./build/bin/llama-server \
    --alias anikifoss/MiniMax-M2.1-HQ4_K \
    --model /models/anikifoss/MiniMax-M2.1-HQ4_K-GGUF/MiniMax-M2.1-HQ4_K-00001-of-00004.gguf \
    --temp 1.0 --top-k 0 --top-p 1.0 --min-p 0.02 \
    --repeat-penalty 1.04 --repeat-last-n 256 \
    --ctx-size 84000 \
    -ctk q8_0 -ctv q8_0 \
    -fa on \
    -b 1024 -ub 1024 \
    -ngl 99 \
    --device CUDA0 \
    -ot "blk\.([0-9])\.attn_.*=CUDA0" \
    -ot "blk\.([1-6][0-9])\.attn_.*=CUDA0" \
    -ot "blk\.([0-4])\.ffn_.*_exps.*=CUDA0" \
    -ot "blk\.([5-9])\.ffn_.*_exps.*=CPU" \
    -ot "blk\.([1-6][0-9])\.ffn_.*_exps.*=CPU" \
    --jinja \
    --parallel 1 \
    --threads 32 \
    --host 127.0.0.1 \
    --port 8090

llama.cpp - experts offload to 4x32G AMD GPUs (MI50 or R9700)

bash

./build/bin/llama-server \
    --alias anikifoss/MiniMax-M2.1-HQ4_K \
    --model /models/anikifoss/MiniMax-M2.1-HQ4_K-GGUF/MiniMax-M2.1-HQ4_K-00001-of-00004.gguf \
    --temp 1.0 --top-k 0 --top-p 1.0 --min-p 0.02 \
    --repeat-penalty 1.04 --repeat-last-n 256 \
    --ctx-size 84000 \
    -ctk q8_0 -ctv q8_0 \
    -fa on \
    -b 1024 -ub 1024 \
    -ngl 99 \
    --device CUDA0,ROCm0,ROCm1,ROCm2,ROCm3 \
    --tensor-split 1,0,0,0,0 \
    -ot "blk\.([0-9])\.attn_.*=CUDA0" \
    -ot "blk\.([1-6][0-9])\.attn_.*=CUDA0" \
    -ot "blk\.([0-9])\.ffn_.*_exps.*=ROCm0" \
    -ot "blk\.(1[0-9])\.ffn_.*_exps.*=ROCm1" \
    -ot "blk\.(2[0-9])\.ffn_.*_exps.*=ROCm2" \
    -ot "blk\.(3[0-9])\.ffn_.*_exps.*=ROCm3" \
    -ot "blk\.(4[0-3])\.ffn_.*_exps.*=ROCm0" \
    -ot "blk\.(4[4-7])\.ffn_.*_exps.*=ROCm1" \
    -ot "blk\.(4[8-9])\.ffn_.*_exps.*=ROCm2" \
    -ot "blk\.(5[0-1])\.ffn_.*_exps.*=ROCm2" \
    -ot "blk\.(5[2-5])\.ffn_.*_exps.*=ROCm3" \
    -ot "blk\.(5[6-9])\.ffn_.*_exps.*=CUDA0" \
    -ot "blk\.(6[0-9])\.ffn_.*_exps.*=CUDA0" \
    --jinja \
    --parallel 1 \
    --threads 32 \
    --host 127.0.0.1 \
    --port 8090

Provided that you have ROCm setup on Linux, you can build llama.cpp with both CUDA and ROCm support at the same time using this command:

bash

HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
    cmake -S . -B build -DGGML_HIP=ON \
    -DGGML_BACKEND_DL=ON -DGGML_CUDA=ON -DGGML_CPU_ALL_VARIANTS=ON \
    -DAMDGPU_TARGETS="gfx906;gfx1201" \
    -DGGML_BLAS=OFF -DLLAMA_CURL=OFF -DGGML_RPC=ON -DCMAKE_BUILD_TYPE=Release

HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
    cmake --build ./build --config Release --clean-first -j $(nproc)

Adjust DAMDGPU_TARGETS:

gfx906 for MI50
gfx1201 for R9700

Quantization Recipe

Quantized with llama.cpp. See Custom Quants section in this detailed guide for all the quantization steps.

bash

TARGET_MODEL="MiniMax-M2.1-HQ4_K"
mkdir -p /models/anikifoss/$TARGET_MODEL
./build/bin/llama-quantize \
    --output-tensor-type Q8_0 \
    --token-embedding-type Q8_0 \
    --tensor-type attn_q=Q8_0 \
    --tensor-type attn_k=Q8_0 \
    --tensor-type attn_v=Q8_0 \
    --tensor-type ffn_down_exps=Q6_K \
    --tensor-type ffn_gate_exps=Q4_K \
    --tensor-type ffn_up_exps=Q4_K \
    /models/MiniMaxAI/MiniMax-M2.1-GGUF/MiniMax-M2.1-256x4.9B-BF16-00001-of-00010.gguf \
    /models/anikifoss/$TARGET_MODEL/$TARGET_MODEL.gguf \
    Q8_0 \
    32