Instructions to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Use Docker

docker model run hf.co/aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

LM Studio
Jan

vLLM

How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "aogavrilov/diffusiongemma-agent-iq3-cuda13"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aogavrilov/diffusiongemma-agent-iq3-cuda13",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Ollama
How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with Ollama:
```
ollama run hf.co/aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M
```

Unsloth Studio

How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for aogavrilov/diffusiongemma-agent-iq3-cuda13 to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for aogavrilov/diffusiongemma-agent-iq3-cuda13 to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for aogavrilov/diffusiongemma-agent-iq3-cuda13 to start chatting

How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with Docker Model Runner:
```
docker model run hf.co/aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M
```

Lemonade

How to use aogavrilov/diffusiongemma-agent-iq3-cuda13 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull aogavrilov/diffusiongemma-agent-iq3-cuda13:Q4_K_M

Run and chat with the model

lemonade run user.diffusiongemma-agent-iq3-cuda13-Q4_K_M

List all available models

lemonade list

diffusiongemma-agent-iq3-cuda13 / payload /app /docs /local_agent_stack.md

aogavrilov

Add files using upload-large-folder tool

ef2127a verified 16 days ago

preview code

Raw

History Blame Contribute Delete

12.6 kB

Local Agent Stack

This repo now has fifteen layers around the local DiffusionGemma runtime. The preferred entrypoint is:

cd /root/diffusiongemma-agent
scripts/dg_agent.sh status
scripts/dg_agent.sh up
scripts/dg_agent.sh doctor
scripts/dg_agent.sh capabilities --live

Docs: docs/dg_agent_cli.md.

OSS wrapper matrix: docs/oss_agent_wrappers.md.

Last Verified Runtime State

Checked on 2026-07-14 from the Windows checkout C:\Users\alexg\Downloads\diffusiongemma-agent with the active backend in WSL. Runtime availability is time-sensitive; recheck with scripts/dg_agent.sh status or scripts/dg_agent.sh doctor before relying on these endpoints.

Runtime:

backend  http://127.0.0.1:4100/healthz
proxy    http://127.0.0.1:8090/healthz
litellm  http://127.0.0.1:4100/v1/models
model    diffusiongemma-26b-a4b-it-iq3m-fullgpu
gpu      full offload profile, ngl=999, maxtok=768, about 14.7 GiB VRAM loaded

Installed and audited OSS agent surfaces:

Aider        active WSL runtime, 0.86.2
OpenCode     .tools/opencode/node_modules/.bin/opencode plus opencode-linux-x64, 1.17.20; provider/run/MCP/ACP smoke pass
MCP SDK      active WSL runtime, 1.28.1
Serena       WSL bridge, 1.5.3, live Pyright symbols
Qwen Code    .tools/qwen-code/node_modules/.bin/qwen, 0.19.10, WSL runner + MCP
LangGraph    WSL Python 3.14 venv from local cp314 wheelhouse; smoke-import/dry-run pass
Watchdog     repo-local status/up/watchdog stack-control scripts restored and smoke pass

Core verification completed on this checkout:

python scripts/dg_agent.py smoke --suite qwen-code --timeout 180
python scripts/dg_agent.py smoke --suite mcp --suite serena-mcp --timeout 180
python scripts/dg_agent.py smoke --suite opencode-provider --suite opencode-run --suite opencode-mcp --suite opencode-acp --timeout 180
python scripts/dg_agent.py smoke --suite langgraph --timeout 240
python scripts/dg_agent.py smoke --suite watchdog --suite stack-control --timeout 60
python scripts/dg_agent.py smoke --suite wrappers --timeout 180

OpenCode is installed from the Windows npm payload for PowerShell usage and the Linux optional package opencode-linux-x64 for WSL MCP process spawning. Live opencode mcp list verifies dg_agent and repomix as connected. Serena has a separate MCP smoke with live Pyright symbols and is intended for IDE bundles.

The verified Aider end-to-end scenario changed only an allowed Python file through the GPU gateway, then passed an independent WSL Python assertion. The temporary Aider histories stayed outside the target repository. Optional OpenHands, AutoGen, smolagents, CrewAI, Open Interpreter, LlamaIndex, and SWE-family wrappers remain separate experiments and must be smoke-tested individually before use.

External clients should prefer the client pack plus MCP resources instead of large prompts. The exported local_agent profile includes artifact commands: scripts/dg_agent.sh agent-runs list, scripts/dg_agent.sh agent-runs show --latest, and scripts/dg_agent.sh agent-runs artifact transcript --latest.

1. Fast Backend

http://127.0.0.1:4100/v1

Start from Windows:

powershell -ExecutionPolicy Bypass -File \\wsl.localhost\Ubuntu-24.04\root\diffusiongemma-agent\scripts\start_agent_fast_service_windows.ps1

2. Aider-Compatible Model Proxy

http://127.0.0.1:8090/v1

This proxy compresses Aider prompts, repairs malformed output fences, applies exact single-file return constraints when the model output is malformed, and limits the model output to a size the current DG profile can handle.

