keeping shortnames up to date

[juhp]

Feature request description

I am wondering what/how to keep the list of shortnames up-to-date.
It can be challenging due to the fast pace of AI model development.
What is Ramalama general policy here?

Some examples:

• latest Granite is 3.3 (vs 3.1) [4.0 preview even exists now]
• Mistral 3.1 (mistral-small-2503 exists)
• Qwen 3 exists (vs 2.5)
• latest Phi v4 (vs 2)

I guess there is a bit of a trade-off too - stability vs latest-greatest.

Suggest potential solution

I wonder if it makes sense to focus more on hf://ggml-org/ in general say?

Is there any way to provide more unversioned or latest aliases in general?

Have you considered any alternatives?

No response

Additional context

No response

[rhatdan]

We have no great way yet other then community suggesting changes, I don't know if renovate would help here.

I would love to get a way to do better then just plain shortnames, and actually pick out the model which would work best on a particular hardware. But I would need data scientist help for that.

If there are newer released models, please open a PR to update the shortnames.conf file.

Tell me more about hf://ggml-org/

[ericcurtin]

Feature request description

I am wondering what/how to keep the list of shortnames up-to-date. It can be challenging due to the fast pace of AI model development. What is Ramalama general policy here?

Some examples:

• latest Granite is 3.3 (vs 3.1) [4.0 preview even exists now]
• Mistral 3.1 (mistral-small-2503 exists)
• Qwen 3 exists (vs 2.5)
• latest Phi v4 (vs 2)

I guess there is a bit of a trade-off too - stability vs latest-greatest.

Suggest potential solution

I wonder if it makes sense to focus more on hf://ggml-org/ in general say?

Is there any way to provide more unversioned or latest aliases in general?

Have you considered any alternatives?

No response

Additional context

No response

We rely on shortnames and fallback to ollama server-side shortnames.

I mean it's just

ramalama run granite3.3
ramalama run mistral-small3.1

etc. to pull these (they are in ollama).

Generally I wouldn't always default to ggml-org, the problem is the models there aren't pushed by the creators of the model, gemma being an example. Plus ggml-org has just a very small group of models.

Contribute to:

https://github.com/containers/ramalama/blob/main/shortnames/shortnames.conf

if we need more aliases. Fallback to Ollama has benefits, we don't have to maintain such lists. hf just doesn't have this server-side shortnames, so we have to populate a list ourselves, but I recommend only doing as required. Ollama does a great job server-side here.

[ericcurtin]

I do think we should be avoiding aliases like this:

"granite:2b" = "ollama://granite3.1-dense:2b"
"granite:8b" = "ollama://granite3.1-dense:8b"

I mean they had a name already as designated by the creator/author of the model. And now people complain the version is out of date because we brought that burden upon ourselves and can lead to questions like, why dense and not moe?

[rhatdan]

The question I have is could we do better, and have some smarts around when to use dense versus moe (I have no idea what the difference is) But users would like a tool like RamaLama to pick the best one for their particular hardware.

[ericcurtin]

dense vs moe is not a hardware thing, it's just the type of model you choose. Like:

dense
moe
reasoning
etc.

[ericcurtin]

If IBM want to create some super generic one called granite, that's on them I guess, why rename?

[rhatdan]

It is not on IBM, the basic idea is Humans do not know the differenece. Between dense, mode or reasoning. Or versions, or compression levels.

They just heard of granite or llama3 or deepseek.

[github-actions[bot]]

A friendly reminder that this issue had no activity for 30 days.

[rhatdan]

I think is more of a discussion at this point then an issue.

[rjeffman]

@rhatdan but humans do choose a model, so they should either know the difference or test the different models to check which one gives the best outcome (and they would do it, even if they know the difference).

In my opinion, we should keep the same names as the original models. It is kinda weird to ask for "granite" and don't get the latest version, for example. And it may happen in case the configuration file has not been updated yet.