DFlash Console

Local control panel for DFlash speculative-decoding stacks and a unified local model runtime — manage llama-server engines, Piper TTS, Whisper STT, embeddings, model libraries, Hugging Face downloads, and GPU settings from a single web UI.

Status: Public preview. This project is intended for local, single-user Windows deployments.

Developer: ILAN AVIV · UI: http://127.0.0.1:8900/ · Version: v0.3.106

Download (Windows)

Package	Link
Setup installer (recommended)	Latest GitHub Release — DFlash-Console-Setup-*-x64.exe
Portable	Same Releases page — DFlash-Console-Portable-*-x64.exe when published
CLI only	pip install dflash-console then dflash serve

Installed desktop apps check for updates through a separate signed feed (not GitHub). New installs: use the GitHub Release installer above.

---

What it does

DFlash Console is a standalone FastAPI + vanilla JavaScript app that sits beside your DFlash installation. It gives you an LM Studio–style workbench focused on your engine profiles—not a generic chat client.

Area	Highlights
Engines	Start/stop llama-server routers, load & eject models in parallel, live boot progress, token stats on cards, developer logs with clear button
Runtimes	Unified multi-modal layer: Piper TTS, whisper.cpp STT, embeddings — Console-proxied OpenAI routes, adapter registry, shared ports, process identity
Models	Full PC library: DFlash, GGUF, Ollama, LM Studio, Piper, Whisper, OCR, and embeddings
Model catalog	Browse and download Hugging Face models into configured library locations
Settings	GPU strategy, model storage, engine network/API, MCP client preview, Locations panel with config/preset import-export
Documentation	In-app API reference, user guide, terminal CLI, and release notes
Terminal CLI	dflash command: list, load, chat, embed, delete, nodes, settings, search, pull. Install with pip install dflash-console
About	Developer attribution, version, license, runtime boundary, and public project links
Discovery	Scan PC finds model folders; Add folder opens a drive-aware browser (C:, D:, …)

Router mode uses --models-preset with load/unload over HTTP so engines stay listening while models swap.

---

Multi-modal runtime

DFlash Console is now a unified local model runtime, not just a GGUF control panel. A small adapter registry (core/runtimes/) lets the Console discover, load, monitor, and proxy OpenAI-shaped APIs for many model families, while llama-server / DFlash stacks stay the first-class GGUF path.

Modality	Runtime	Engine	Mode
LLM chat / instruct	llama-server	GGUF (existing)	server
LLM chat (SafeTensors)	vLLM or Transformers	Hugging Face folders (runtimes/vllm/, runtimes/transformers/)	server
Embeddings	llama-server embedding profile	GGUF (existing)	server
Text-to-speech	Piper	ONNX voices (runtimes/piper/)	CLI
Speech-to-text	whisper.cpp whisper-server	GGUF-whisper (runtimes/stt/)	server
Vision / OCR	llama-server + mmproj	GGUF + projector (existing)	server

Runtime bundles (thin installer / fat data root)

Engines and voices live outside the Electron installer under the Console data root:

runtimes/
├── piper/                 # piper.exe + espeak-ng-data + voices/*.onnx(+json)
│   └── manifest.json
├── stt/                   # whisper-server.exe + ggml-cuda.dll + ...
│   └── manifest.json
└── process-tokens.json    # shared managed-process identity (read by server.ps1)

Each adapter writes a manifest.json at boot, exposed read-only at GET /api/runtimes/manifests. Piper and Whisper ship as native bundles. vLLM and Transformers download on demand from Settings or the first-run wizard so the Windows installer stays small.

Playground modes

The Playground has four modes:

Mode	What it does
Chat	Chat completions against llama-server / DFlash, or vLLM / Transformers when that engine is loaded
Speak	Text → Piper WAV (voice picker, speed, download)
Transcribe	Pick a Whisper model → upload audio → transcript
Embed	One item per line → vectors; Export .jsonl

Console-proxied OpenAI routes

Children stay on internal loopback ports; clients only talk to the Console:

Route	Upstream
/api/servers/{id}/v1/chat/completions	llama-server (existing)
/api/runtimes/piper/v1/audio/speech	Piper (CLI)
/api/runtimes/stt/v1/audio/transcriptions	whisper-server (multipart → /inference)
/api/servers/{id}/v1/embeddings	llama-server embedding profile
/api/servers/{id}/embed/batch	batch embed + .jsonl export

Console OpenAI gateway (port 8001)

Point any OpenAI-compatible app at one stable base URL — http://127.0.0.1:8001/v1 — and the gateway proxies to the loaded engine (chat, embeddings, TTS, STT). Model names are tolerant (engine id, checkpoint id, or any alias like gpt-4o); chat JIT-loads and streams SSE. Configure in Settings → Engine profiles → Console OpenAI gateway (gateway_port, gateway_server_id); see USER-GUIDE §8.

