Assertion error when offloading Llama 4 layers to CPU

[BrianStucky-USDA]

Prerequisites

Please answer the following questions for yourself before submitting an issue.

• I am running the latest code. Development is very rapid so there are no tagged versions as of now.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new bug or useful enhancement to share.

Expected Behavior

Please provide a detailed written description of what you were trying to do, and what you expected llama-cpp-python to do.

Running a quantized Llama 4 model (e.g., https://huggingface.co/unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF?show_file_info=Llama-4-Scout-17B-16E-Instruct-UD-IQ1_S.gguf) with some layers on GPU and others offloaded to CPU should be possible.

Current Behavior

Please provide a detailed written description of what llama-cpp-python did, instead.

Attempting to offload 1 or more layers to CPU results in

llama-cpp-python/vendor/llama.cpp/ggml/src/ggml-cuda/mmvq.cu:519: GGML_ASSERT(!src0->view_src) failed

before any tokens are generated.

Inference works as expected if all layers are on CPU or all layers are on GPU.

Environment and Context

Please provide detailed information about your computer setup. This is important in case the issue is not reproducible except for under certain specific conditions.

• Physical (or virtual) hardware you are using, e.g. for Linux:

$ lscpu

AMD EPYC 7713 CPU, NVIDIA A100 GPU

• Operating System, e.g. for Linux:

Rocky Linux

• SDK version, e.g. for Linux:

$ python3 --version
$ make --version
$ g++ --version

Python 3.12.5
GNU Make 4.3
g++ (GCC) 11.3.1 20220421 (Red Hat 11.3.1-2)

Failure Information (for bugs)

This is caused by a bug in llama.cpp that was corrected in a recent commit. The submodule for llama.cpp is pinned to an older commit and does not include the more recent bug fix. I tried building llama-cpp-python with llama.cpp updated to the bug-fix commit, but that resulted in an immediate segmentation fault, so there seem to have been other changes to llama.cpp that are not immediately compatible. I haven't had a chance to explore further.

Please help provide information about the failure if this is a bug. If it is not a bug, please remove the rest of this template.

Steps to Reproduce

Please provide detailed steps for reproducing the issue. We are not sitting in front of your screen, so the more detail the better.

1. Build llama-cpp-python from latest source.
2. Attempt to run inference on a Llama 4 Scout quantized model (e.g., https://huggingface.co/unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF?show_file_info=Llama-4-Scout-17B-16E-Instruct-UD-IQ1_S.gguf) with some layers on GPU and the rest on CPU. E.g., initialize the model as follows:

   from langchain_community.llms import LlamaCpp
   
   llm = LlamaCpp(
       model_path='/path/to/Llama-4-Scout-17B-16E-Instruct-UD-IQ1_S.gguf',
       n_gpu_layers=20, n_batch=512, n_ctx=2048, f16_kv=True, max_tokens=1024,
       verbose=True
   )
   

Note: Many issues seem to be regarding functional or performance issues / differences with llama.cpp. In these cases we need to confirm that you're comparing against the version of llama.cpp that was built with your python package, and which parameters you're passing to the context.

Try the following:

1. git clone https://github.com/abetlen/llama-cpp-python
2. cd llama-cpp-python
3. rm -rf _skbuild/ # delete any old builds
4. python -m pip install .
5. cd ./vendor/llama.cpp
6. Follow llama.cpp's instructions to cmake llama.cpp
7. Run llama.cpp's ./main with the same arguments you previously passed to llama-cpp-python and see if you can reproduce the issue. If you can, log an issue with llama.cpp

Failure Logs

Please include any relevant log snippets or files. If it works under one configuration but not under another, please provide logs for both configurations and their corresponding outputs so it is easy to see where behavior changes.

Also, please try to avoid using screenshots if at all possible. Instead, copy/paste the console output and use Github's markdown to cleanly format your logs for easy readability.

Example environment info:

llama-cpp-python$ git log | head -1
commit 47b0aa6e957b93dbe2c29d53af16fbae2dd628f2

llama-cpp-python$ python3 --version
Python 3.10.10

llama-cpp-python$ pip list | egrep "uvicorn|fastapi|sse-starlette|numpy"
fastapi                  0.95.0
numpy                    1.24.3
sse-starlette            1.3.3
uvicorn                  0.21.1

llama-cpp-python/vendor/llama.cpp$ git log | head -3
commit 66874d4fbcc7866377246efbcee938e8cc9c7d76
Author: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
Date:   Thu May 25 20:18:01 2023 -0600