Support conversion of Qwen3-Embedding models

[iamlemec]

This adds GGUF conversion support for the Qwen3-Embedding class of models and makes pooling work properly by default. I'm not sure how the official Qwen GGUFs were produced, but I'll try to get them updated if this gets merged. More details:

• Though the pre-tokenizer is in fact the same qwen2 as usual, the HF tokenizer adds an EOT token that makes the checksum different from the usual text generation models
• For some reason the tensor names are not prefixed with model. like most other models in this class
• Actually loads the default pooling mode so it works out of the box

[ggerganov]

Thanks for looking into this - it would be great to get this model supported.

[CISC]

Ooops, was a bit trigger happy, forgot convert_hf_to_gguf_update.py...

@iamlemec care to add the model there in a new PR?

Edit: It has to go in pre_computed_hashes (since qwen2 is in models already), which means the entry in convert_hf_to_gguf.py will be moved towards the top as well.

[iamlemec]

@CISC Ah I see, will have that up in a moment. See #15030.

[Mushoz]

@iamlemec Did you ever speak to Qwen about the official GGUFs? Are they okay to use in their current form or would they have to be updated?

[iamlemec]

@Mushoz looks like they uploaded new GGUFs to HF! Just tested the 0.6B version and it works as expected with correct pooling.

[Mushoz]

@iamlemec which ones are you talking about?

I am looking at this one: https://huggingface.co/Qwen/Qwen3-Embedding-0.6B-GGUF/tree/main
But that has been updated 14th of July, way before this PR was merged.

Same holds true for the 8b version: https://huggingface.co/Qwen/Qwen3-Embedding-8B-GGUF/tree/main

[iamlemec]

@Mushoz Huh, you're right. Maybe I missed that update before and was still using the old ones? Anyway, the ones you linked to have the pooling type properly specified and seem to work as expected!