DeepSeek-V3.2-Speciale-GGUF

#!/usr/bin/env bash

custom="
## Attention [0-60] (GPU)
# attn_kv_b is only used for PP so keep it q8_0 for best speed and accuracy
blk\..*\.attn_kv_b\.weight=q8_0

# ideally k_b and v_b are smaller than q8_0 as they are is used for TG with -mla 3
# https://github.com/ikawrakow/ik_llama.cpp/issues/651
# blk.*.attn_k_b.weight is not divisible by 256 so only supports iq4_nl or legacy qN_0
blk\..*\.attn_k_b\.weight=q8_0
blk\..*\.attn_v_b\.weight=q8_0

# Balance of attn tensors
blk\..*\.attn_kv_a_mqa\.weight=q8_0
blk\..*\.attn_q_a\.weight=q8_0
blk\..*\.attn_q_b\.weight=q8_0
blk\..*\.attn_output\.weight=q8_0

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

## Shared Expert (1-60) (GPU)
blk\..*\.ffn_down_shexp\.weight=q8_0
blk\..*\.ffn_(gate|up)_shexp\.weight=q8_0

## Routed Experts (1-60) (CPU)
blk\..*\.ffn_down_exps\.weight=iq6_k
blk\..*\.ffn_(gate|up)_exps\.weight=iq5_k

## Token embedding and output tensors (GPU)
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/DeepSeek-V3.2-Speciale-GGUF/imatrix-DeepSeek-V3.2-Speciale-Q8_0.dat \
    /mnt/data/models/ubergarm/DeepSeek-V3.2-Speciale-GGUF/DeepSeek-V3.2-Speciale-256x20B-safetensors-BF16-00001-of-00030.gguf \
    /mnt/data/models/ubergarm/DeepSeek-V3.2-Speciale-GGUF/DeepSeek-V3.2-Speciale-IQ5_K.gguf \
    IQ5_K \
    128

#!/usr/bin/env bash

custom="
## Attention [0-60] (GPU)
blk\..*\.attn_k_b\.weight=q8_0
blk\..*\.attn_v_b\.weight=q8_0

# Balance of attn tensors
blk\..*\.attn_kv_a_mqa\.weight=q8_0
blk\..*\.attn_q_a\.weight=q8_0
blk\..*\.attn_q_b\.weight=q8_0
blk\..*\.attn_output\.weight=iq6_k

## First Three Dense Layers [0-2] (GPU)
blk\.0\.ffn_down\.weight=q8_0
blk\.0\.ffn_(gate|up)\.weight=q8_0
blk\..*\.ffn_down\.weight=q8_0
blk\..*\.ffn_(gate|up)\.weight=iq6_k

## Shared Expert [3-60] (GPU)
blk\..*\.ffn_down_shexp\.weight=q8_0
blk\..*\.ffn_(gate|up)_shexp\.weight=iq6_k

## Routed Experts [3-60] (CPU)
blk\..*\.ffn_down_exps\.weight=iq4_kss
blk\..*\.ffn_(gate|up)_exps\.weight=iq3_k

## Token embedding and output tensors (GPU)
token_embd\.weight=iq4_k
output\.weight=iq6_k
"

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/DeepSeek-V3.2-Speciale-GGUF/imatrix-DeepSeek-V3.2-Speciale-Q8_0.dat \
    /mnt/data/models/ubergarm/DeepSeek-V3.2-Speciale-GGUF/DeepSeek-V3.2-Speciale-256x20B-safetensors-BF16-00001-of-00030.gguf \
    /mnt/data/models/ubergarm/DeepSeek-V3.2-Speciale-GGUF/DeepSeek-V3.2-Speciale-IQ3_K.gguf \
    IQ3_K \
    128

#!/usr/bin/env bash

custom="
## Attention [0-60] (GPU)
blk\..*\.attn_k_b\.weight=q6_0
blk\..*\.attn_v_b\.weight=iq6_k

# Balance of attn tensors
blk\..*\.attn_kv_a_mqa\.weight=iq6_k
blk\..*\.attn_q_a\.weight=iq6_k
blk\..*\.attn_q_b\.weight=iq6_k
blk\..*\.attn_output\.weight=iq6_k

## First Three Dense Layers [0-2] (GPU)
blk\..*\.ffn_down\.weight=iq5_ks
blk\..*\.ffn_(gate|up)\.weight=iq4_kss

## Shared Expert [3-60] (GPU)
blk\..*\.ffn_down_shexp\.weight=iq5_ks
blk\..*\.ffn_(gate|up)_shexp\.weight=iq4_kss

## Routed Experts [3-60] (CPU)
blk\..*\.ffn_down_exps\.weight=iq1_kt
blk\..*\.ffn_(gate|up)_exps\.weight=iq1_kt

## Token embedding and output tensors (GPU)
token_embd\.weight=iq4_k
output\.weight=iq6_k
"

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/DeepSeek-V3.2-Speciale-GGUF/imatrix-DeepSeek-V3.2-Speciale-Q8_0.dat \
    /mnt/data/models/ubergarm/DeepSeek-V3.2-Speciale-GGUF/DeepSeek-V3.2-Speciale-256x20B-safetensors-BF16-00001-of-00030.gguf \
    /mnt/data/models/ubergarm/DeepSeek-V3.2-Speciale-GGUF/DeepSeek-V3.2-Speciale-smol-IQ1_KT.gguf \
    IQ1_KT \
    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 Hybrid CPU + 2x CUDA GPUs (48GB VRAM each older RTX A6000 non-PROs)
## no -sm graph for DeepSeek yet. Is there an easy way to disable thinking or need to prefill response?
## might be better way to do this with --n-cpu-moe 46 -ts 48,48 etc...
./build/bin/llama-server \
    --model "$model" \
    --alias ubergarm/DeepSeek-V3.2-Speciale-GGUF \
    --ctx-size 32768 \
    -ctk q8_0 \
    -ger \
    --merge-qkv \
    -mla 3 -amb 1024 \
    -ot "blk\.(3|4|5|6|7|8|9|10)\.ffn_(gate|up|down)_exps.*=CUDA0" \
    -ot "blk\.(52|53|54|55|56|57|58|59|60)\.ffn_(gate|up|down)_exps.*=CUDA1" \
    --cpu-moe \
    -ub 4096 -b 4096 \
    --threads 24 \
    --host 127.0.0.1 \
    --port 8080 \
    --no-mmap \
    --jinja

# CPU Only
numactl -N ${SOCKET} -m ${SOCKET} \
./build/bin/llama-server \
    --model "$model"\
    --alias ubergarm/DeepSeek-V3.2-Speciale-GGUF \
    --merge-qkv \
    --ctx-size 131072 \
    -ctk q8_0 \
    -mla 3 \
    --parallel 1 \
    --threads 96 \
    --threads-batch 128 \
    --numa numactl \
    --host 127.0.0.1 \
    --port 8080 \
    --no-mmap \
    --jinja

    # --validate-quants