homelab/active/software_ai_stack/ai_stack.md

# Self Hosted AI Stack

- [Self Hosted AI Stack](#self-hosted-ai-stack)
  - [Notes](#notes)
    - [Podman Volume Locations](#podman-volume-locations)
  - [Setup](#setup)
    - [Create the AI user](#create-the-ai-user)
    - [Helper aliases](#helper-aliases)
    - [Create the models dir](#create-the-models-dir)
    - [Install the Hugging Face CLI](#install-the-hugging-face-cli)
    - [Samba Model Storage](#samba-model-storage)
    - [Download models](#download-models)
      - [Text models](#text-models)
        - [GPT-OSS](#gpt-oss)
        - [Mistral](#mistral)
        - [Qwen](#qwen)
        - [GLM](#glm)
        - [Gemma](#gemma)
        - [Dolphin](#dolphin)
        - [LiquidAI](#liquidai)
        - [Level 1 Techs](#level-1-techs)
      - [Image models](#image-models)
        - [Z-Image](#z-image)
        - [Flux](#flux)
      - [Embedding Models](#embedding-models)
        - [Qwen Embedding](#qwen-embedding)
        - [Nomic Embedding](#nomic-embedding)
  - [llama.cpp](#llamacpp)
  - [stable-diffusion.cpp](#stable-diffusioncpp)
  - [STT whisper.cpp](#stt-whispercpp)
  - [TTS coqui-tts](#tts-coqui-tts)
  - [open-webui](#open-webui)
    - [Upgrade Open Webui](#upgrade-open-webui)
  - [lite-llm](#lite-llm)
  - [Install the update script](#install-the-update-script)
  - [Benchmark Results](#benchmark-results)
  - [Testing with Curl](#testing-with-curl)
    - [OpenAI API](#openai-api)
  - [VLLM](#vllm)
    - [Run VLLM with Podman](#run-vllm-with-podman)
  - [Misc](#misc)
    - [Quantizing your own Models](#quantizing-your-own-models)
    - [Qwen3.5 Settings](#qwen35-settings)

## Notes

```bash
# Shortcut for downloading models
hf-download ()
{
    if [ $# -ne 3 ]; then
        echo "ERROR: Expected 3 arguments, but only got $#" 1>&2
        return 1
    fi
    BASE_DIR='/opt/ai/models'
    mkdir -p $BASE_DIR/$1
    pushd $BASE_DIR/$1 2>&1 >/dev/null
    hf download --local-dir . $2 $3
    popd 2>&1 >/dev/null
}
```

### Podman Volume Locations

`~/.local/share/containers/storage/volumes/`

## Setup

### Create the AI user

```bash
# Create your local ai user. This will be the user you launch podman processes from.
useradd -m ai
loginctl enable-linger ai
su -l ai
mkdir -p /home/ai/.config/containers/systemd/
mkdir -p /home/ai/.ssh
```

Models are big. You'll want some tools to help find large files quickly when space runs out.

### Helper aliases

Add these to your .bashrc:

```bash
# Calculate all folder sizes in current dir 
alias {dudir,dud}='du -h --max-depth 1 | sort -h'

# Calculate all file sizes in current dir
alias {dufile,duf}='ls -lhSr'

# Restart llama-server / follow logs
alias llama-reload="systemctl --user daemon-reload && systemctl --user restart llama-server.service"
alias llama-logs="journalctl --user -fu llama-server"

# Restart stable diffusion gen and edit server / follow logs
alias sd-gen-reload='systemctl --user daemon-reload && systemctl --user restart stable-diffusion-gen-server'
alias sd-gen-logs='journalctl --user -xeu stable-diffusion-gen-server'
alias sd-edit-reload='systemctl --user daemon-reload && systemctl --user restart stable-diffusion-edit-server'
alias sd-edit-logs='journalctl --user -xeu stable-diffusion-edit-server'
```

### Create the models dir

```bash
mkdir -p /home/ai/models/{text,image,video,embedding,tts,stt}
```

### Install the Hugging Face CLI

<https://huggingface.co/docs/huggingface_hub/en/guides/cli#getting-started>

```bash
# Install
curl -LsSf https://hf.co/cli/install.sh | bash

# Login
hf auth login
```

