... to make it easier to read Vespa documentation on an e-reader / offline
Vespa documentation is generated using Jekyll from .md and .html files, look into options for generating the artifact as part of site generation (there might be plugins we can use here)
The documentation for creating complex inference graphs should include what components can be connected to what, and how the overall graph should behave.
For example, all inference graphs need to end with a single node, whether its a combiner combining multiple inputs, or a transformer, or a model--the graph can't split and then never rejoin, etc.
In a multi-user system, this line can raise Permission Error because of the same file name in the shared /tmp/ directory which maybe created by another user. Maybe we can give more unique name when we store it using timestamp or something.
ReloadableInferEngine::proc_initialize中使用TaskExecutor的是个单例，那么多个模型会一次去设置、启动同一个TaskExecutor；
在调用schedule()之后，TaskExecutor如何确定去调用哪个模型呢？
The format of current master branch UDF output, eva-reuse and my SSD object detector is different. Need a sync format.
Add a description, image, and links to the serving topic page so that developers can more easily learn about it.
To associate your repository with the serving topic, visit your repo's landing page and select "manage topics."
Description
There are multiple user requests of using GraphNN data (node and edge lists) as sample batches into a custom RLlib model.
https://discuss.ray.io/t/rllib-variable-length-observation-spaces-without-padding/726
https://discuss.ray.io/t/working-with-graph-neural-networks-varying-state-space/5730/2
The recommended method today is to use Repeated observation space and VariableVal