AI Toolkit para VS Code reúne varios modelos del Catálogo de Azure AI Studio y otros catálogos como Hugging Face. Esta herramienta simplifica las tareas comunes de desarrollo para construir aplicaciones de IA con herramientas y modelos de IA generativa a través de:
- Comenzar con el descubrimiento de modelos y un entorno de pruebas.
- Ajuste fino e inferencia de modelos utilizando recursos de computación locales.
- Ajuste fino e inferencia remota utilizando recursos de Azure.
Instalar AI Toolkit para VSCode
[Vista previa privada] Aprovisionamiento con un clic para Azure Container Apps para realizar ajuste fino e inferencia de modelos en la nube.
Ahora comencemos con el desarrollo de tu aplicación de IA:
- Asegúrate de que el controlador de NVIDIA esté instalado en el host.
- Ejecuta
huggingface-cli loginsi estás utilizando HF para la utilización del conjunto de datos. Oliveexplica configuraciones clave para cualquier cosa que modifique el uso de memoria.
Dado que usamos un entorno WSL compartido, necesitas activar manualmente el entorno conda. Después de este paso, puedes ejecutar el ajuste fino o la inferencia.
conda activate [conda-env-name] Si solo deseas probar el modelo base sin ajuste fino, puedes ejecutar este comando después de activar conda.
cd inference
# Web browser interface allows to adjust a few parameters like max new token length, temperature and so on.
# User has to manually open the link (e.g. http://0.0.0.0:7860) in a browser after gradio initiates the connections.
python gradio_chat.py --baseonlyUna vez que el espacio de trabajo esté abierto en un contenedor de desarrollo, abre un terminal (la ruta predeterminada es la raíz del proyecto) y ejecuta el siguiente comando para ajustar un LLM al conjunto de datos seleccionado.
python finetuning/invoke_olive.py Los puntos de control y el modelo final se guardarán en models folder.
Next run inferencing with the fune-tuned model through chats in a console, web browser or prompt flow.
cd inference
# Console interface.
python console_chat.py
# Web browser interface allows to adjust a few parameters like max new token length, temperature and so on.
# User has to manually open the link (e.g. http://127.0.0.1:7860) in a browser after gradio initiates the connections.
python gradio_chat.pyPara usar prompt flow in VS Code, please refer to this Quick Start.
Next, download the following model depending on the availability of a GPU on your device.
To initiate the local fine-tuning session using QLoRA, select a model you want to fine-tune from our catalog.
| Platform(s) | GPU available | Model name | Size (GB) |
|---|---|---|---|
| Windows | Yes | Phi-3-mini-4k-directml-int4-awq-block-128-onnx | 2.13GB |
| Linux | Yes | Phi-3-mini-4k-cuda-int4-onnx | 2.30GB |
| Windows Linux |
No | Phi-3-mini-4k-cpu-int4-rtn-block-32-acc-level-4-onnx | 2.72GB |
Note You do not need an Azure Account to download the models
The Phi3-mini (int4) model is approximately 2GB-3GB in size. Depending on your network speed, it could take a few minutes to download.
Start by selecting a project name and location. Next, select a model from the model catalog. You will be prompted to download the project template. You can then click "Configure Project" to adjust various settings.
We use Olive to run QLoRA fine-tuning on a PyTorch model from our catalog. All of the settings are preset with the default values to optimize to run the fine-tuning process locally with optimized use of memory, but it can be adjusted for your scenario.
- Fine tuning Getting Started Guide
- Fine tuning with a HuggingFace Dataset
- Fine tuning with Simple DataSet
- To run the model fine-tuning in your remote Azure Container App Environment, make sure your subscription has enough GPU capacity. Submit a support ticket to request the required capacity for your application. Get More Info about GPU capacity
- If you are using private dataset on HuggingFace, make sure you have a HuggingFace account and generate an access token
- Enable Remote Fine-tuning and Inference feature flag in the AI Toolkit for VS Code
- Open the VS Code Settings by selecting File -> Preferences -> Settings.