### Samba Model Storage

I recommend adding network storage for keeping models offloaded. This mounts a samba share at `/srv/models`.

```bash
dnf install -y cifs-utils

# Add this to /etc/fstab
//driveripper.reeselink.com/smb_models /srv/models cifs _netdev,nofail,uid=1001,gid=1001,credentials=/etc/samba/credentials 0 0

# Then mount
systemctl daemon-reload
mount -a --mkdir
```

Here are some sync commands that I use to keep the samba share in sync with the home directory:

```bash
# Sync models from home dir to the samba share
rsync -av --progress /home/ai/models/ /srv/models/
```

### Download models

In my completely subjective opinion: 5 bit quant is usually the sweet spot for
unsloth models. Q5_K_S is usually just fine.

I usually download the BF16 mmproj files. This is also completely subjective.

#### Text models

<https://huggingface.co/ggml-org/collections>

##### GPT-OSS

<https://unsloth.ai/docs/models/gpt-oss-how-to-run-and-fine-tune#recommended-settings>

```bash
# gpt-oss-120b
mkdir gpt-oss-120b && cd gpt-oss-120b
hf download --local-dir . ggml-org/gpt-oss-120b-GGUF

# gpt-oss-20b
mkdir gpt-oss-20b && cd gpt-oss-20b
hf download --local-dir . ggml-org/gpt-oss-20b-GGUF
```

##### Mistral

```bash
# devstral-small-2-24b
mkdir devstral-small-2-24b && cd devstral-small-2-24b
hf download --local-dir . ggml-org/Devstral-Small-2-24B-Instruct-2512-GGUF Devstral-Small-2-24B-Instruct-2512-Q8_0.gguf

# ministral-3-14b
mkdir ministral-3-14b && cd ministral-3-14b
hf download --local-dir . ggml-org/Ministral-3-14B-Reasoning-2512-GGUF

# ministral-3-3b-instruct
mkdir ministral-3-3b-instruct && cd ministral-3-3b-instruct
hf download --local-dir . ggml-org/Ministral-3-3B-Instruct-2512-GGUF
```

##### Qwen

```bash
# qwen3.6-35b-a3b
mkdir qwen3.6-35b-a3b && cd qwen3.6-35b-a3b
hf download --local-dir . unsloth/Qwen3.6-35B-A3B-GGUF Qwen3.6-35B-A3B-UD-Q5_K_M.gguf
hf download --local-dir . unsloth/Qwen3.6-35B-A3B-GGUF mmproj-F16.gguf

# qwen3.5-27b-opus
mkdir qwen3.5-27b-opus && cd qwen3.5-27b-opus
hf download --local-dir . Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF Qwen3.5-27B.Q4_K_M.gguf
hf download --local-dir . Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF mmproj-BF16.gguf

# qwen3.5-4b
mkdir qwen3.5-4b && cd qwen3.5-4b
hf download --local-dir . unsloth/Qwen3.5-4B-GGUF Qwen3.5-4B-Q8_0.gguf
hf download --local-dir . unsloth/Qwen3.5-4B-GGUF mmproj-F16.gguf

# qwen3.5-35b-a3b
mkdir qwen3.5-35b-a3b && cd qwen3.5-35b-a3b
hf download --local-dir . unsloth/Qwen3.5-35B-A3B-GGUF Qwen3.5-35B-A3B-Q8_0.gguf
hf download --local-dir . unsloth/Qwen3.5-35B-A3B-GGUF mmproj-F16.gguf

# qwen3-30b-a3b-instruct
mkdir qwen3-30b-a3b-instruct && cd qwen3-30b-a3b-instruct
hf download --local-dir . ggml-org/Qwen3-30B-A3B-Instruct-2507-Q8_0-GGUF

# qwen3-vl-30b-a3b-thinking
mkdir qwen3-vl-30b-a3b-thinking && cd qwen3-vl-30b-a3b-thinking
hf download --local-dir . Qwen/Qwen3-VL-30B-A3B-Thinking-GGUF Qwen3VL-30B-A3B-Thinking-Q8_0.gguf
hf download --local-dir . Qwen/Qwen3-VL-30B-A3B-Thinking-GGUF mmproj-Qwen3VL-30B-A3B-Thinking-F16.gguf

# qwen3-vl-30b-a3b-instruct
mkdir qwen3-vl-30b-a3b-instruct && cd qwen3-vl-30b-a3b-instruct
hf download --local-dir . Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF Qwen3VL-30B-A3B-Instruct-Q8_0.gguf
hf download --local-dir . Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF mmproj-Qwen3VL-30B-A3B-Instruct-F16.gguf

# qwen3-coder-30b-a3b-instruct
mkdir qwen3-coder-30b-a3b-instruct && cd qwen3-coder-30b-a3b-instruct
hf download --local-dir . ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF

# qwen3-coder-next
mkdir qwen3-coder-next && cd qwen3-coder-next
hf download --local-dir . unsloth/Qwen3-Coder-Next-GGUF --include "Q8_0/*.gguf"

# qwen3-8b (benchmarks)
mkdir qwen3-8b && cd qwen3-8b
hf download --local-dir . Qwen/Qwen3-8B-GGUF Qwen3-8B-Q8_0.gguf
```

