What it takes to support bidirectional Llama3 for LLM2Vec

[nv-jsaito]

Can someone help me understand llama.cpp modifications needed to support Llama3-based LLM2Vec?

Background

We have a project making use of Llama3-based LLM2Vec as text embedding for a text-to-content model. While the model is trained with the original Llama3-based LLM2Vec on HF, we would like to make the deployment easier by running LLM2Vec with llama.cpp.

AFAIK, LLM2Vec should be just finetuning on top of the original LLM architecture (Llama3 in our case) but with bidirectional attention. Because llama.cpp supports Llama3, I am hoping running Llama3-based LLM2Vec with llama.cpp is not too much work.

What has been done so far

• Merged LLM2Vec finetuned parameters to the original Meta-Llama-3-8B-Instruct and saved as an HF model
  • I confirmed this merged model outputs the same exact results as the original non-merged model
• Converted the above to GGUF with a modified convert_hf_to_gguf.py like below to set the flag for non-causal attention

@Model.register("LlamaBiModel")
class LlamaBiModel(LlamaModel):
    model_arch = gguf.MODEL_ARCH.LLAMA
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        self.gguf_writer.add_causal_attention(False)

• Made LLM_ARCH_LLAMA be aware of the optional causal attention flag in llama-model.cpp

Observations

• The embeddings from above seems to be garbage
• Interestingly, when I turn on the causal attention, the model performs close to expectations while losing some details in the text prompts
• With or without the causal attention, the embeddings are surely significantly off numerically from the original HF model

Let me know what I am missing here and/or ideas to try.

[ggerganov]

There could be different causes:

• Tokenization problem
• Incorrect prompt format
• Incorrect pooling type (most likely)
• etc.

Could you provide a basic example using llama-server + curl where you generate the embeddings of some text and it does not match a reference result that you expect? This will help find the root cause of the issue.

Try use this and see if the results are good:

llama-server -m model.gguf --pooling last

curl http://localhost:8080/embeddings -H "Content-Type: application/json" -d '{"input": ["hello"]}'