Skip to content

Instantly share code, notes, and snippets.

@rsrini7

rsrini7/llama.cpp-mtp-ngram-mac.md

Last active May 21, 2026 16:45
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save rsrini7/8b38283478fa48cc291f683f843dc50e to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save rsrini7/8b38283478fa48cc291f683f843dc50e to your computer and use it in GitHub Desktop.
Download ZIP
mac mini installed
Raw
llama.cpp-mtp-ngram-mac.md

brew install make

git clone https://github.com/ggml-org/llama.cpp

cd llama.cpp

cmake -B build

cmake --build build --config Release -j $(sysctl -n hw.ncpu)

brew install hf

hf auth login

open browser -> https://huggingface.co/settings/tokens get token

mkdir -p ~/llama.cpp/models cd ~/llama.cpp/models

hf download ggml-org/Qwen3.6-27B-MTP-GGUF
Qwen3.6-27B-MTP-Q8_0.gguf
--local-dir .

hf download RDson/Qwen3.6-27B-MTP-Q4_K_M-GGUF
Qwen3.6-27B-MTP-Q4_K_M.gguf
--local-dir .

or local dir inside llama.cpp

hf download Radamanthys11/Qwen3.6-27B-MTP-Q8_0-GGUF
Qwen3.6-27B-MTP-Q8_0.gguf
--local-dir ./models

only mtp - 16.5gb model

./build/bin/llama-server
-m models/qwen3.6-27b-mtp-Q4_K_M.gguf
-ngl 999
-c 4096
--spec-type draft-mtp
--spec-draft-n-max 2
--port 9000

only mtp - 30gb model

./build/bin/llama-server
-m ~/llama.cpp/models/Qwen3.6-27B-MTP-Q8_0.gguf
-ngl 999
-c 16384
--flash-attn on
--spec-type draft-mtp
--spec-draft-n-max 2
--host 0.0.0.0
--port 9000

mtp + ngram - 16.5gb model

./build/bin/llama-server
-m ~/llama.cpp/models/Qwen3.6-27B-MTP-Q4_K_M.gguf
-ngl 999
-c 16384
--flash-attn on
--spec-type ngram-mod,draft-mtp
--spec-draft-n-max 2
--spec-ngram-mod-n-match 24
--spec-ngram-mod-n-min 48
--spec-ngram-mod-n-max 64
--host 0.0.0.0
--port 9000

mtp + ngram - 30gb model

./build/bin/llama-server
-m ~/llama.cpp/models/Qwen3.6-27B-MTP-Q8_0.gguf
-ngl 999
-c 16384
--flash-attn on
--spec-type ngram-mod,draft-mtp
--spec-draft-n-max 2
--spec-ngram-mod-n-match 24
--spec-ngram-mod-n-min 48
--spec-ngram-mod-n-max 64
--host 0.0.0.0
--port 9000

curl http://localhost:9000/v1/chat/completions
-H "Content-Type: application/json"
-d '{"model":"qwen3","messages":[{"role":"user","content":"Hello"}]}'

Raw
llama.cpp-router.txt
micro ~/llama.cpp/models/models.ini
models.ini
version = 1
[*]
threads = 12
ctx-size = 8192
n = -1
flash-attn = on
no-mmap = true
[Qwen3.6-27B-MTP-Q8_0]
ctx-size = 16384
n-gpu-layers = 999
temperature = 0.6
top-p = 0.95
top-k = 20
min-p = 0.0
repeat-penalty = 1.0
cache-type-k = q8_0
cache-type-v = q8_0
batch-size = 2048
ubatch-size = 512
spec-type = ngram-mod,draft-mtp
spec-draft-n-max = 2
spec-ngram-mod-n-match = 24
spec-ngram-mod-n-min = 48
spec-ngram-mod-n-max = 64
[Qwen3.6-27B-MTP-Q4_K_M]
ctx-size = 16384
n-gpu-layers = 999
temperature = 0.6
top-p = 0.95
top-k = 20
min-p = 0.0
repeat-penalty = 1.0
cache-type-k = q4_0
cache-type-v = q4_0
batch-size = 2048
ubatch-size = 512
spec-type = ngram-mod,draft-mtp
spec-draft-n-max = 2
spec-ngram-mod-n-match = 24
spec-ngram-mod-n-min = 48
spec-ngram-mod-n-max = 64
---
Command 1: Q8_0 (maximum quality)
./build/bin/llama-server \
--host 0.0.0.0 \
--port 9000 \
--threads 12 \
--models-dir ~/llama.cpp/models/ \
--models-autoload \
--models-max 1 \
--models-preset ~/llama.cpp/models/models.ini \
-m ~/llama.cpp/models/Qwen3.6-27B-MTP-Q8_0.gguf
---
Command 2: Q4_K_M (faster, lighter)
./build/bin/llama-server \
--host 0.0.0.0 \
--port 9000 \
--threads 12 \
--models-dir ~/llama.cpp/models/ \
--models-autoload \
--models-max 1 \
--models-preset ~/llama.cpp/models/models.ini \
-m ~/llama.cpp/models/Qwen3.6-27B-MTP-Q4_K_M.gguf
Raw
mac mini installed.md

