I've had some issues running Gemma 4 31B with llama.cpp, even after updating the model weights, pulling the latest codebase and recompiling everything. I've run into some bugs and troubleshot them one by one until I could finally run autonomous long running tasks.

Hope someone finds this helpful.

The Setup:

Hardware: RTX 6000 Pro 96GB, CUDA 13.1, 128GB RAM (DDR5)

Model: Gemma 4 31B Unsloth GGUF BF16, from April 10th (This is the re-upload).

llama.cpp commit: d132f22fc92f36848f7ccf2fc9987cd0b0120825

My launch script:

#!/bin/bash

export GGML_CUDA_NO_VMM=1

llama-server \
    --model /gemma-4-31B-it/BF16/gemma-4-31B-it-BF16-00001-of-00002.gguf \
    --chat-template-file /models/templates/google-gemma-4-31B-it-interleaved.jinja \
    --temp 1.0 \
    --top-p 0.95 \
    --top-k 64 \
    --no-webui \
    --no-mmap \
    --parallel 1 \
    --ctx-size 65576 \
    --flash-attn off

Here's the reason for some of the settings:

These are the recommended parameters from Google:

    --temp 1.0 \
    --top-p 0.95 \
    --top-k 64 \

This was a lot of trial and error. Apparently there are some bugs in llama.cpp where using memory mapping might not free the model weights from RAM, and this caused OOM when trying to use memory which was apparently free, but crashed in run time:

    --no-mmap \
    --parallel 1 \
    --ctx-size 65576 \

Apparently there is a bug in the llama.cpp CUDA implementation where FA kernel fails to synchronize properly when the context is too large:

    --flash-attn off

These are just for my use case:

    --parallel 1 \
    --ctx-size 65576 \
    --no-webui \

For some cases I also use --reasoning-off to save time.

So this is it, with these settings I got Gemma 4 running pretty well with 64K context length. When I get the chance, I'll try TurboQuant to see if I can get even more context length.