##### GLM

```bash
# glm-4.7-flash-30b
mkdir glm-4.7-flash-30b && cd glm-4.7-flash-30b
hf download --local-dir . unsloth/GLM-4.7-Flash-GGUF GLM-4.7-Flash-Q8_0.gguf
```

##### Gemma

```bash
# Note "it" vs "pt" suffixes. "it" is instruction following, "pt" is the base model (not as good for out-of-the-box use)

# gemma-4-26b-a4b
mkdir gemma-4-26b-a4b && cd gemma-4-26b-a4b
hf download --local-dir . ggml-org/gemma-4-26B-A4B-it-GGUF gemma-4-26B-A4B-it-Q8_0.gguf
hf download --local-dir . ggml-org/gemma-4-26B-A4B-it-GGUF mmproj-gemma-4-26B-A4B-it-f16.gguf

# gemma-4-31b
mkdir gemma-4-31b && cd gemma-4-31b
hf download --local-dir . ggml-org/gemma-4-31B-it-GGUF gemma-4-31B-it-Q8_0.gguf
hf download --local-dir . ggml-org/gemma-4-31B-it-GGUF mmproj-gemma-4-31B-it-f16.gguf

# gemma-3-27b-it
mkdir gemma-3-27b-it && cd gemma-3-27b-it
hf download --local-dir . unsloth/gemma-3-27b-it-GGUF gemma-3-27b-it-Q8_0.gguf
hf download --local-dir . unsloth/gemma-3-27b-it-GGUF mmproj-F16.gguf
```

##### Dolphin

```bash
# dolphin-mistral-24b-venice
mkdir dolphin-mistral-24b-venice && cd dolphin-mistral-24b-venice
hf download --local-dir . bartowski/cognitivecomputations_Dolphin-Mistral-24B-Venice-Edition-GGUF cognitivecomputations_Dolphin-Mistral-24B-Venice-Edition-Q8_0.gguf
```

##### LiquidAI

```bash
# lfm2-24b
mkdir lfm2-24b && cd lfm2-24b
hf download --local-dir . LiquidAI/LFM2-24B-A2B-GGUF LFM2-24B-A2B-Q8_0.gguf
```

##### Level 1 Techs

```bash
# kappa-20b
# https://huggingface.co/eousphoros/kappa-20b-131k-GGUF-Q8_0/tree/main
mkdir kappa-20b && cd kappa-20b
hf download --local-dir . eousphoros/kappa-20b-131k-GGUF-Q8_0
```

#### Image models

##### Z-Image

```bash
# z-turbo
# Fastest image generation in 8 steps. Great a text and prompt following.
# Lacks variety.
mkdir /home/ai/models/image/z-turbo && cd /home/ai/models/image/z-turbo
hf download --local-dir . leejet/Z-Image-Turbo-GGUF z_image_turbo-Q8_0.gguf
hf download --local-dir . black-forest-labs/FLUX.1-schnell ae.safetensors
hf download --local-dir . unsloth/Qwen3-4B-Instruct-2507-GGUF Qwen3-4B-Instruct-2507-Q8_0.gguf
```