Process identity & cleanup

Every adapter contributes path-specific process-identity tokens (runtimes\piper\piper.exe, runtimes\stt\whisper-server) to a shared registry. managed_process_identity(), server.ps1 stop/shutdown cleanup, and restart adoption recognise Console-managed children without ever adopting foreign processes (e.g. another app's bundled Piper). Stop/unload kills children and frees VRAM with no orphans.

Config

Non-llama runtimes are configured in config.json under runtimes[]; the servers[] list stays exclusively for llama-server / DFlash / GGUF:

"runtimes": [
  { "id": "tts-main", "runtime_id": "piper", "port": 0, "device_policy": "cpu" },
  { "id": "stt-main", "runtime_id": "stt",  "port": 8910, "device_policy": "gpu" }
]

GET /api/runtimes merges servers[] (synthesised runtime_id: llama-server) with runtimes[]. Ports are unique across ui/servers/runtimes (shared registry); CLI runtimes use port: 0.

---

Recent improvements

v0.3.105 — public preview documentation

Feature	Description
Docs	README, user guide, CLI, About page, and in-app Documentation aligned for launch
Install	GitHub Releases + PyPI + single-server rule documented everywhere

v0.3.104 — pip on PyPI and single-server takeover

Feature	Description
PyPI	pip install dflash-console is live on PyPI
dflash serve	Stops a foreign Console on port 8900 (dev, EXE, or older pip) before starting
Public preview	GitHub Releases for the Windows installer; Discussions and Issues for feedback

v0.3.98 — pip install and extra terminal commands

Feature	Description
pip package	pip install dflash-console then dflash serve. The command is dflash.
Terminal	dflash embed, dflash delete, dflash nodes, and dflash settings.

v0.3.30 — choose vLLM in the toolbar

Feature	Description
Engine menu	Pick vLLM, Transformers, or DFlash on Engines, Playground, and Models before you load.

v0.3.29 — vLLM engine

Feature	Description
vLLM	Choose vLLM to load Hugging Face models. The engine downloads after install. The first-run wizard can install it on a new PC.

v0.3.28 — Windows installer

Feature	Description
Desktop update	Production installer with the terminal CLI, full local library list, source filters, and Downloads page.

v0.3.27 — full library list and source filters

Feature	Description
dflash list	Shows the same full local library as the Models tab. Filter with --ollama, --lmstudio, --dflash, or --source.

v0.3.26 — dflash works in PowerShell

Feature	Description
dflash command	Type dflash list in any PowerShell window. The installer registers the command so you do not need .\dflash.

v0.3.25 — terminal CLI

Feature	Description
dflash command	PowerShell can list models, check what is loaded, search, download, and chat with the running Console, the same way ollama talks to its server.

v0.3.24 — last downloads from this PC

Feature	Description
Last downloads	The Downloads page lists models already on this PC and lets you filter the last 24 hours, 7 days, 30 days, 90 days, or 12 months.

v0.3.6 — full mobile model names

Feature	Description
Engine cards	Narrow views keep the complete model name visible before the compact status details.

v0.3.4 — clearer catalog model variants

Feature	Description
Catalog model cards	Cards show approximate disk size beneath the update age and identify accelerator checkpoints, GGUF models, and full model weights at a glance.

v0.3.3 — higher sidebar restore control

Feature	Description
Collapsed sidebar arrow	The restore control is raised while preserving its original colors and appearance.

v0.3.2 — responsive engine cards

Feature	Description
Portrait/split-window layout	Engine cards switch to compact summaries at narrow browser widths; long names stay on one line while status, memory, token, and unload information remain visible. Desktop card layout is unchanged.

