Problem on Qwen3-VL lora SFT

[huaiyizhao]

[rank1]: Traceback (most recent call last):
[rank1]: File "/app/ms-swift/swift/cli/sft.py", line 10, in
[rank1]: sft_main()
[rank1]: File "/app/ms-swift/swift/llm/train/sft.py", line 340, in sft_main
[rank1]: return SwiftSft(args).main()
[rank1]: ^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/app/ms-swift/swift/llm/base.py", line 49, in main
[rank1]: result = self.run()
[rank1]: ^^^^^^^^^^
[rank1]: File "/app/ms-swift/swift/llm/train/sft.py", line 178, in run
[rank1]: self.model = self.prepare_model(self.args, self.model, template=self.template, train_dataset=train_dataset)
[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/app/ms-swift/swift/llm/train/tuner.py", line 350, in prepare_model
[rank1]: model = prepare_adapter(
[rank1]: ^^^^^^^^^^^^^^^^
[rank1]: File "/app/ms-swift/swift/llm/train/tuner.py", line 204, in prepare_adapter
[rank1]: model = Swift.prepare_model(model, lora_config)
[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/app/ms-swift/swift/tuners/base.py", line 716, in prepare_model
[rank1]: return get_peft_model(model, config, **kwargs)
[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/usr/local/lib/python3.11/site-packages/peft/mapping_func.py", line 125, in get_peft_model
[rank1]: return MODEL_TYPE_TO_PEFT_MODEL_MAPPING[peft_config.task_type](
[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/usr/local/lib/python3.11/site-packages/peft/peft_model.py", line 1815, in init
[rank1]: super().init(model, peft_config, adapter_name, **kwargs)
[rank1]: File "/usr/local/lib/python3.11/site-packages/peft/peft_model.py", line 130, in init
[rank1]: self.base_model = cls(model, {adapter_name: peft_config}, adapter_name)
[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/app/ms-swift/swift/tuners/peft.py", line 304, in new_init
[rank1]: self.init_origin(model, config, adapter_name)
[rank1]: File "/usr/local/lib/python3.11/site-packages/peft/tuners/tuners_utils.py", line 209, in init
[rank1]: self.inject_adapter(self.model, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage, state_dict=state_dict)
[rank1]: File "/usr/local/lib/python3.11/site-packages/peft/tuners/tuners_utils.py", line 578, in inject_adapter
[rank1]: self._create_and_replace(
[rank1]: File "/app/ms-swift/swift/tuners/peft.py", line 103, in _create_and_replace_hook
[rank1]: return self._create_and_replace_origin(peft_config, adapter_name, target, *args, **kwargs)
[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/usr/local/lib/python3.11/site-packages/peft/tuners/lora/model.py", line 259, in _create_and_replace
[rank1]: new_module = self._create_new_module(lora_config, adapter_name, target, device_map=device_map, **kwargs)
[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/usr/local/lib/python3.11/site-packages/peft/tuners/lora/model.py", line 365, in _create_new_module
[rank1]: new_module = dispatcher(target, adapter_name, lora_config=lora_config, **kwargs)
[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: File "/usr/local/lib/python3.11/site-packages/peft/tuners/lora/awq.py", line 105, in dispatch_awq
[rank1]: from awq.modules.linear import WQLinear_GEMM
[rank1]: File "/usr/local/lib/python3.11/site-packages/awq/init.py", line 24, in
[rank1]: from awq.models.auto import AutoAWQForCausalLM
[rank1]: File "/usr/local/lib/python3.11/site-packages/awq/models/init.py", line 1, in
[rank1]: from .mpt import MptAWQForCausalLM
[rank1]: File "/usr/local/lib/python3.11/site-packages/awq/models/mpt.py", line 1, in
[rank1]: from .base import BaseAWQForCausalLM
[rank1]: File "/usr/local/lib/python3.11/site-packages/awq/models/base.py", line 49, in
[rank1]: from awq.quantize.quantizer import AwqQuantizer
[rank1]: File "/usr/local/lib/python3.11/site-packages/awq/quantize/quantizer.py", line 11, in
[rank1]: from awq.quantize.scale import apply_scale, apply_clip
[rank1]: File "/usr/local/lib/python3.11/site-packages/awq/quantize/scale.py", line 12, in
[rank1]: from transformers.activations import NewGELUActivation, PytorchGELUTanh, GELUActivation
[rank1]: ImportError: cannot import name 'PytorchGELUTanh' from 'transformers.activations' (/usr/local/lib/python3.11/site-packages/transformers/activations.py)

Uninstall autoawq works for me. Maybe we need a more specific environment requirements for training Qwen3-VL or a docker image.
hiyouga/LLaMA-Factory#9247

[hjh0119]

https://swift.readthedocs.io/en/latest/BestPractices/Qwen3-VL-Best-Practice.html#environment-setup

[hjh0119]

AutoAWQ is no longer maintained (https://github.com/casper-hansen/AutoAWQ),

so compatibility issues with Transformers may arise.

[huaiyizhao]

It seems the swift 3.8.3 docker image still has autoawq preinstalled.

[Jintao-Huang]

Try the image provided by ms-swift 3.9.1

[wgzhendong]

Try the image provided by ms-swift 3.9.1

I encountered this error even when using this image： modelscope-registry.cn-beijing.cr.aliyuncs.com/modelscope-repo/modelscope:ubuntu22.04-cuda12.8.1-py311-torch2.8.0-vllm0.11.0-modelscope1.31.0-swift3.9.3

ImportError: cannot import name 'PytorchGELUTanh' from 'transformers.activations' (/usr/local/lib/python3.11/site-packages/transformers/activations.py)

[Jintao-Huang]

pip uninstall autoawq

[wgzhendong]

pip uninstall autoawq
This image does not have autoqwq.