- Everywhere: web, iOS, macOS, Android, Windows, Linux.
- Fast: exceeds average reading speed on all platforms except web.
- Private: No network connection, server, cloud required.
- Forward compatible: Any model compatible with llama.cpp. (so, every model.)
- Full OpenAI compatability: chat messages, multimodal/image support via LLaVa models, and function calling. (constrain outputs to valid JSON based on a JSON schema)
- Bare metal interface: call LLMs without being constrained to a chat implementation.
- Use with FONNX for RAG
- Combine with FONNX to have a full retrieval-augmented generation stack available on all platforms.
Platform | Status |
---|---|
Android | |
iOS | |
Linux | |
macOS | |
Web | |
Windows |
- Web is now based on WASM compiled from FLLAMA itself, rather than just llama.cpp, guaranteeing native/web parity.
- Tokenizing strings based on the model is 1000x faster, via caching the model. Went from O(300 ms) on native only to O(0.2 ms) on web, O(0.00001 ms) on native. This enables calculating what strings will be in context based on the context size.
- Method renames for consistency, correctness, and clarity. (ex. remove *Async from names, because all methods are async; rename
fllamaChatCompletionAsync
tofllamaChat
) - Document methods and updated example. TL;DR: Use
fllamaChat
unless you're doing something funny with LLMs that isn't user-facing, it will act like a true text completion engine instead of a chatbot.
- Add this to your package's pubspec.yaml file:
dependencies:
fllama:
git:
url: https://github.com/Telosnex/fllama.git
ref: main
- Run inference:
import 'package:fllama/fllama.dart';
String latestResult = "";
final request = OpenAiRequest(
maxTokens: 256,
messages: [
Message(Role.system, 'You are a chatbot.'),
Message(Role.user, messageText),
],
numGpuLayers: 99, /* this seems to have no adverse effects in environments w/o GPU support, ex. Android and web */
modelPath: _modelPath!,
mmprojPath: _mmprojPath,
frequencyPenalty: 0.0,
// Don't use below 1.1, LLMs without a repeat penalty
// will repeat the same token.
presencePenalty: 1.1,
topP: 1.0,
// Proportional to RAM use.
// 4096 is a good default.
// 2048 should be considered on devices with low RAM (<8 GB)
// 8192 and higher can be considered on device with high RAM (>16 GB)
// Models are trained on <= a certain context size. Exceeding that # can/will lead to completely incoherent output.
contextSize: 2048,
// Don't use 0.0, some models will repeat the same token.
temperature: 0.1,
logger: (log) {
// ignore: avoid_print
print('[llama.cpp] $log');
},
);
fllamaChat(request, (response, done) {
setState(() {
latestResult = response;
});
});
Web is extremely slow, ex. on a MBP M2 Max with 64 VRAM, it does ~2 tokens/second with a 3B parameter model. It's best seen as validation for your users that you will strive to support free LLMs everywhere.
To install: copy the fllama_wasm* files from example/web
to your app's web
directory, then add the <script>
tag in example/web/index.html
.
3 top-tier open models are in the fllama HuggingFace repo.
- Stable LM 3B is the first LLM model that can handle RAG, using documents such as web pages to answer a query, on all devices.
Mistral models via Nous Research. They trained and finetuned the Mistral base models for chat to create the OpenHermes series of models.
- Mistral 7B is best on 2023 iPhones or 2024 Androids or better. It's about 2/3 the speed of Stable LM 3B and requires 5 GB of RAM.
- Mixtral 8x7B should only be considered on a premium laptop or desktop, such as an M-series MacBook or a premium desktop purchased in 2023 or later. It's about 1/3 the speed of Stable LM 3B and requires 26 GB of RAM.
Roughly: you'll need as much RAM as the model file size. If inference runs on CPU, that much regular RAM is required. If inference runs on GPU, that much GPU RAM is required.
HuggingFace, among many things, can be thought of as GitHub for AI models. You can download a model from HuggingFace and use it with fllama. To get a download URL at runtime, see below.
String getHuggingFaceUrl(
{required String repoId,
required String filename,
String? revision,
String? subfolder}) {
// Default values
const String defaultEndpoint = 'https://huggingface.co';
const String defaultRevision = 'main';
// Ensure the revision and subfolder are not null and are URI encoded
final String encodedRevision =
Uri.encodeComponent(revision ?? defaultRevision);
final String encodedFilename = Uri.encodeComponent(filename);
final String? encodedSubfolder =
subfolder != null ? Uri.encodeComponent(subfolder) : null;
// Handle subfolder if provided
final String fullPath = encodedSubfolder != null
? '$encodedSubfolder/$encodedFilename'
: encodedFilename;
// Construct the URL
final String url =
'$defaultEndpoint/$repoId/resolve/$encodedRevision/$fullPath';
return url;
}
FLLAMA is licensed under a dual-license model.
The code as-is on GitHub is licensed under GPL v2. That requires distribution of the integrating app's source code, and this is unlikely to be desirable for commercial entities. See LICENSE.md.
Commercial licenses are also available. Contact [email protected]. Expect very fair terms: our intent is to charge only entities, with a launched app, making a lot of money, with FLLAMA as a core dependency. The base agreement is here: https://github.com/lawndoc/dual-license-templates/blob/main/pdf/Basic-Yearly.pdf
- Best practice is:
- Check out latest llama.cpp.
- Copy checked out llama.cpp to macos/llama.cpp. Commit. DO NOT PUSH.
- Try running example app, fix, repeat.
- Commit. DO NOT PUSH.
- Copy macos/llama.cpp to ios/llama.cpp. DO NOT PUSH.
- Try running example app in iOS simulator.
- Fix, repeat.
- Commit. DO NOT PUSH.
- Copy checked out llama.cpp to src/llama.cpp.
- Commit. Push. Watch codemagic, particularly for errors on Android/Windows/Linux.
- When codemagic is green, copy src/llama.cpp to wasm_build/build/llama.cpp.
- Read note at top of build-wasm.sh. (TL;DR: need emscripten sdk in env).
- Run build-wasm.sh.
- Copy these files in wasm_build/build to example/web:
- fllama_wasm.worker.mjs
- fllama_wasm.js
- fllama_wasm.wasm
- Test using instructions below, then commit and push.
- Note: Drop usually requires restoring build-info.cpp for macOS and iOS. It's a stock set of values with ex. git commit. That should be updated. Other than that you can leave it alone.
- When C++ changes, use ./build-wasm.sh to build updated WASM files.
- Copy: fllama_wasm.js, fllama_wasm.wasm, fllama_wasm.worker.mjs from wasm_build/build to example/web directory.
- From example directory: flutter run -d web-server --web-hostname=localhost --web-port=1234
- Open Chrome and go to https://localhost:1234.
- Enable WASM via modheader extension, set: Cross-Origin-Embedder-Policy: require-corp Cross-Origin-Opener-Policy: same-origin
- When changes are made to C++ bindings, run
flutter pub run ffigen --config ffigen.yaml
to make them available in Dart. rm -rvf Podfile.lock && rm -rvf Podfile && rm -rvf Pods && flutter clean
^ run in example/macos / example/ios when upgrading cpp files, or when getting cryptic errors about build cache.