##### Flux

```bash
# flux2-klein
# Capable of editing images in 4 steps (though 5 is my recommended steps)
mkdir /home/ai/models/image/flux2-klein && cd /home/ai/models/image/flux2-klein
hf download --local-dir . leejet/FLUX.2-klein-9B-GGUF flux-2-klein-9b-Q8_0.gguf
hf download --local-dir . black-forest-labs/FLUX.2-dev ae.safetensors
hf download --local-dir . unsloth/Qwen3-8B-GGUF Qwen3-8B-Q8_0.gguf
```

#### Embedding Models

##### Qwen Embedding

```bash
mkdir qwen3-embed-4b && cd qwen3-embed-4b
hf download --local-dir . Qwen/Qwen3-Embedding-4B-GGUF Qwen3-Embedding-4B-Q8_0.gguf
```

##### Nomic Embedding

```bash
# nomic-embed-text-v2
mkdir /home/ai/models/embedding/nomic-embed-text-v2
hf download --local-dir /home/ai/models/embedding/nomic-embed-text-v2 ggml-org/Nomic-Embed-Text-V2-GGUF
```

## llama.cpp

<https://github.com/ggml-org/llama.cpp/tree/master/tools/server>

```bash
# Build the llama.cpp container image
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
export BUILD_TAG=$(date +"%Y-%m-%d-%H-%M-%S")

# Vulkan (better performance as of Feb 2026)
podman build -f .devops/vulkan.Dockerfile -t llama-cpp-vulkan:${BUILD_TAG} -t llama-cpp-vulkan:latest .

# ROCM
podman build -f .devops/rocm.Dockerfile -t llama-cpp-rocm:${BUILD_TAG} -t llama-cpp-rocm:latest .

# Run llama demo server (Available on port 8010)
podman run \
--rm \
--name llama-server-demo \
--device=/dev/kfd \
--device=/dev/dri \
-v $(pwd):/models:z \
-p 8010:8000 \
--ipc host \
localhost/llama-cpp-vulkan:latest \
--host 0.0.0.0 \
--port 8000 \
--perf \
--n-gpu-layers all \
--jinja \
--model /models/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf \
--mmproj /models/mmproj-F16.gguf \
--temp 1.0 \
--top-p 0.95 \
--top-k 64 \
--chat-template-kwargs '{"enable_thinking":true}'

podman run \
--rm \
--name llama-server-demo \
--device=/dev/kfd \
--device=/dev/dri \
-v $(pwd):/models:z \
-p 8010:8000 \
--ipc host \
localhost/llama-cpp-vulkan:latest \
--host 0.0.0.0 \
--port 8000 \
--perf \
--n-gpu-layers all \
--jinja \
--model /models/Qwen3.5-9B-UD-Q4_K_XL.gguf \
--mmproj /models/mmproj-BF16.gguf \
--temp 0.6 \
--top-p 0.95 \
--top-k 20 \
--min-p 0.00 \
--chat-template-kwargs '{"enable_thinking":true}'
```

Embedding models

```bash
podman run \
--rm \
--name llama-server-demo \
--device=/dev/kfd \
--device=/dev/dri \
-v /home/ai/models/text:/models:z \
-p 8000:8000 \
localhost/llama-cpp-vulkan:latest \
--host 0.0.0.0 \
--port 8001 \
-c 512 \
--perf \
--n-gpu-layers all \
--models-max 1 \
--models-dir /models \
--embedding
```

```bash
# Test with curl
curl -X POST "https://llama-embed.reeselink.com/embedding" --data '{"model": "qwen3-embed-4b", "content":"Star Wars is better than Star Trek"}'
```

## stable-diffusion.cpp

Server: <https://github.com/leejet/stable-diffusion.cpp/tree/master/examples/server>

