Explanations about mtmd are needed (critical problem found)

[aropb]

Description

@SignalRT Thanks a lot for MTMD!

I am trying to use qwen3 VL in multimodal mode for a web application (many sessions). MtmdWeights.ClearMedia() is confusing; It turns out that you need to create MtmdWeights for each user? Is it impossible to use one instance for all users?

I would really like to have one instance of the model!

Adding media to executor:
// without adding to _pendingMedia
using SafeMtmdEmbed embed = MtmdWeights.LoadMedia(memoryStream.ToArray());
executor.Embeds.Add(embed);

embed - It has nothing to do with the model, but just a resource.

Ideally, there should be a SafeMtmdEmbed method:
TokenCount()

Is it possible?
Thanks.

[aropb]

I can be use:

SafeMtmdEmbed.FromMediaBuffer(MtmdWeights.NativeHandle, data)

But there is a problem:

class StatefulExecutorBase:
protected Task PreprocessMtmd(string text, InferStateArgs args, bool addBos, bool replaceExisting)

...
this is
var status = ClipModel.Tokenize(prompt, addBos, parseSpecial: true, out chunks);

use embeds from executor
var status = ClipModel.Tokenize(prompt, addBos, parseSpecial: true, Embeds, out chunks);

Only the executor should have embeds.
In addition, I want to store my SafeMtmdEmbed list and insert them into executor without LoadMedia().

There should be nothing related to session resources in the model.
It is better to remove this approach from _pendingMedia (remove all related to _pendingMedia).

[SignalRT]

@aropb I'm working in a PR#1316 to change the web. I will review it next week.

#1316

[aropb]

@SignalRT

Please just clear the list, don't do Dispose() for Embeds!
I may still need them.

    protected void DisposeEmbeds()
    {
        if (Embeds.Count == 0)
            return;

        foreach (var embed in Embeds)
            embed.Dispose();

        Embeds.Clear();
    }

[aropb]

Implemented my own executor