Start:

powershell -ExecutionPolicy Bypass -File \\wsl.localhost\Ubuntu-24.04\root\diffusiongemma-agent\scripts\start_aider_proxy_windows.ps1

3. Reliable Automated Edits

Use the agent profile for the normal Codex-like local workflow:

/root/diffusiongemma-agent/scripts/dg_agent.sh agent \
  --repo /path/to/repo \
  --task "Fix the failing test and keep the public API unchanged" \
  --file path/to/file.py

This wraps the context packer, task runner, automatic verification, rollback on failure, deterministic repair fallbacks for exact small edits, and preserved session artifacts into one command.

Use the lower-level task runner for explicit multi-step plans:

/root/diffusiongemma-agent/scripts/run_task_runner.sh \
  --repo /path/to/repo \
  --plan /path/to/plan.json \
  --report /tmp/dg-task-report.json \
  --rollback-on-failure

This is the most reliable current mode because it splits work into bounded steps, verifies each step, can repair common exact Python return edits when the model/Aider output is malformed, can handle explicit single-occurrence text replacement tasks, and can rollback a failed step patch.

Docs: docs/task_runner.md.

4. Interactive Web/API Agent

Use AgentAPI when you want a ready-made web/API surface over Aider:

/root/diffusiongemma-agent/scripts/run_agentapi_aider.sh \
  --repo /path/to/repo \
  --port 3284

Open:

http://127.0.0.1:3284/chat

Docs: docs/agentapi_aider.md.

OpenCode Compact Delegate

The primary interactive terminal route is the compact OpenCode delegate:

Set-Location C:\path\to\repo
C:\Users\alexg\Downloads\diffusiongemma-agent\scripts\run_opencode_agent_windows.ps1

or through the unified CLI:

scripts/dg_agent.sh opencode-agent -- /path/to/repo

It uses the upstream OpenCode interface but exposes one bash tool. The safe gateway immediately delegates that tool to the existing read/persistent-supervisor runner, so a task gets ranked retrieval, narrow deterministic repairs when applicable, Aider for broader file edits, verification, and rollback. This avoids relying on native tool-calling from the diffusion model. The compact profile uses a 450-second OpenCode bash deadline to cover the 420-second bounded session.

The model remains on the GPU; OpenCode, retrieval, Git checks, and MCP process management run on CPU and do not add model VRAM residency.

For semantic navigation before wide work, use the default IDE MCP bundle with Serena, or repo-map/code-outline. Serena is not auto-mounted here because its cold start exceeds OpenCode's MCP connect timeout.

Experimental: Generic OpenCode

configs/opencode.dg.json points generic OpenCode at the same local proxy. Node.js and OpenCode are installed locally under .tools/, and provider discovery is validated with scripts/smoke_opencode_local.sh. A minimal opencode run fallback is validated with scripts/smoke_opencode_run_fallback.sh; it confirms that the CLI does not hang or expose proxy 500s, not that OpenCode edits are reliable with this model.

configs/opencode.dg-mcp.json mounts the DG MCP server and upstream Repomix MCP. scripts/dg_agent.sh opencode-mcp -- /repo launches that profile as a TUI, and scripts/dg_agent.sh opencode-acp -- --cwd /repo --hostname 127.0.0.1 --port 3295 exposes OpenCode's ACP server over the same profile.

Docs: docs/opencode_local.md.

Experimental: Goose

Goose is installed locally under .tools/goose and launched through the same OpenAI-compatible proxy:

/root/diffusiongemma-agent/scripts/run_goose_local.sh run \
  --no-profile \
  --max-turns 1 \
  --text "Reply exactly OK."

It is useful as a ready-made OSS agent shell with sessions, MCP/extensions, review, and TUI commands. Treat it as experimental with the current DG model: native tool-calling is still the weak point, so reliable edits should stay on the Aider/task-runner path.

Docs: docs/goose_local.md.

OpenAI Gateway: LiteLLM

LiteLLM is available as a stable external compatibility layer:

/root/diffusiongemma-agent/scripts/run_litellm_gateway.sh

External clients can use:

base_url: http://127.0.0.1:4100/v1
api_key: dummy
model: diffusiongemma-local

Docs: docs/litellm_gateway.md.

Checked-in client profiles:

configs/client_profiles/openai-compatible.local.json
configs/client_profiles/openai.env
configs/client_profiles/continue.config.yaml

Docs: docs/ide_client_profiles.md.