v0.3.0-dev — multi-modal runtime

Feature	Description
Piper TTS	Native bundle, Playground Speak, POST .../v1/audio/speech, per-runtime logs
Whisper STT	whisper.cpp whisper-server bundle, Playground Transcribe, proxied multipart transcriptions
Embeddings	Proxied /v1/embeddings, batch embed + .jsonl export, Playground Embed
Runtime registry	Adapter protocol/registry, runtimes[] config, shared port registry, GET /api/runtimes
Process identity	Path-specific tokens; server.ps1 + supervisor clean up only Console-owned children
Vision polish	Inspector vision row; image attach gated on model capability
Hardening	Bundle manifests (/api/runtimes/manifests), sandbox/no-shell, per-adapter smoke tests

v0.2.9

Feature	Description
Model catalog	HF README titles/descriptions, lab filter, size/age on list cards, install detection, card download progress
Download queue	Global downloads tray, Models tab Downloading filter with live progress bars
External API	/api/endpoints, /api/installed, /api/console/logs, and request logging for integrations
Playground	Load models from catalog via engine + model picker
UI polish	Themed dropdowns, settings/docs as main views, terminology cleanup (checkpoint → model)
Portable runtime	Electron uses an external Console data root so the installer stays small and model files remain local
Production boundary	Loopback validation, path-safe model/projector handling, release preflight, and public repository security policy

Earlier improvements (v0.0.23)

Feature	Description
Live token stats	Loaded engine cards show generated tokens and speed; Generating Xs updates every second during inference
Parallel loading	Load multiple engines at once without blocking the UI
Engine card UX	Click a loaded card to open the runtime inspector; right-click for context menu (details, copy URL, unload, etc.)
Accurate CPU meter	System bar CPU reading matches Task Manager (process-time delta, not inflated WMI load)
Cleaner load progress	Solid progress bar without conflicting slide animation
Locations settings	One panel for config file, DFlash install, logs, presets, and console URL; import/export config and launch presets
Clear engine logs	One-click clear in the developer log header
Chat proxy fix	Long completions no longer freeze live stats polling
In-app docs	Documentation tab with overview, user guide, and full API catalog

---

Requirements

• Windows (primary target; PowerShell startup scripts)
• Python 3.10+
• PowerShell 7+ (pwsh)
• Node.js 22.12+ (Electron shell and packaging)
• Built llama-server and compatible model files in the configured DFLASH_ROOT
• NVIDIA GPU optional but recommended for multi-GPU load settings

---

Quick start

Install with pip

pip install dflash-console
dflash serve

Open http://127.0.0.1:8900/. The PyPI package is dflash-console; the terminal command is dflash. Model weights and llama-server stay on this PC (DFLASH_ROOT or Settings). pip install dflash is a different project.

From source

git clone https://github.com/ilan4ever/Dflash-Console.git
cd Dflash-Console

# First-time setup
copy config.example.json config.json
# Edit config.json — set dflash_root, server ports, model paths, and enable a server profile.
# Keep model weights, logs, and local credentials out of Git.

.\server.ps1

The example server profile is disabled until you point it at a model available on your machine. This prevents a fresh public checkout from trying to launch excluded model assets.

Open http://127.0.0.1:8900/ in your browser.

Command line (like ollama)

The Console is the source of truth for models on this PC. dflash list shows the same full library as the Models tab, including Ollama and LM Studio.

pip install dflash-console
dflash serve
dflash help
dflash list
dflash list --ollama
dflash list --lmstudio
dflash list --dflash
dflash list --source library
dflash ps
dflash load qwen
dflash chat "hello"
dflash embed "index this"
dflash delete old-model
dflash nodes
dflash settings
dflash search qwen

From a checkout you can also run .\dflash.ps1 install.

Command	What it does
dflash serve / open	Start the UI / open the browser
dflash list / ps / show	Library, loaded models, one model
dflash load / unload / start / stop	Engines
dflash chat / embed	Prompt or vectors
dflash delete	Remove a local model from disk
dflash nodes	Other Consoles on your network
dflash settings	Show or change ports and paths
dflash search / pull / downloads	Hugging Face catalog
dflash hardware / stats / report / logs	This PC
dflash api GET /api/health	Any Console HTTP route

Use dflash <command> --help for flags. --json prints raw server data. The Console must be running (dflash serve or .\server.ps1) except for help, version, serve, and install. Full command list: docs/CLI.md and the public site onevoiceai.in/dflash-console/docs.

run.ps1 performs a full developer reset: it releases managed model VRAM, stops configured engines, and starts a clean Console instance.

Foreground mode (attach logs to the terminal):

.\server.ps1 -Foreground

Full restart (release GPU + stop engines):

.\server.ps1 -Restart

Restart API after backend edits:

.\scripts\restart-console-server.ps1

This is a gentle API restart. It preserves engine listeners, then the new Console process adopts them and leaves router models unloaded. The UI auto-refreshes when the API restarts (/api/health boot id).

Desktop app (Electron)

The desktop shell opens the same Console UI in a native window and starts the local API if it is not already running. The Windows installer is a desktop shell; the Console data root remains separate so model weights and llama-server binaries are not copied into every install. Set DFLASH_CONSOLE_ROOT to a checkout containing server.ps1, api, and static, or choose that folder when the packaged app first starts. The data root still requires Python 3.10+ and PowerShell 7+.

.\scripts\run-electron.ps1

Build Windows packages (branded DFlash setup + portable exe):

.\scripts\run-electron.ps1 -Build

Output lands in dist-electron/. The shell uses the configured ui_port (8900 by default) and talks to loopback, so the look and behavior match the browser. Closing the window leaves the Console API and engines running, same as closing a browser tab. Windows artifacts are not code-signed by default; publishers should sign them before distributing outside a trusted team.

---

Using the app

See docs/USER-GUIDE.md for a full walkthrough, or open Documentation → User guide in the sidebar. The About page is available in both the browser UI and the Electron desktop app.

Typical workflow:

1. Open Engines and turn on an engine profile (toggle or Load).
2. Pick a model from the dropdown and click Load, or load from the Models tab.
3. Watch boot progress and live stats on the card; click the card for runtime settings in the side panel.
4. Point your app at the engine OpenAI URL shown on the card, or use the console proxy at /api/servers/{id}/v1/chat/completions.
5. Use the Playground mode switcher (Chat · Speak · Transcribe · Embed) for TTS, STT, and embedding workflows.
6. Use Settings → Locations to back up or restore config.json and launch presets.
7. Open About for the current release, developer attribution, license, and public project links.

---

Third-party model libraries and trademarks

DFlash Console can scan model folders that the user explicitly enables, including local LM Studio libraries. It reads and loads model files already present on the user's machine; it does not include, copy, or redistribute the LM Studio application or model weights.

“LM Studio library” identifies the detected folder source only. LM Studio is a trademark of Element Labs, Inc.; DFlash Console is independent and is not affiliated with or endorsed by LM Studio. Model files remain subject to the licenses and terms provided by their respective model authors or distributors. Users are responsible for checking those terms before copying, publishing, or commercially deploying model files.

---

Configuration

Variable / file	Purpose
config.json	Local settings (not committed — copy from config.example.json)
models_root / ./models	Developer model library inside the Console app folder
DFLASH_ROOT_OVERRIDE	Optional explicit override for the configured DFlash root
config.json → dflash_root	Path to DFlash repo with llama-server binaries and launch scripts
config.json → servers[]	llama-server / DFlash engine profiles: port, GPU layers, context, idle unload
config.json → runtimes[]	Non-llama runtimes (Piper, STT): runtime_id, port, device_policy
config.json → model_libraries[]	Folders scanned for local models and HF download targets

Default UI port is 8900. Engine ports (e.g. 8090, 8092) are configured per server block.

---

Security boundary

The Console binds to loopback by default and validates configured engine URLs as loopback-only. It is designed for one trusted user on one machine; it does not provide multi-user authentication or CSRF protection. Do not expose the Console or engine ports to a LAN, reverse proxy, or public network without adding an authenticated access layer.

Hugging Face downloads use HF_TOKEN or HUGGING_FACE_HUB_TOKEN when private repositories require them. Keep those values in the process environment, never in config.json or committed files.

---

Project layout

Dflash-Console/
├── api/app.py              # FastAPI routes
├── core/                   # Config, runtime, model discovery, HF, GPU, inference stats
│   └── runtimes/           # Adapter registry + protocol (piper, stt, noop, contention)
├── runtimes/               # Installed engine bundles (piper/, stt/) + manifests (gitignored)
├── static/                 # UI (HTML, CSS, JS)
├── electron/               # Desktop shell (Electron)
├── scripts/                # Server start / restart / Electron helpers
├── tests/                  # pytest suite
├── dflash_cli/             # `dflash` terminal command
├── pyproject.toml          # pip package `dflash-console`
├── docs/
│   ├── USER-GUIDE.md       # End-user walkthrough
│   ├── CLI.md              # Full terminal command list
│   ├── ARCHITECTURE.md     # Public architecture overview
│   ├── LICENSING.md        # AGPL and redistribution guide
│   ├── RELEASING.md        # GitHub and Windows release process
│   └── ui/                 # UI panel design notes
├── run.ps1                 # Full reset and background API startup
├── package.json            # Electron desktop packaging
├── config.example.json     # Template configuration
├── CONTRIBUTING.md         # Contribution workflow and AGPL terms
├── CODE_OF_CONDUCT.md      # Community standards
├── SUPPORT.md              # Questions, bugs, and contact channels
├── SECURITY.md             # Vulnerability reporting
├── TRADEMARKS.md           # DFlash name and logo policy
├── NOTICE.md               # Copyright and third-party notices
└── LICENSE                 # GNU AGPL v3 or later