CLI: <https://github.com/leejet/stable-diffusion.cpp/tree/master/examples/cli>

```bash
git clone https://github.com/leejet/stable-diffusion.cpp.git
cd stable-diffusion.cpp
git submodule update --init --recursive
export BUILD_TAG=$(date +"%Y-%m-%d-%H-%M-%S")

# Vulkan
podman build -f Dockerfile.vulkan -t stable-diffusion-cpp:${BUILD_TAG} -t stable-diffusion-cpp:latest .
```

```bash
# Generate an image with z-turbo
podman run --rm \
-v /home/ai/models:/models:z \
-v /home/ai/output:/output:z \
--device /dev/kfd \
--device /dev/dri \
localhost/stable-diffusion-cpp:latest \
--diffusion-model /models/image/z-turbo/z_image_turbo-Q8_0.gguf \
--vae /models/image/z-turbo/ae.safetensors  \
--llm /models/image/z-turbo/Qwen3-4B-Instruct-2507-Q8_0.gguf \
-v \
--cfg-scale 1.0 \
--vae-conv-direct \
--diffusion-conv-direct \
--fa \
--mmap \
--seed -1 \
--steps 8 \
-H 1024 \
-W 1024 \
-o /output/output.png \
-p "A photorealistic dragon"

# Edit the generated image with flux2-klein
podman run --rm \
-v /home/ai/models:/models:z \
-v /home/ai/output:/output:z \
--device /dev/kfd \
--device /dev/dri \
localhost/stable-diffusion-cpp:latest \
--diffusion-model  /models/image/flux2-klein/flux-2-klein-9b-Q8_0.gguf \
--vae /models/image/flux2-klein/ae.safetensors \
--llm /models/image/flux2-klein/Qwen3-8B-Q8_0.gguf \
-v \
--cfg-scale 1.0 \
--sampling-method euler \
--vae-conv-direct \
--diffusion-conv-direct \
--fa \
--mmap \
--steps 5 \
-H 1024 \
-W 1024 \
-r /output/output.png \
-o /output/edit.png \
-p "Replace the dragon with an old car"

# Video generation with wan2.2
podman run --rm \
-v /home/ai/models:/models:z \
-v /home/ai/output:/output:z \
--device /dev/kfd \
--device /dev/dri \
localhost/stable-diffusion-cpp:latest \
-M vid_gen \
--diffusion-model /models/video/wan2.2/Wan2.2-T2V-A14B-LowNoise-Q5_K_M.gguf \
--high-noise-diffusion-model /models/video/wan2.2/Wan2.2-T2V-A14B-HighNoise-Q5_K_M.gguf \
--vae /models/video/wan2.2/wan_2.1_vae.safetensors \
--t5xxl /models/video/wan2.2/umt5-xxl-encoder-Q5_K_M.gguf \
--cfg-scale 3.5 \
--sampling-method euler \
--steps 10 \
--high-noise-cfg-scale 3.5 \
--high-noise-sampling-method euler \
--high-noise-steps 8 \
--vae-conv-direct \
--diffusion-conv-direct \
--vae-tiling \
-v \
-n "Colorful tones, overexposed, static, blurred details, subtitles, style, artwork, painting, picture, still, overall graying, worst quality, low quality, JPEG compression residue, ugly, mutilated, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, deformed limbs, finger fusion, still pictures, messy backgrounds, three legs, many people in the background, walking backwards" \
-W 512 \
-H 512 \
--diffusion-fa \
--video-frames 24 \
--flow-shift 3.0 \
-o /output/video_output \
-p "A normal business meeting. People discuss business for 2 seconds. Suddenly, a horde of furries carrying assault rifles bursts into the room and causes a panic. Hatsune Miku leads the charge screaming in rage."
```

## STT whisper.cpp

Whisper stt service.

```bash
podman run -it --rm -p "8081:8080" \
-v /home/ai/models/stt:/models:z \
--device /dev/kfd \
--device /dev/dri \
ghcr.io/ggml-org/whisper.cpp:main-vulkan \
"whisper-server \
--host 0.0.0.0 \
-m /models/ggml-large-v3-turbo/ggml-large-v3-turbo-q5_0.bin \
--convert"
```

