model : support sentence-transformers CrossEncoder heads for ModernBERT

[boshjerns]

My name is Josh. I'm a software developer. I've been working a lot with AI tools, on an iOS app that lets people run models locally on their device. I really care about the security and sovereignty of individuals when they're using LLMs: if they want to interact with them without their interactions leaving their device, they should be able to. I think that's really cool.

The app I'm building is 100% free, and there are no in-app purchases. Long story short, I'm absolutely using AI to create everything that I'm doing at this point. While working on it, I seem to have come across something that would be a real benefit to get into the core infrastructure of llama.cpp, in that it would enable more models to be used better. If this is something that's not welcome, I will totally understand and take that message loud and clear: I'm not well-equipped enough to contribute, and that is actually a fair statement.

That being said, I've been working on this application for over a year now. I'm a huge fan of Georgi Gerganov, if anything as an idea: putting things out in the world that are good for people to use. This is just me trying to be a part of that. I don't mean to step on any toes here. Again, if it's not useful, I will not do this anymore. I'm just working with these tools so much that the agent I'm working with and I came across a few small changes that look like they could actually be of great benefit. For that reason, I'm submitting them, and I acknowledge 100% that this is on the line of a potential ban for contributing. Again, I will take the message loud and clear and never do this again. I just want to be helpful, and this is the way I've found that I can potentially do that.

Thank you for reading my PRs, and just know that there is absolutely a human in the loop, even though I am certainly using my agent to put this PR through.

-josh - this is a continuation off of another PR right before this same one as they're related...

The technical sections below were drafted by my agent from our testing, and I've reviewed them.

Overview

Builds on #29627: only the second commit is new here.

sentence-transformers CrossEncoder models saved with sentence-transformers v5, such as the cross-encoder/ettin-reranker-*-v1 family, keep the scoring head outside the transformer, as separate modules listed in modules.json: Pooling → Dense (GELU) → LayerNorm → Dense. The converter only reads the main model.safetensors, so these models convert without a head and their rank output isn't a relevance score.

This PR:

• detects a CrossEncoder from config_sentence_transformers.json and modules.json, and maps the head modules onto the existing ModernBERT rank head: Dense → cls, LayerNorm → cls_norm, the last Dense → cls_out
• takes the head's pooling from 1_Pooling, using the key added in model : support classifier_pooling for ModernBERT rerankers #29627
• adds an optional cls_norm_b tensor so the LayerNorm bias is applied. Folding the bias into the classifier bias would also work numerically, but only while the classifier directly follows the norm, and it would change the stored weights. Dropping the bias shifts the Ettin 68m scores by about 0.13.
• refuses module chains the rank head can't represent (another activation, a first Dense with a bias, other module orders, a module without model.safetensors) with a clear error instead of writing a broken file

Additional information

Scores for 6 query/passage pairs (f16, llama-server --reranking, CPU) compared with CrossEncoder(...).predict in fp32 (sentence-transformers 5.4.1, transformers 5.7.0):

Model	Max difference
cross-encoder/ettin-reranker-17m-v1	0.0023
cross-encoder/ettin-reranker-32m-v1	0.0041
cross-encoder/ettin-reranker-68m-v1	0.0050
cross-encoder/ettin-reranker-150m-v1	0.0179
cross-encoder/ettin-reranker-400m-v1	0.0073

On master the same models load, but without a head their outputs don't track the reference at all. Q8_0 of the 68m model stays within 0.22 and keeps the same ranking. gte-reranker-modernbert-base and granite-embedding-reranker-english-r2 convert to byte-identical GGUFs with and without this PR.

These checkpoints' tokenizer_config.json names the transformers v5 TokenizersBackend class, so converting them needs transformers 5.x. The rope settings they keep in rope_parameters are already handled by the converter.

Checks: the same as #29627 (warning-free build, flake8, editorconfig-checker, gguf-py tests, test-llama-archs).

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. The code, the verification and the technical sections of this description were done with an AI agent; I reviewed them and I'm responsible for the changes.

[boshjerns]

Closing for now to stay within the one-open-PR limit for new contributors. I'll reopen this once #29627 is done.