---

API (selected)

Method	Endpoint	Description
GET	/api/health	Liveness + boot_id (UI reload watch)
GET	/api/servers	All engine statuses + inference stats
POST	/api/servers/{id}/load	Load model (optional runtime JSON body)
POST	/api/servers/{id}/unload	Eject model (router stays up)
POST	/api/servers/{id}/v1/chat/completions	Proxy chat; updates live token stats
POST	/api/servers/{id}/v1/embeddings	Text → vectors (embedding profile, OpenAI shape)
POST	/api/servers/{id}/embed/batch	Batch embed text items; optional .jsonl export
GET	/api/runtimes	Unified runtime list (servers[] + runtimes[] + adapters)
GET	/api/runtimes/manifests	Installed runtime plugin manifests + process tokens
POST	/api/runtimes/piper/v1/audio/speech	Text → speech (Piper, OpenAI shape)
POST	/api/runtimes/stt/v1/audio/transcriptions	Audio → text (whisper.cpp, OpenAI shape)
GET	/api/models	Full local library (same as Models tab). source=ollama|lmstudio|dflash|library
DELETE	/api/models/file	Delete a local model file or Hugging Face folder
GET / POST / DELETE	/api/nodes	Remote Console nodes
GET / PUT	/api/config	Settings (dflash settings)
GET	/api/hf/downloads	Current downloads and last-download history
POST	/api/models/load	Unified loader — load ANY catalog model by path; dispatches by modality
GET	/api/gateway	Console OpenAI gateway status (port, url, running, default server, routes)
GET	/api/docs/catalog	In-app documentation JSON
GET	/api/model-libraries/scan	PC scan for model folders
GET	/api/fs/browse	Folder picker for library paths
GET	/api/hf/search	Hugging Face model search
POST	/api/hf/download	Download GGUF into a library
POST	/api/hf/install	Search → download → load in one call
GET	/api/status/loaded	Currently loaded models (engines + runtimes)
GET	/api/status/report	Full report: CPU/RAM/VRAM, engines, loaded models
DELETE	/api/logs/{id}	Clear engine log file

Full route list: api/app.py or Documentation tab in the UI.

---

Development

pip install -e .[dev]
pip install -r requirements-dev.txt
pytest
dflash serve

Cursor agents: see .cursor/rules/dev-server-restart.mdc — restart the API after Python backend changes.

---

Related projects

• DFlash — speculative decoding stack and llama-server profiles this console manages
• Model weights and native build outputs are intentionally excluded from this repository.

---

Community and support

Use the repository's public channels so questions and fixes are easy for other users to find:

• Questions and setup help: GitHub Discussions
• Bug reports: Issue tracker
• Feature ideas: Feature request
• Security vulnerabilities: follow SECURITY.md and use GitHub's private reporting channel; do not post exploit details publicly.
• Developer: ILAN AVIV

Please include the Console version from About, Windows and Python versions, the relevant logs with secrets removed, and a minimal reproduction. Never include config.json, access tokens, model weights, or private logs.

---

License

This project is licensed under the GNU Affero General Public License version 3 or later (AGPL-3.0-or-later). See LICENSE.

AGPL-covered source may be used, studied, modified, and redistributed under the license terms. If you distribute a modified version, or run a modified version as a network service, the AGPL source-sharing obligations apply. The DFlash name, logo, and other marks are not granted by the software license; see TRADEMARKS.md. Third-party runtimes, binaries, model weights, and other assets remain under their own licenses; see NOTICE.md. The AGPL migration begins with the 0.3.0 release; earlier versions retain the license that accompanied their distribution.

---

Credits

DFlash Console is developed and maintained by ILAN AVIV. The project is free software under the GNU AGPL v3 or later. See the developer profile, source repository, and related DFlash project.

When you redistribute or build on DFlash Console, retain the copyright and license notices and provide the corresponding source as required by the AGPL. Please also include a link to the DFlash Console source repository in your README, About page, or other project documentation.

---

Roadmap

See docs/ARCHITECTURE.md for the public architecture overview. Run python scripts/release-preflight.py before packaging a checkout. Public release discussions and roadmap items belong in the GitHub issue tracker.