LlamaCpp crash when embedding (in beta)

[vodkaslime]

Issue description

LlamaCpp crash when embedding

Expected Behavior

The code generates correct embedding vector.

Actual Behavior

LlamaCpp crashed with error code:

zsh: segmentation fault  node server/test.js

Steps to reproduce

Download the embedding model from https://huggingface.co/CompendiumLabs/bge-large-en-v1.5-gguf/tree/main

Run the following code:

import { getLlama } from 'node-llama-cpp';

const modelPath = '/path/to/bge-large-en-v1.5-q8_0.gguf';

const llama = await getLlama();
const model = await llama.loadModel({
  modelPath,
});

const embeddingContext = await model.createEmbeddingContext();
const text = 'Hello world';
const embedding = await embeddingContext.getEmbeddingFor(text);
console.log(text, embedding.vector);

My Environment

Dependency	Version
Operating System	MacOS
CPU	Apple M1
Node.js version	"node-llama-cpp": "^3.0.0-beta.17",
Typescript version	N/A
node-llama-cpp version	"node-llama-cpp": "^3.0.0-beta.17",

Additional Context

No response

Relevant Features Used

• [X] Metal support
• [ ] CUDA support
• [ ] Grammar

Are you willing to resolve this issue by submitting a Pull Request?

Yes, I have the time, but I don't know how to start. I would need guidance.

[vodkaslime]

Confirmed that the embedding functionality works when using model functionary-small-v2.2.q4_0.gguf, but not working with bge or acge models.

It might be worth noting that when loading bge and acge models, the LlamaCpp would output warning like llm_load_vocab: mismatch in special tokens definition ( 8754/21128 vs 5/21128 ). But even with this warning, the embedding still works on these models when using python counterpart python-llama-cpp. So it should not be problem with models.

[chadkirby]

I've run into the same problem with prithivida/all-MiniLM-L6-v2-gguf, CompendiumLabs/bge-small-en-v1.5-gguf, and nomic-ai/nomic-embed-text-v1.5-GGUF.

[giladgd]

I found the cause for this issue and will release a new version with the fix in the next few days. Thanks for reporting this issue!

[github-actions[bot]]

:tada: This issue has been resolved in version 3.0.0-beta.18 :tada:

The release is available on:

• npm package (@beta dist-tag)
• GitHub release
• v3.0.0-beta.18

Your semantic-release bot :package::rocket:

[bitterspeed]

Edit: found this issue in llama.cpp:

Hello, while I confirm this fixes Mac Os BGE models, this causes a crash on Windows. Running the test code above with bge-large-en-v1.5-q4_k_m.gguf causes the following error message:

[node-llama-cpp] llm_load_vocab: mismatch in special tokens definition ( 7104/30522 vs 5/30522 ).
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: NVIDIA GeForce RTX 3080 | uma: 0 | fp16: 1 | warp size: 32
WARNING: failed to allocate 0.00 MB of pinned memory
GGML_ASSERT: D:\a\node-llama-cpp\node-llama-cpp\llama\llama.cpp\ggml-backend.c:100: base != NULL && "backend buffer base cannot be NULL"

[giladgd]

I'll release a new version in the next few days that will include prebuilt binaries with the fix for Vulkan

[bitterspeed]

On Mac (mac-arm64-metal): 3.0 beta 18 + Electron (with Electron-forge + vite) + BGE models run on Electron development (npm run start) but there is a failure with no error message on production (npm run make) at this line: const embeddingContext = await model.createEmbeddingContext();

EDIT: A non-crashing workaround (but effectively useless b/c gibberish is outputted) is using const llama = await getLlama({ gpu: false }); which runs on mac production (npm run make) - after building from source and putting the build files llama/localBuilds/mac-arm64-release-b2953/Release/ into llamaBins/mac-arm64

[giladgd]

@bitterspeed Eventually, it's going to take more than a few days until I can publish a new beta version due to a bug in the llama.cpp tokenizer with some models.

In the meantime, you can run this command to download and build the most recent release of llama.cpp prior to the tokenizer issue, which includes the fix for the Vulkan issue:

npx --no node-llama-cpp download --release b2952

[bitterspeed]

@giladgd Thanks. I have run that command, and while that above Vulkan error does not show up anymore, there is now a crash on runtime (with no error message) when using Vulkan ( CUDA works fine). This applies to not only using BGE embeddings but also Llama 3 inference.

OS: Windows_NT 10.0.19045 (x64)
Node: 18.20.1 (x64)
TypeScript: 5.4.3
node-llama-cpp: 3.0.0-beta.22

CUDA: available
Vulkan: available

CUDA device: NVIDIA GeForce RTX 3080
CUDA used VRAM: 11.39% (1.14GB/10GB)
CUDA free VRAM: 88.6% (8.86GB/10GB)

Vulkan device: NVIDIA GeForce RTX 3080
Vulkan used VRAM: 2.1% (212MB/9.82GB)
Vulkan free VRAM: 97.89% (9.61GB/9.82GB)

CPU model: Intel(R) Core(TM) i7-10700K CPU @ 3.80GHz
Used RAM: 37.03% (11.81GB/31.89GB)
Free RAM: 62.96% (20.08GB/31.89GB)

[giladgd]

@bitterspeed I've managed to use BGE and run inference with Llama 3 on release b2952 with Vulkan, but it only worked when I didn't create more than one context at the same time, or disposed the previous context before creating a new one. If you don't dispose a context manually, the context will be disposed via the garbage collection of V8, which can happen after you create another context, which leads to a crash with Vulkan. I've created an issue on llama.cpp for this.

[bitterspeed]

Amazing. Thank you for the guidance, works perfectly!

[github-actions[bot]]

:tada: This PR is included in version 3.0.0 :tada:

The release is available on:

• npm package (@latest dist-tag)
• GitHub release
• v3.0.0

Your semantic-release bot :package::rocket: