Fetch from hub #1160
Fetch from hub #1160
Conversation
…get model hub infos.
|
Could you please paste some screenshots to show your PR effect? |
|
Add the "Model Source" option in "Register Model" Page. "Self Hosted" is same as the original design, "Hub" is the new function for registering new model by model hub and model id directly.
If the model format is "GGML" or "GGUF", the extra fields "Model File Name Template" and "Model File Name Split Template" will show out. If "Model File Name Split Template" is not empty, the extra "Quantization Parts" field shows.
When the correct model id and model hub are provided, click "IMPORT MODEL" button, it will fetch the model info and fill the corresponding fields, like "Context Length", "Model Size in Billions", "Model File Name Template", "Model File Name Split Template", "Quantization" and "Quantization Parts", you can modify them if there is anything wrong.
When the other options are selected correctly, the model can be registered like a built-in model, you do not download it until the first running. |
|
add the support for embedding and rerank model register from hub
|
… run on python < 3.9
Add the function that user can register custom the model from model hub (
HuggingFace,ModelScope) by selecting model type Hub, and input the model Id, and user can import basic info like context length, model size, quantization by click "Import Model" button.The model will be download from model hub at first running.
TODO: the cached model will not show (cached) tag in model card in "Launch Model" page.
在自定义模型时加入了从Model Hub下载的功能，只要选择Model Source是Hub，并从HuggingFace或ModelScope中选择一个，然后指定Model Id就能注册新模型，并且可以通过点击"Import Model"从Hub中导入诸如context length， model size和量化信息。
注册后的模型会在第一次启动时从Model Hub下载。
TODO： 在"Launch Model"页面的模型卡片中，自定义模型在模型下载后不会显示cached标签。