Open ysmu opened 1 year ago
Hi @ysmu, can you elaborate a bit on why is this needed? Are you uploading from local machine to the cluster? or from head node to the worker nodes?
I was hoping this would upload the entire directory from the local/dev machine to the entire cluster. I can't use the git+zip method because my repo has submodules, which isn't included the zip file.
Hi @ysmu , that is known limitation and expected behavior. it is not allowed to use local working_dir in runtime env unless it is used in ray.init(). Internally the issue is also related to the https://github.com/ray-project/ray/issues/30666.
Hi,
I'm facing the same issue (remote cluster + serve app in local) and I was wondering how a local directory can be sent to a cluster without using the git+zip method ?
Thanks,
does this mean that the functionality of ray serve (automatically uploading files mentioned in --working-dir --app-dir) is not available via multi-app deployments?
Hello,
Anyone was able to workaround this?
I built my app as below:
serve build --app-dir "./" ray_serve_app:rent_predictor_app -o rent_predictor_app_config.yaml
Then I tried to deploy the app as below:
serve deploy --address "http://kuberay-head-svc.kuberay:8265" rent_predictor_app_config.yaml
As I am not able to pass working dir as "./", the module ray_serve_app
is missing in the Ray cluster.
What is the right way to handle this?
For me anything other than providing a zip AND using the FILE:// scheme failed.
Am 26.03.2024 um 01:31 schrieb Sercan Tekin @.***>:
Hello,
Anyone was able to workaround this?
I built my app as below:
serve build --app-dir "./" ray_serve_app:rent_predictor_app -o rent_predictor_app_config.yaml Then I tried to deploy the app as below:
serve deploy --address "http://kuberay-head-svc.kuberay:8265" rent_predictor_app_config.yaml As I am not able to pass working dir as "./", the module ray_serve_app is missing in the Ray cluster. What is the right way to handle this?
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.
For me anything other than providing a zip AND using the FILE:// scheme failed. … Am 26.03.2024 um 01:31 schrieb Sercan Tekin @.***>: Hello, Anyone was able to workaround this? I built my app as below: serve build --app-dir "./" ray_serve_app:rent_predictor_app -o rent_predictor_app_config.yaml Then I tried to deploy the app as below: serve deploy --address "http://kuberay-head-svc.kuberay:8265" rent_predictor_app_config.yaml As I am not able to pass working dir as "./", the module ray_serve_app is missing in the Ray cluster. What is the right way to handle this? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.
Thank you @chris-aeviator.
After your suggestion, I tried the following and it worked.
I executed ray.init
with working dir option between serve build
and serve deploy
. It failed as it is deprecated and not stable, but it pushed the files to GCS and printed URI anyways.
I specified the mentioned URI in my config file, and now the deployment is not good.
What happened + What you expected to happen
I have a deployment file that specifies the
.
as the working directory. Following the documentation fromI was expecting
serve deploy deployment.yml
to upload the current directory to all workers in the cluster. However, I'm hitting some exceptions. Full repro is below. Is this a known limitation in ray serve?Versions / Dependencies
Reproduction script
Issue Severity
Medium: It is a significant difficulty but I can work around it.