hellork/DeepSeek-R1-Distill-Qwen-7B-IQ3_XXS-GGUF

TESTING...TESTING! The quantization used on this model may reduce quality, but it is hopefully faster, and maybe usable with 4GB VRAM. TESTING...

So far so good! We were able to use all 29 layers with -ngl 29 and it reserves less than 3.5GiB of VRAM with -c 2048 context window. Quite usable. Use the llama-server and navigate to the web interface at http://127.0.0.1:8080 for best results. Happy AI.

hellork/DeepSeek-R1-Distill-Qwen-7B-IQ3_XXS-GGUF

This model was converted to GGUF format from unsloth/DeepSeek-R1-Distill-Qwen-7B using llama.cpp via the ggml.ai's GGUF-my-repo space. Refer to the original model card for more details on the model.

Usage Recommendations

We recommend adhering to the following configurations when utilizing the DeepSeek-R1 series models, including benchmarking, to achieve the expected performance:

Set the temperature within the range of 0.5-0.7 (0.6 is recommended) to prevent endless repetitions or incoherent outputs. Avoid adding a system prompt; all instructions should be contained within the user prompt. For mathematical problems, it is advisable to include a directive in your prompt such as: "Please reason step by step, and put your final answer within \boxed{}." When evaluating model performance, it is recommended to conduct multiple tests and average the results.

Use with llama.cpp

Install llama.cpp through brew (works on Mac and Linux)

brew install llama.cpp

Or compile it to take advantage of Nvidia CUDA hardware:

git clone https://github.com/ggerganov/llama.cpp.git
cd llama*
# look at docs for other hardware builds or to make sure none of this has changed.

cmake -B build -DGGML_CUDA=ON
CMAKE_ARGS="-DGGML_CUDA=on" cmake --build build --config Release # -j6 (optional: use a number less than the number of cores)

# If your version of gcc is > 12 and it gives errors, use conda to install gcc-12 and activate it.
# Run the above cmake commands again.
# Then run conda deactivate and re-run the last line once more to link the build outside of conda.

# Add the -ngl 33 flag to the comamnds below to take advantage of all the GPU layers.
# If that uses too much GPU and crashes, use some lower number.

Invoke the llama.cpp server or the CLI.

CLI:

llama-cli --hf-repo hellork/DeepSeek-R1-Distill-Qwen-7B-IQ3_XXS-GGUF --hf-file deepseek-r1-distill-qwen-7b-iq3_xxs-imat.gguf -p "The meaning to life and the universe is"

Server:

llama-server --hf-repo hellork/DeepSeek-R1-Distill-Qwen-7B-IQ3_XXS-GGUF --hf-file deepseek-r1-distill-qwen-7b-iq3_xxs-imat.gguf -c 2048

Note: You can also use this checkpoint directly through the usage steps listed in the Llama.cpp repo as well.

Step 1: Clone llama.cpp from GitHub.

git clone https://github.com/ggerganov/llama.cpp

Step 2: Move into the llama.cpp folder and build it with LLAMA_CURL=1 flag along with other hardware-specific flags (for ex: LLAMA_CUDA=1 for Nvidia GPUs on Linux).

cd llama.cpp && LLAMA_CURL=1 make

Step 3: Run inference through the main binary.

./llama-cli --hf-repo hellork/DeepSeek-R1-Distill-Qwen-7B-IQ3_XXS-GGUF --hf-file deepseek-r1-distill-qwen-7b-iq3_xxs-imat.gguf -p "The meaning to life and the universe is"

or

./llama-server --hf-repo hellork/DeepSeek-R1-Distill-Qwen-7B-IQ3_XXS-GGUF --hf-file deepseek-r1-distill-qwen-7b-iq3_xxs-imat.gguf -c 2048