Step-3.5-Flash-GGUF

#!/usr/bin/env bash

custom="
# 45 Repeating Layers [0-44]

# Attention [0-44] GPU
blk\..*\.attn_gate.*=q8_0
blk\..*\.attn_q.*=q8_0
blk\..*\.attn_k.*=q8_0
blk\..*\.attn_v.*=q8_0
blk\..*\.attn_output.*=q8_0

# First 3 Dense Layers [0-2] GPU
blk\..*\.ffn_down\.weight=q8_0
blk\..*\.ffn_(gate|up)\.weight=q8_0

# Shared Expert Layers [3-44] GPU
blk\..*\.ffn_down_shexp\.weight=q8_0
blk\..*\.ffn_(gate|up)_shexp\.weight=q8_0

# Routed Experts Layers [3-44] CPU
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/Step-3.5-Flash-GGUF/imatrix-Step-3.5-Flash-BF16.dat \
    /mnt/data/models/ubergarm/Step-3.5-Flash-GGUF/Step-3.5-Flash-288x7.4B-BF16-00001-of-00009.gguf \
    /mnt/data/models/ubergarm/Step-3.5-Flash-GGUF/Step-3.5-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)

# Run full offload on >2 GPUs with `-sm graph` Graph Parallel
## https://github.com/ikawrakow/ik_llama.cpp/pull/1236
## https://github.com/ikawrakow/ik_llama.cpp/pull/1231
## https://github.com/ikawrakow/ik_llama.cpp/pull/1239
## https://github.com/ikawrakow/ik_llama.cpp/pull/1240
CUDA_VISIBLE_DEVICES="0,1" \
./build/bin/llama-server \
  --model "$model" \
  --alias ubergarm/Step-Fun-3.5-Flash \
  -c 65536 \
  -ger \
  -sm graph \
  -ngl 99 \
  -ub 4096 -b 4096 \
  -ts 47,48 \
  --threads 1 \
  --host 127.0.0.1 \
  --port 8080 \
  --jinja \
  --no-mmap

# CPU-only Mainline llama.cpp Example
numactl -N "$SOCKET" -m "$SOCKET" \
./build/bin/llama-server \
    --model "$model"\
    --alias ubergarm/Step-3.5-Flash \
    --ctx-size 65536 \
    -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 \
    --jinja