AMD SDXL changes resolution and steps

[HemanLin-cl]

Hi all,

I've tried to download nexa-cli_windows_x86_64.exe and install and then execute the

nexa infer NexaAI/sdxl-turbo-amd-npu --prompt "a cartoon cat"

It can run with NPU usage under AMD Ryzen AI 9 HX370. However, I've found it cannot set the arguments to change the output resolution and steps cause I want to test 1024x1024 actually

Therefore, I try to use "pip install nexaai" instead
And then I implement a python file to run the model.

from nexaai import ImageGen, ModelConfig
imagegen = ImageGen.from_(model="NexaAI/sdxl-turbo-amd-npu", plugin_id="cpu_gpu", device_id="gpu")
result = imagegen.txt2img(prompt="a beautiful mountain landscape, sunset, 4k",
output_path="test_sdxl_amd.png",
height=1024,
width=1024,
steps=5,
guidance_scale=7.5)

After running this python file, it shows: (plugin_id="nexaai" also shows the errors)

[DEBUG] [model_hub.py:449:ensure_model_downloaded] Model NexaAI/sdxl-turbo-amd-npu already exists locally
[DEBUG] [model_loader.py:212:_select_model_file] Selected quant: N/A, file: lib/xclbin/stx/SD15_SD3_VAE_2x4x4.xclbin
[DEBUG] [image_gen.py:68:init] is_native_plugin(nexa_dml) = False
[DEBUG] [image_gen.py:96:init] Plugin nexa_dml is not a native plugin, using C interface
[WARN ] [core.py:148:_log_callback] Plugin id 'nexa_dml' is deprecated, use 'nexaai' instead
[WARN ] [core.py:148:log_callback] Failed to list directory contents: directory_iterator::directory_iterator: The directory name is invalid.: "C:\Users.cache\nexa.ai\nexa_sdk\models\NexaAI\sdxl-turbo-amd-npu\lib\xclbin\stx\SD15_SD3_VAE_2x4x4.xclbin"
[ERROR] [core.py:148:log_callback] Required file model_index.json not found in model directory: C:\Users.cache\nexa.ai\nexa_sdk\models\NexaAI\sdxl-turbo-amd-npu\lib\xclbin\stx\SD15_SD3_VAE_2x4x4.xclbin/model_index.json
Traceback (most recent call last):
File "C:\Users\Heman\AIPC\nexa\nexa_infer.py", line 8, in
imagegen = ImageGen.from(model="NexaAI/sdxl-turbo-amd-npu", plugin_id="cpu_gpu", device_id="gpu") #nexaai
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\isaac\miniconda3\Lib\site-packages\nexaai\internal\model_loader.py", line 27, in from
return cls._from_huggingface_repo(model, quant, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\isaac\miniconda3\Lib\site-packages\nexaai\internal\model_loader.py", line 97, in _from_huggingface_repo
return cls._create_instance(**params)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\isaac\miniconda3\Lib\site-packages\nexaai\image_gen.py", line 261, in _create_instance
return cls(**params)
^^^^^^^^^^^^^
File "C:\Users\isaac\miniconda3\Lib\site-packages\nexaai\image_gen.py", line 135, in init
check_error(error_code)
File "C:\Users\isaac\miniconda3\Lib\site-packages\nexaai\nexa_sdk\error.py", line 131, in check_error
raise exception_class(error_code)
nexaai.nexa_sdk.error.NexaModelLoadError: [-100201] Model loading failed
[DEBUG] [core.py:148:_log_callback] deinitializing ml

Is there any sample code we can run sdxl-turbo-amd-npu by python API?
or is there any method I can test using other resolution and steps under AMD?

I also try to let the directory_iterator::directory_iterator: be "C:\Users.cache\nexa.ai\nexa_sdk\models\NexaAI" and then the model_index.json can be read but it still cannot generate the images.

Thanks