Watchdog

Use the watchdog to check and recover the local runtime stack:

/root/diffusiongemma-agent/scripts/dg_agent.sh status
/root/diffusiongemma-agent/scripts/dg_agent.sh up --restart
/root/diffusiongemma-agent/scripts/dg_agent.sh watchdog -- status
/root/diffusiongemma-agent/scripts/dg_agent.sh watchdog -- ensure --restart

It covers backend 4100, Aider proxy 8090, and LiteLLM-compatible 4100/v1.

Docs: docs/stack_watchdog.md.

Recommended Use

Default coding-agent mode: scripts/dg_agent.sh agent --repo /repo --task "..." --file path
One-command coding-agent mode: scripts/dg_agent.sh run --repo /repo --task "..." --file path --start
OSS wrapper status: scripts/dg_agent.sh wrappers
OSS wrapper audit/install: scripts/dg_agent.sh bootstrap and scripts/dg_agent.sh bootstrap --install
Client/IDE profile export: scripts/dg_agent.sh client-pack --write
Target repo setup: scripts/dg_agent.sh workspace-init --repo /repo
Repo-local launchers: .dg-agent/bin/run --task "..." --file path --start
Repo-local health: .dg-agent/bin/capabilities --latest, .dg-agent/bin/doctor, .dg-agent/bin/status
Repo-local work: .dg-agent/bin/plan, .dg-agent/bin/edit, .dg-agent/bin/task, .dg-agent/bin/supervisor, .dg-agent/bin/web
Target repo preflight: scripts/dg_agent.sh preflight --repo /repo --task "..." --file path
Capability audit: scripts/dg_agent.sh capabilities or scripts/dg_agent.sh capabilities --live
Last capability report: scripts/dg_agent.sh capabilities --latest
Small direct code edit: scripts/run_supervisor_agent.sh.
Repo context pack: scripts/dg_agent.sh context --repo /repo --task "..."
Repo-scale compact RAG retrieval: scripts/dg_agent.sh rag --repo /repo --task "..." --print-context
OSS repo packer: scripts/dg_agent.sh repo-pack --repo /repo --include "src/**" --style markdown --stdout
Generated plan: scripts/dg_agent.sh plan --repo /repo --task "..."
Full session with artifacts: scripts/dg_agent.sh session --repo /repo --task "..." --auto-test
Session history: scripts/dg_agent.sh sessions list and scripts/dg_agent.sh sessions show --latest
Session diff/artifacts: scripts/dg_agent.sh sessions diff --latest and scripts/dg_agent.sh sessions artifact context_md --latest
Natural-language edit: scripts/dg_agent.sh edit --repo /repo --task "..." --auto-test
Standalone verification: scripts/dg_agent.sh verify --repo /repo --file path
Multi-file or multi-step task: scripts/dg_agent.sh task with an explicit plan.
Browser/API interaction: scripts/dg_agent.sh web.
Primary OpenCode terminal route: scripts/dg_agent.sh opencode-agent.
Generic Codex-like TUI experiment: scripts/dg_agent.sh opencode.
Goose/MCP agent experiment: scripts/dg_agent.sh goose -- run --no-profile --max-turns 1 --text "..."
OpenHands MCP setup: scripts/dg_agent.sh openhands-mcp -- --repo /repo --reset
Qwen Code read-only inspection: scripts/dg_agent.sh qwen-code -- --repo /repo -- --prompt "Read path/to/file.py. Summarize ..."
AutoGen AgentChat framework experiment: scripts/dg_agent.sh autogen -- --repo /repo --dry-run
smolagents CodeAgent framework experiment: scripts/dg_agent.sh smolagents -- --repo /repo --dry-run
LangGraph graph-agent framework experiment: scripts/dg_agent.sh langgraph -- --repo /repo --dry-run
CrewAI multi-agent framework experiment: scripts/dg_agent.sh crewai -- --repo /repo --dry-run
Open Interpreter shell experiment: scripts/dg_agent.sh open-interpreter -- --repo /repo --dry-run
LlamaIndex RAG/agent framework experiment: scripts/dg_agent.sh llamaindex -- --repo /repo --dry-run
Haystack BM25 RAG: scripts/dg_agent.sh haystack -- --repo /repo --retrieve-only --task "..."
Persistent local coding controller: scripts/dg_agent.sh autonomous -- --repo /repo --task "..."
MCP HTTP endpoint: scripts/dg_agent.sh mcp-http -- --host 127.0.0.1 --port 8765
Serena semantic MCP: scripts/dg_agent.sh serena-mcp
External OpenAI-compatible clients: scripts/dg_agent.sh litellm.
Validate IDE/OpenAI-compatible profiles: scripts/dg_agent.sh smoke --suite gateway-clients.
Recover local services: scripts/dg_agent.sh watchdog -- ensure --restart.
One-command stack control: scripts/dg_agent.sh status, scripts/dg_agent.sh up, scripts/dg_agent.sh down.