaptation.

New files

• examples/ascend/train/qwen3_8/qwen3_8_27b_lora_fsdp2.sh — training entry script
• examples/ascend/train/qwen3_8/qwen3_8_27b_fsdp2.json — Accelerate FSDP2 config

Key points

• Transformers path + FLA GDN (USE_MCORE_GDN=0), with packing=true and max_length=2048
• Requires explicit --model_type qwen3_5 (Qwen3.8 config also matches ovis_ocr2, so auto-detection fails)
• flash-linear-attention commit: ab181c671576fd9cdeed678d216a6273185be700

Environment setup

pip install torch==2.7.1 torch_npu==2.7.1.post4 torchvision==0.22.1 torchaudio==2.7.1
pip install triton-ascend==3.2.1 --extra-index-url=https://mirrors.huaweicloud.com/ascend/repos/pypi
source /home/cann/cann9.1.0/ascend-toolkit/set_env.sh

git clone https://github.com/fla-org/flash-linear-attention.git
pip install -e ./ --no-deps

git clone https://github.com/modelscope/ms-swift.git
pip install -e ./

Usage

bash examples/ascend/train/qwen3_8/qwen3_8_27b_lora_fsdp2.sh

# Use local weights
MODEL=/path/to/Qwen3.8-27B bash examples/ascend/train/qwen3_8/qwen3_8_27b_lora_fsdp2.sh

Results

[zheliuyu]

@hjh0119 @tastelikefeet The failing citest / unittest job looks unrelated to this PR. This change only adds Ascend example scripts; lint and citest-npu both passed.

The GPU unittest failure matches the known CUDA CI issue in #9862: pip install . in ci_container_test.sh re-resolves transformers onto 5.x (and/or later suites get polluted by the StructBERT remote-code install), then imports such as transformers.models.gemma3 fail.

Should we wait for that CI fix to land and then re-run unittest before requesting review, or is it fine to review this PR as-is?