## TTS coqui-tts

```bash
podman run --rm -it -p 5002:5002 --entrypoint /bin/bash ghcr.io/coqui-ai/tts-cpu
```

## open-webui

```bash
mkdir /home/ai/.env
# Create a file called open-webui-env with `WEBUI_SECRET_KEY="some-random-key"
scp active/software_ai_stack/secrets/open-webui-env deskwork-ai:.env/

# Will be available on port 8080
podman run \
-d \
-p 8080:8080 \
-v open-webui:/app/backend/data \
--name open-webui \
--restart always \
ghcr.io/open-webui/open-webui:main
```

Use the following connections:

| Service                   | Endpoint                                  |
| ------------------------- | ----------------------------------------- |
| llama.cpp server          | <http://host.containers.internal:8000>    |
| llama.cpp embed           | <http://host.containers.internal:8001>    |
| stable-diffusion.cpp      | <http://host.containers.internal:1234/v1> |
| stable-diffusion.cpp edit | <http://host.containers.internal:1235/v1> |

### Upgrade Open Webui

```bash
podman image pull ghcr.io/open-webui/open-webui:main
systemctl --user restart open-webui.service
```

## lite-llm

<https://docs.litellm.ai/docs/proxy/configs>

```bash
podman run \
--rm \
--name litellm \
-p 4000:4000
```

## Install the update script

```bash
# 1. Builds the latest llama.cpp and stable-diffusion.cpp
# 2. Pulls the latest open-webui
# 3. Restarts all services
scp active/software_ai_stack/update-script.sh deskwork-ai:
ssh deskwork-ai
chmod +x update-script.sh
./update-script.sh
```

## Benchmark Results

Benchmarks are run with [unsloth gpt-oss-20b Q8_0](https://huggingface.co/unsloth/gpt-oss-20b-GGUF/blob/main/gpt-oss-20b-Q8_0.gguf)

```bash
# Run the llama.cpp pod (AMD)
podman run -it --rm \
--device=/dev/kfd \
--device=/dev/dri \
-v /home/ai/models/text:/models:z \
--entrypoint /bin/bash \
ghcr.io/ggml-org/llama.cpp:full-vulkan

# Benchmark command
./llama-bench -m /models/gpt-oss-20b/gpt-oss-20b-Q8_0.gguf -p 4096 -n 1024
```

Framework Desktop

| model            |      size |  params | backend |  ngl |   test |           t/s |
| ---------------- | --------: | ------: | ------- | ---: | -----: | ------------: |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | Vulkan  |   99 | pp4096 | 992.74 ± 6.07 |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | Vulkan  |   99 | tg1024 |  75.82 ± 0.07 |

AMD R9700

| model            |      size |  params | backend |  ngl |   test |            t/s |
| ---------------- | --------: | ------: | ------- | ---: | -----: | -------------: |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | Vulkan  |   99 | pp4096 | 3190.85 ± 8.24 |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | Vulkan  |   99 | tg1024 |  168.73 ± 0.15 |

NVIDIA GeForce RTX 4080 SUPER

| model            |      size |  params | backend |  ngl |  test |           t/s |
| ---------------- | --------: | ------: | ------- | ---: | ----: | ------------: |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | CUDA    |   99 | tg128 | 193.28 ± 1.03 |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | CUDA    |   99 | tg256 | 193.55 ± 0.34 |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | CUDA    |   99 | tg512 | 187.39 ± 0.10 |

NVIDIA GeForce RTX 3090

| model            |      size |  params | backend     |  ngl |   test |             t/s |
| ---------------- | --------: | ------: | ----------- | ---: | -----: | --------------: |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | CUDA,Vulkan |   99 | pp4096 | 3034.03 ± 80.36 |
| gpt-oss 20B Q8_0 | 11.27 GiB | 20.91 B | CUDA,Vulkan |   99 | tg1024 |   181.05 ± 9.01 |

Apple M4 max

| model                         |   test |            t/s |
| :---------------------------- | -----: | -------------: |
| unsloth/gpt-oss-20b-Q8_0-GGUF | pp2048 | 1579.12 ± 7.12 |
| unsloth/gpt-oss-20b-Q8_0-GGUF |   tg32 |  113.00 ± 2.81 |