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

brew install mise

ls ~/.local/share/mise/installs/

brew install --cask meld brew install --cask sublime-text

brew tap jundot/omlx https://github.com/jundot/omlx

brew install omlx --with-grammar

(or)

brew install omlx &&

brew reinstall omlx --with-grammar

/opt/homebrew/opt/omlx/libexec/bin/pip install "omlx[mcp]"

/opt/homebrew/opt/omlx/libexec/bin/pip install "omlx[modelscope]"

omlx serve

openssl rand -hex 12

ssh-keygen -t ed25519 -C "7rsrini@gmail.com"

pbcopy < ~/.ssh/id_ed25519.pub

git config --global user.email "7rsrini@gmail.com"

git config --global user.name "Srinivasan Ragothaman"

curl -fsSL https://bun.com/install | bash

bun add -g opencode-ai

npm install -g @gitlawb/openclaude

brew install podman

podman machine init

podman machine start

bun add -g @earendil-works/pi-coding-agent

'omlx' launch pi --model 'GLM-4.7-Flash-4bit' --api-key '66a5aecc85a403471667cc64'

'omlx' launch opencode --model 'gpt-oss-20b-MXFP4-Q8' --api-key '66a5aecc85a403471667cc64'

'omlx' launch opencode --model 'Qwen3.6-35B-A3B-4bit' --api-key '66a5aecc85a403471667cc64'

'omlx' launch pi --model 'Qwen3.6-35B-A3B-4bit' --api-key '66a5aecc85a403471667cc64'

curl -fsSL https://herdr.dev/install.sh | sh

brew install fastfetch

brew install micro

brew install podman

brew install btop

brew install dos2unix

brew install mise

brew install node

brew install ripgrep

brew install btop

brew install macmon

brew install glow

brew install fzf

brew install zoxide

brew install sqlite3

brew install lnav

brew install tailspin

brew install gemini-cli (now replaced by antigravity cli) --> curl -fsSL https://antigravity.google/cli/install.sh | bash

brew install mas

direct installed

Antigravity 2.0 Citrix Workspace DevCleaner GarageBand Google Chrome iTerm Podman Desktop Telegram TRAE WhatsApp Xcode zoom Zoom VDI Plugin Management

curl -fsSL https://ollama.com/install.sh | sh

npm install -g opensrc

curl -s "https://get.sdkman.io" | zsh

source "/Users/srinivasanragothaman/.sdkman/bin/sdkman-init.sh"

sdk install java 25.0.3-tem

curl -fsSL https://antigravity.google/cli/install.sh | bash

Raw
others.md

pwd | pbcopy && echo "/$(pbpaste)/arcclaw-dev.log" | pbcopy

system_profiler SPApplicationsDataType -json
| jq -r '.SPApplicationsDataType[] | select(.obtained_from != "apple") | ._name'
| sort > appslist.txt

Raw
zshrc
export PATH="$HOME/.local/share/mise/shims:$PATH"
# mise - ArcClaw: activate (added by setup:profile)
eval "$(mise activate zsh)"
# mise - ArcClaw: tab completion (added by setup:profile)
autoload -Uz compinit && compinit; eval "$(mise completion zsh)"
# bun completions
[ -s "/Users/srinivasanragothaman/.bun/_bun" ] && source "/Users/srinivasanragothaman/.bun/_bun"
# bun
export BUN_INSTALL="$HOME/.bun"
export PATH="$BUN_INSTALL/bin:$PATH"
eval "$(zoxide init zsh)"
export PATH="/Users/srinivasanragothaman/.local/bin:$PATH"
# Added by Antigravity
export PATH="/Users/srinivasanragothaman/.antigravity/antigravity/bin:$PATH"
export GEMINI_SANDBOX=podman
export SANDBOX_SET_UID_GID=false
export SANDBOX_FLAGS="--user root --env HOME=/home/$USER"
alias docker=podman
# Added by Antigravity IDE
export PATH="/Users/srinivasanragothaman/.antigravity-ide/antigravity-ide/bin:$PATH"
#THIS MUST BE AT THE END OF THE FILE FOR SDKMAN TO WORK!!!
export SDKMAN_DIR="$HOME/.sdkman"
[[ -s "$HOME/.sdkman/bin/sdkman-init.sh" ]] && source "$HOME/.sdkman/bin/sdkman-init.sh"
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment