Qwen3.5-35B-A3B-GGUF

#!/usr/bin/env bash

custom="
# 60 Repeating Layers [0-59]

## Gated Attention/Delta Net [Blended 0-59]
blk\..*\.attn_gate\.weight=q8_0
blk\..*\.attn_qkv\.weight=q8_0
blk\..*\.attn_output\.weight=q8_0
blk\..*\.attn_q\.weight=q8_0
blk\..*\.attn_k\.weight=q8_0
blk\..*\.attn_v\.weight=q8_0
blk\..*\.ssm_alpha\.weight=f32
blk\..*\.ssm_beta\.weight=f32
blk\..*\.ssm_out\.weight=q8_0

# Shared Expert Layers [0-59]
blk\..*\.ffn_down_shexp\.weight=q8_0
blk\..*\.ffn_(gate|up)_shexp\.weight=q8_0

# Routed Experts Layers [0-59]
blk\..*\.ffn_down_exps\.weight=iq5_ks
blk\..*\.ffn_(gate|up)_exps\.weight=iq4_ks

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

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

    #--dry-run \
numactl -N ${SOCKET} -m ${SOCKET} \
./build/bin/llama-quantize \
    --custom-q "$custom" \
    --imatrix /mnt/data/models/ubergarm/Qwen3.5-35B-A3B-GGUF/imatrix-Qwen3.5-35B-A3B-BF16.dat \
    /mnt/data/models/ubergarm/Qwen3.5-35B-A3B-GGUF/Qwen3.5-35B-A3B-BF16-00001-of-00002.gguf \
    /mnt/data/models/ubergarm/Qwen3.5-35B-A3B-GGUF/Qwen3.5-35B-A3B-IQ4_KS.gguf \
    IQ4_KS \
    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)

# Download Desired Quants
$ pip install huggingface_hub
$ hf download --local-dir ./ --include=*IQ4_KS.gguf ubergarm/Qwen3.5-35B-A3B-GGUF

# Full GPU Offload
# NOTE: https://github.com/ikawrakow/ik_llama.cpp/pull/1198
./build/bin/llama-server \
  --alias Qwen3.5-35B-A3B \
  --model "$model" \
  -c 131072 \
  -ctk f16 -ctv q8_0 \
  -fa on \
  -cuda fa-offset=0 \
  -ub 1024 -b 2048 \
  --merge-qkv \
  -muge \
  -ngl 999 \
  --no-mmap \
  --parallel 1 \
  --threads 1 \
  --host 127.0.0.1 \
  --port 8080 \
  --jinja \
  --ctx-checkpoints 8