Bug: Invalid Embeddings if GPU offloaded (CUDA)

[adrianliechti]

Update:
Disabling GPU Offloading (--n-gpu-layers 83 to --n-gpu-layers 0) seems to "fix" my issue with Embeddings.

Dear Llama Community,

I might need a hint about embeddings API on the (example)server. And already say thanks a lot for taking your time and any help

I'm sending a string (e.g. "hello") to the /embedding endpoint of the llama-example-server against a 70b llama2 model.
In the raw result, I see this json:

{"embedding":[3.1069589551009844e-41,4.1016006050787396e-42,4.736388809417882e-43,3.8956097308229915e-43,5.1834030195374983e-42,4.200111887120774e-41,1.0165019060212223e-41,4.1883409800204457e-41,0.0883883461356163,7.370829922348538e-43,3.685414961174269e-43,1.1832564232758755e-41,4.188761369559743e-41,4.75040179406113e-42,1.8483126744444337e-42,4.512181055125911e-43,0.0,0.0,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,1.401298464324817e-45,0.0,0.0,null,null,null,null,null,0.009384572505950928,-0.029291629791259766,0.007542848587036133,0.025960087776184082,-0.005349218845367432,0.014909744262695313,0.007542848587036133,-0.0035074949264526367,-0.0016657710075378418,-0.0163995623588562,-0.012716114521026611,...

My goal is to covert this to an OpenAI compatible format and use it with langchain and a chroma db.
(see here for a llama.cpp-to-openai-server: https://github.com/adrianliechti/llama/blob/main/llama-openai/provider/llama/llama.go)

Currently the vector db search returns no results when using the llama-embeddings (but works fine using the openai embeddings).

My assumption is that I somehow have to convert or format this result... I am not very sure about the "null" since I haven' see such values in OpenAI results...

Do these embeddings number look normal or weird to you? do you have a hint how to proper convert them? If i parse them as float32 in golang, the "null" would get "0". would that make sense?

Thanks a ton!! Adrian

[FSSRepo]

For a long time, it has been known that llama models are not good at generating usable embeddings. Furthermore, it seems that the server example provides incorrect embeddings (in comparison to the example/embedding).

[adrianliechti]

@FSSRepo thanks for the fast reply! and confirming my wild guess.

at this point i would be more than happy if i get usable embeddings. but i would follow your hint to use a different model for embeddings later. thanks for that

[FSSRepo]

@adrianliechti I am working on improving the server example, but the generation of embeddings is not working for me, and I haven't found the reason yet. #3589

[shibe2]

For usable embeddings, you need an encoder. Many LLMs these days, including Llama, don't have encoders. For local inference, I'm using one of these: https://mteb-leaderboard.hf.space/?__theme=dark

[adrianliechti]

@shibe2
This seems convincing to me - also considering that e.g. OpenAI uses a different model for embedding...
I will experiment with another embedder directly my code instead of the llama.cpp endpoint then.

Or is there an encoding/tokenize model to use in llama.cpp? Or what is the /embedding endpoint for another usecase and I just made a wrong assumption?

[adrianliechti]

@FSSRepo I press all my thumbs for you and am curious for the result.
Since my c++ is not where it needs to be for this project - I probably only can help with testing / validate your version.

[adrianliechti]

I tested the very same with the python-llama-cpp version... and here the result looks much better in a very first small test!

docker run --rm -it -p 8080:8000 -v $(pwd):/models -e MODEL=/models/codellama-7b-instruct.gguf ghcr.io/abetlen/llama-cpp-python:latest

since this is just a wrapper, I have a feeling that having usable embeddings in the example server is feasible :)

[adrianliechti]

hmmmm new observation

1. llama-2-70b-chat.Q5_K_M.gguf running on a Nvidia A100 -> embeddings contain these weird null fields / long values
2. codellama-7b-instruct.Q4_K_M.gguf locally on a MacBook M1 -> embeddings look good / seem to work with chromadb

[adrianliechti]

I can confirm that --n-gpu-layers 0 changed the result on embeddings for the better...

[FSSRepo]

@adrianliechti I can confirm this, the embedding example too have this problem.

embedding.exe -m ..\models\vicuna-7b.gguf -p "Which is your dog?"

Embedding with -ngl 0 (correct):

2.071708 0.094213 -0.879323 -0.868794 -1.877190 0.149166 0.456107 -1.180170 -1.899972 0.324895 0.137798 1.375898 -0.929520 -0.724631 -3.895488 0.835485 2.534587 -1.334827 -1.068944 -1.522295 -2.574473 -1.089735 0.195261 0.207192 0.607332 -0.160022 

Embedding with -ngl 33:

0.001389 0.025688 -0.022911 0.001389 0.001389 -0.022911 0.001389 0.001389 0.001389 -0.022911 -0.022911 -0.022911 0.001389 0.001389 0.025688 0.025688 0.001389 0.001389 0.001389 -0.022911 0.025688 0.001389 -0.047211 0.001389 -0.054474 0.013565 -0.020454 0.013565 -0.020454

Possibly the embeddings tensor is not being downloaded from the GPU.

[ggerganov]

There is a TODO referencing the embeddings layer, but I don't recall what it means:

https://github.com/ggerganov/llama.cpp/blob/e74c705e15cd228ad696c4a3cdea6d6fb4ff434c/llama.cpp#L3438-L3441

We are intentionally not offloading that layer (see that the offload_func_nr is commented out) in order to be able to read the embeddings, but there might be some issue still.