[
https://issues.apache.org/jira/browse/FLINK-40692?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Danish updated FLINK-40692:
---------------------------
Description:
When we don't define `model-version` in `CREATE MODEL` SQL, it tends to create
URL with `version: latest`. i.e.
{code:java}
https://{model endpoint}/v2/models/{model name}/versions/latest/infer{code}
This is not necessarily the case for all model servers in KServe protocol and
it fails for them. The protocol defines that if there is no version defined, we
fall back to using the default model defined by the server itself. i.e. the url
does not contain any version info
{code:java}
https://{model endpoint}/v2/models/{model name}/infer{code}
The server model can define either the latest or any other model version to be
used when no version is mentioned using server side policies.
was:
When we don't define `model-version` in `CREATE MODEL` SQL, it tends to create
URL with `version: latest`. i.e.
https://\{model endpoint}/v2/models/\{model name}/versions/latest/infer
This is not necessarily the case for all model servers in KServe protocol and
it fails for them. The protocol defines that if there is no version defined, we
fall back to using the default model defined by the server itself. i.e. the url
does not contain any version info
https://\{model endpoint}/v2/models/\{model name}/infer
The server model can define either the latest or any other model version to be
used when no version is mentioned using server side policies.
> [Model] Use default model version from server when no version is defined in
> model config
> ----------------------------------------------------------------------------------------
>
> Key: FLINK-40692
> URL: https://issues.apache.org/jira/browse/FLINK-40692
> Project: Flink
> Issue Type: Improvement
> Reporter: Danish
> Priority: Major
>
> When we don't define `model-version` in `CREATE MODEL` SQL, it tends to
> create URL with `version: latest`. i.e.
> {code:java}
> https://{model endpoint}/v2/models/{model name}/versions/latest/infer{code}
>
> This is not necessarily the case for all model servers in KServe protocol and
> it fails for them. The protocol defines that if there is no version defined,
> we fall back to using the default model defined by the server itself. i.e.
> the url does not contain any version info
> {code:java}
> https://{model endpoint}/v2/models/{model name}/infer{code}
> The server model can define either the latest or any other model version to
> be used when no version is mentioned using server side policies.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
