---
license: apache-2.0
base_model:
- deepseek-ai/DeepSeek-R1
---

# Q2_K Quant of Deepseek-R1 for the MLA fork pull request

## Requires this custom build of llama.cpp:

https://github.com/ggerganov/llama.cpp/pull/11446

** IMPORTANT NOTE **

If you try to load this with the `main` branch of llama.cpp you'll see an error like this:

```
load_tensors: loading model tensors, this can take a while... (mmap = true)
llama_model_load: error loading model: done_getting_tensors: wrong number of tensors; expected 1147, got 1025
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model '/mount/checkpoints/DeepSeek-R1-11446-Q2_K-00001-of-00030.gguf'
srv    load_model: failed to load model, '/mount/checkpoints/DeepSeek-R1-11446-Q2_K-00001-of-00030.gguf'
srv    operator(): operator(): cleaning up before exit...
main: exiting due to model loading error
terminate called without an active exception
Aborted (core dumped)
```

There's a Q3_K_M version here: [daydream-org/DeepSeek-R1-GGUF-11446](https://huggingface.co/daydream-org/DeepSeek-R1-GGUF-11446)

Created using the script below by [evshiron](https://huggingface.co/evshiron):

```python
export WORK_DIR=$(pwd)
python3 -m venv venv
source venv/bin/activate
pip3 install -U "huggingface_hub[cli]"

# the fp8 checkpoints are around 700GB
mkdir checkpoints
huggingface-cli download --resume-download --local-dir checkpoints/DeepSeek-R1 deepseek-ai/DeepSeek-R1

# my fork of llama.cpp including pr #11446 and some changes to allow converting fp8 hf to bf16 gguf directly using triton(-cpu) without the need of intermediate checkpoints
git clone https://github.com/evshiron/llama.cpp --recursive
pushd llama.cpp
pip3 install -r requirements/requirements-convert_hf_to_gguf.txt
cmake -B build
cmake --build build --config Release
popd

# install triton-cpu for cpu-only dequant
git clone https://github.com/triton-lang/triton-cpu --recursive
pushd triton-cpu
pip3 install ninja cmake wheel pybind11
MAX_JOBS=32 pip3 install -e python
popd

# hopefully it should work, takes an hour or more depending on your hardware, the bf16 checkpoints are around 1.3TB
# the dequant process may take more than 64GB RAM, but should be doable within 360GB RAM
python3 llama.cpp/convert_hf_to_gguf.py --outtype bf16 --split-max-size 50G checkpoints/DeepSeek-R1

# removing the fp8 checkpoints gives us 700GB back
mkdir checkpoints/DeepSeek-R1-BF16
mv checkpoints/DeepSeek-R1/*.gguf checkpoints/DeepSeek-R1-BF16
rm -r checkpoints/DeepSeek-R1

# then use llama-quantize to make the quants you want, Q4_K_M should be around 400GB?
./llama.cpp/build/bin/llama-quantize --keep-split checkpoints/DeepSeek-R1-BF16/<THE_FIRST_OF_DeepSeek-R1-BF16_GGUF>.gguf Q4_K_M
```

It took 16 hours on an EC2 instance so I figured I'd share it.

Script Credit/Source: [daydream-org/DeepSeek-R1-GGUF-11446](https://huggingface.co/daydream-org/DeepSeek-R1-GGUF-11446/discussions/1#67a327570051a98a96ded9e6)