Better log message for the case when error happens inside of TRITONBACKEND_ModelBatcherInitialize()

[lazykyama]

Is your feature request related to a problem? Please describe.

This request is related to a feature introduced since v2.30.0 (23.01). Since this version, users can make their own customized batching strategy.

As far as I checked, it looks like a log message for an error happens in TRITONBACKEND_ModelBatcherInitialize() is confusing and should be improved. For example, at least v2.30.0 shows following message when an error happens during initialization.

E0501 04:51:22.480857 1 model_lifecycle.cc:597] failed to load 'modelname' version 1: Invalid argument:

But, this error message is always almost same regardless of the error reasons (e.g., config parse error, no required parameter missing, etc). That is, a message given into TRITONSERVER_ErrorNew() seems to be ignored.

Describe the solution you'd like

Like other errors, a given message into TRITONSERVER_ErrorNew() is expected to be displayed as a server log.
For example, when an error happens from the code below,

    if (!params.Find("EXAMPLE_LIMIT", &example_limit))
    {
        return TRITONSERVER_ErrorNew(
            TRITONSERVER_ERROR_NOT_FOUND,
            "Unable to find EXAMPLE_LIMIT parameter in model config");
    }

the log message should be like below.

E0428 13:10:46.284968 1717 model_lifecycle.cc:597] failed to load 'length_limit' version 1: Invalid argument: "Unable to find EXAMPLE_LIMIT parameter in model config"

Describe alternatives you've considered

If users could easily understand/identify what error happens, that would be enough.

Additional context

No other additional contexts.

[dyastremsky]

Thanks for your detailed report. It looks like the error is being deleted prematurely here. I created a ticket to fix this bug.

[dyastremsky]

Fixed with #5729.