Repository navigation
PaddleX hps server pipeline.yaml contains server_url #5214
Description
Following these instructions, I grabbed the various services involved in paddle hps orchestration - the gateway, pipeline and vlm. I've built separate images for deploying this via kubernetes. For the pipeline, the pipeline.yaml configuration file at hps/sdk/server has a setting for Submodules > VLRecognition > gen_ai_config > server_url.
Is this setting overridable via environment variable or does this setting need to be separately set for every different environment prior to image packaging? Or is the code that I have old and needs to be updated?
Please advise as this having to be configured as an absolute value via a config file that gets packaged into docker images make managing deployments more challenging.
Thank you for your feedback and insights!
Activity
jstorq commentedon Oct 6, 2026
jstorq commentedon Oct 10, 2026
@Bobholamovic Separately I found Bundled PaddleX hps server page, which appears to bundle the triton inference server with the sdk.
Originally, I had been working from this set of instructions which I've been able to deploy as three separate images - the gateway api, the triton inference server and the vlm. But it would be preferable, if that is what the bundled server instructions imply, to simply run these both together if that works, sharing one GPU for the inference and vlm handling. Is that what the
/release/3.7/deploy/hpsis saying?So, I am a bit confused. Trying to create the images for the bundled server we are running into issues with old Python 3.10 and a version of Ubuntu (22.10) nearing end of life. Is the bundled server going to be callable at the /layout-parsing endpoint? Would I need to build the image for the bundled server only and still use the api gateway but point both the triton server and vlm server to the running bundled server container?
If there is any clarity you could provide, I would really appreciate it. Thank you!