## Testing with Curl

### OpenAI API

```bash
export TOKEN=$(cat active/software_ai_stack/secrets/aipi-token)

# List Models
curl https://llama-instruct.reeseapps.com/v1/models \
-H "Authorization: Bearer $TOKEN" | jq '.data'

# Text
curl https://llama-instruct.reeseapps.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $TOKEN" \
-d '{
  "model": "llama-instruct/instruct",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Hello, how are you?"}
  ],
  "max_tokens": 500
}' | jq

# Completion
curl https://llama-instruct.reeseapps.com/v1/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $TOKEN" \
-d '{
  "model": "llama-instruct/instruct",
  "prompt": "Write a short poem about the ocean.",
  "max_tokens": 500
}' | jq

# Image Gen
curl https://image-gen.reeselink.com/v1/images/generations \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $TOKEN" \
-d '{
  "model": "sdd-gen/sd-cpp-local",
  "prompt": "A futuristic city with flying cars at sunset, digital art",
  "n": 1,
  "size": "1024x1024"
}' | jq

# Image Edit
curl http://aipi.reeseapps.com/v1/images/edits \
-H "Authorization: Bearer $TOKEN" \
-d '{
  "model": "sdd-edit/sd-cpp-local",
  "image": "@path/to/your/image.jpg",
  "prompt": "Add a sunset background",
  "n": 1,
  "size": "1024x1024"
}'

# Embed
curl \
"https://llama-embed.reeseapps.com/v1/embeddings" \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{
  "model": "deskwork-embed/embed",
  "input":"This is the reason you ended up here:",
  "encoding_format": "float"
}'
```

## VLLM

### Run VLLM with Podman

```bash
# 'latest' and 'nightly' are both viable tags
podman run --rm \
--device /dev/kfd \
--device /dev/dri \
-v ~/.cache/huggingface:/root/.cache/huggingface:z \
--env "HF_TOKEN=$HF_TOKEN" \
-p 8010:8000 \
--ipc=host \
-e ROCBLAS_USE_HIPBLASLT=1 \
-e TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 \
-e VLLM_TARGET_DEVICE=rocm \
-e HIP_FORCE_DEV_KERNARG=1 \
-e RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1 \
docker.io/vllm/vllm-openai-rocm:nightly \
--enable-offline-docs \

# Pick your model
Qwen/Qwen3.5-35B-A3B-FP8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
Qwen/Qwen3.5-9B --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
Qwen/Qwen3.5-35B-A3B-FP8
google/gemma-4-26B-A4B-it
openai/gpt-oss-120b
```

## Misc

### Quantizing your own Models

```bash
# Create a scratch dir for downloading models
mkdir scratch && cd scratch

# qwen 3.5 35b
mkdir qwen3.5-35b-a3b && cd qwen3.5-35b-a3b
hf download --local-dir . Qwen/Qwen3.5-35B-A3B

# nemotron cascade
mkdir nemotron-cascade-2-30b-a3b && cd nemotron-cascade-2-30b-a3b
hf download --local-dir . nvidia/Nemotron-Cascade-2-30B-A3B

# Run the full 
podman run -it --rm \
--device=/dev/kfd \
--device=/dev/dri \
-v $(pwd):/models:z \
--entrypoint /bin/bash \
ghcr.io/ggml-org/llama.cpp:full-vulkan

# Run ./llama-quantize to see available quants
# 7 = q_8
# 18 = q_6_k
# 17 = q_5_k
# 15 = q_4_k
./llama-quantize /models/$MODEL_NAME.gguf /models/$MODEL_NAME-Q6_K.gguf 18
./llama-quantize /models/$MODEL_NAME.gguf /models/$MODEL_NAME-Q8_0.gguf 7
```

### Qwen3.5 Settings

> We recommend using the following set of sampling parameters for generation

- Non-thinking mode for text tasks: temperature=1.0, top_p=1.00, top_k=20, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0
- Non-thinking mode for VL tasks: temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
- Thinking mode for text tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
- Thinking mode for VL or precise coding (e.g. WebDev) tasks : temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0

> Please note that the support for sampling parameters varies according to inference frameworks.