MiMo-V2-Flash-GGUF

#!/usr/bin/env bash

custom="
# 48 Repeating Layers [0-47]

# Attention
blk\..*\.attn_q.*=q8_0
blk\..*\.attn_k.*=q8_0
blk\..*\.attn_v.*=q8_0
blk\..*\.attn_output.*=q8_0

# First 1 Dense Layers [0]
blk\..*\.ffn_down\.weight=q8_0
blk\..*\.ffn_(gate|up)\.weight=q8_0

# Routed Experts Layers [1-47]
blk\..*\.ffn_down_exps\.weight=iq6_k
blk\..*\.ffn_(gate|up)_exps\.weight=iq5_k

# Non-Repeating Layers
token_embd\.weight=q8_0
output\.weight=q8_0
"

custom=$(
  echo "$custom" | grep -v '^#' | \
  sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
)

numactl -N ${SOCKET} -m ${SOCKET} \
./build/bin/llama-quantize \
    --custom-q "$custom" \
    --imatrix /mnt/data/models/ubergarm/MiMo-V2-Flash-GGUF/imatrix-MiMo-V2-Flash-BF16.dat \
    /mnt/data/models/ubergarm/MiMo-V2-Flash-GGUF/MiMo-V2-Flash-256x7.2B-BF16-00001-of-00013.gguf \
    /mnt/data/models/ubergarm/MiMo-V2-Flash-GGUF/MiMo-V2-Flash-IQ5_K.gguf \
    IQ5_K \
    128

# Clone and checkout
$ git clone https://github.com/ikawrakow/ik_llama.cpp
$ cd ik_llama.cpp

# Build for hybrid CPU+CUDA
$ cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON
$ cmake --build build --config Release -j $(nproc)

# Full 2x GPU offload
./build/bin/llama-server \
    --model "$model" \
    --alias ubergarm/MiMo-V2-Flash-GGUF \
    --ctx-size 32768 \
    -ctk q8_0 -ctv q8_0 \
    -sm graph \
    -smgs \
    -mea 256 \
    -ts 42,48 \
    -ngl 99 \
    -ub 2048 -b 2048 \
    --threads 1 \
    --host 127.0.0.1 \
    --port 8080 \
    --no-mmap \
    --jinja

# Hybrid CPU + 2 or more GPUs
# using new "-sm graph" 'tensor parallel' feature!
# https://github.com/ikawrakow/ik_llama.cpp/pull/1080
# https://github.com/ikawrakow/ik_llama.cpp/pull/1105
For examples take a look at: https://huggingface.co/ubergarm/GLM-4.7-GGUF#quick-start

# CPU Only
numactl -N "$SOCKET" -m "$SOCKET" \
./build/bin/llama-server \
    --model "$model" \
    --alias ubergarm/MiMo-V2-Flash \
    --ctx-size 65536 \
    --merge-qkv \
    -ctk q8_0 -ctv q8_0 \
    -ub 4096 -b 4096 \
    --parallel 1 \
    --threads 96 \
    --threads-batch 128 \
    --numa numactl \
    --host 127.0.0.1 \
    --port 8080 \
    --no-mmap \
    --no-display-prompt \
    --log-enable \
    --jinja