- Navigate to Extensions and select AI Toolkit.
- Select the "Enable Remote Fine-tuning And Inference" option.
- Reload VS Code to take effect.
- Execute the command palette
AI Toolkit: Focus on Resource View. - Navigate to Model Fine-tuning to access the model catalog. Assign a name to your project and select its location on your machine. Then, hit the "Configure Project" button.
- Project Configuration
- Avoid enabling the "Fine-tune locally" option.
- The Olive configuration settings will appear with pre-set default values. Please adjust and fill in these configurations as required.
- Move on to Generate Project. This stage leverages WSL and involves setting up a new Conda environment, preparing for future updates that include Dev Containers.
- Click on "Relaunch Window In Workspace" to open your remote development project.
Note: The project currently works either locally or remotely within the AI Toolkit for VS Code. If you choose "Fine-tune locally" during project creation, it will operate exclusively in WSL without remote development capabilities. On the other hand, if you forego enabling "Fine-tune locally", the project will be restricted to the remote Azure Container App environment.
To get started, you need to provision the Azure Resource for remote fine-tuning. Do this by running the AI Toolkit: Provision Azure Container Apps job for fine-tuning from the command palette.
Monitor the progress of the provision through the link displayed in the output channel.
If you're using private HuggingFace dataset, set your HuggingFace token as an environment variable to avoid the need for manual login on the Hugging Face Hub.
You can do this using the AI Toolkit: Add Azure Container Apps Job secret for fine-tuning command. With this command, you can set the secret name as HF_TOKEN and use your Hugging Face token as the secret value.
To start the remote fine-tuning job, execute the AI Toolkit: Run fine-tuning command.
To view the system and console logs, you can visit the Azure portal using the link in the output panel (more steps at View and Query Logs on Azure). Or, you can view the console logs directly in the VSCode output panel by running the command AI Toolkit: Show the running fine-tuning job streaming logs.
Note: The job might be queued due to insufficient resources. If the log is not displayed, execute the
AI Toolkit: Show the running fine-tuning job streaming logscommand, wait for a while and then execute the command again to re-connect to the streaming log.
During this process, QLoRA will be used for fine-tuning, and will create LoRA adapters for the model to use during inference. The results of the fine-tuning will be stored in the Azure Files.
After the adapters are trained in the remote environment, use a simple Gradio application to interact with the model.
Similar to the fine-tuning process, you need to set up the Azure Resources for remote inference by executing the AI Toolkit: Provision Azure Container Apps for inference from the command palette.
By default, the subscription and the resource group for inference should match those used for fine-tuning. The inference will use the same Azure Container App Environment and access the model and model adapter stored in Azure Files, which were generated during the fine-tuning step.
If you wish to revise the inference code or reload the inference model, please execute the AI Toolkit: Deploy for inference command. This will synchronize your latest code with Azure Container App and restart the replica.
Once deployment is successfully completed, you can access the inference API by clicking on the "Go to Inference Endpoint" button displayed in the VSCode notification. Or, the web API endpoint can be found under ACA_APP_ENDPOINT in ./infra/inference.config.json y en el panel de salida. Ahora estás listo para evaluar el modelo utilizando este punto de acceso.
Para más información sobre el desarrollo remoto con AI Toolkit, consulta la documentación de Ajuste fino de modelos de forma remota e Inferencia con el modelo ajustado.
Descargo de responsabilidad:
Este documento ha sido traducido utilizando servicios de traducción basados en inteligencia artificial. Si bien nos esforzamos por garantizar la precisión, tenga en cuenta que las traducciones automatizadas pueden contener errores o imprecisiones. El documento original en su idioma nativo debe considerarse la fuente autorizada. Para información crítica, se recomienda una traducción profesional realizada por humanos. No nos hacemos responsables de malentendidos o interpretaciones erróneas que puedan surgir del uso de esta traducción.