node-configs icon indicating copy to clipboard operation
node-configs copied to clipboard

Model Incompatibility: Nemotron and Gemma Architectures Fail to Load

Open Tobivi opened this issue 5 months ago • 0 comments

Hello GaiaNet Team,

I'm reporting a recurring incompatibility issue with models based on the Nemotron and Gemma architectures. When attempting to load them, the gaianet start command fails with a backend error.

1. Models Affected

The issue occurs with the following GGUF models:

  • gaianet/NVIDIA-Nemotron-Nano-9B-v2-GGUF (and other quantizations)
  • Tobivictor/NVIDIA-Nemotron-Nano-12B-v2-GGUF
  • Tobivictor/embeddinggemma-300m-GGUF
  • gaianet/embeddinggemma-300m-GGUF

2. The Error Message

In all cases, the embedding or chat server fails to start with this precise error:

[error] llama_core::graph in crates/llama-core/src/graph.rs:219: Backend Error: WASI-NN Backend Error: Caller module passed an invalid argument
Error: Operation("Backend Error: WASI-NN Backend Error: Caller module passed an invalid argument")

3. System Details

  • Operating System: macOS Sequoia 15.5
  • GaiaNet CLI Version: GaiaNet CLI Tool v0.5.4

4. Steps to Reproduce

The error can be reproduced consistently with the following steps:

  1. Run gaianet stop.
  2. Configure the node for one of the affected models using gaianet config --chat-url ... or gaianet config --embedding-url ....
  3. Manually update the chat_name or embedding_name in ~/gaianet/config.json.
  4. Run gaianet init.
  5. Run gaianet start and observe the error during the "Starting downstream servers" phase.

5. Important Context: What Works

My GaiaNet installation and system are working perfectly with other model architectures. The following models run without any issues:

  • Chat Models: Qwen3-4B-Instruct, Qwen2.5-Omni-7B, Jan-v1-4B, Qwen3-Coder-30B-A3B-Instruct
  • Embedding Models: Nomic-embed-text-v1.5

6. Suggested Cause

The fact that multiple other models work correctly strongly suggests that the version of the LlamaEdge engine used by GaiaNet has not yet been updated with full support for the Nemotron and Gemma architectures.

Could the development team please look into updating the underlying llama.cpp dependency to a more recent version? I'm happy to provide more logs or assist with further testing.

Thank you for your hard work on this project!

Tobivi avatar Sep 09 '25 15:09 Tobivi