Moshi rewrite

.github/workflows/cd.yml

@@ -23,5 +23,4 @@ jobs:
       - name: Deploy
         run: |
           modal token set --token-id $MODAL_TOKEN_ID --token-secret $MODAL_TOKEN_SECRET
-          modal run src.fillers::Fillers.prefill
           modal deploy

README.md

@@ -1,22 +1,23 @@
 # QuiLLMan: Voice Chat with LLMs
 
-A complete chat app that transcribes audio in real-time, streams back a response from a language model, and synthesizes this response as natural-sounding speech.
+A complete voice chat app powered by a speech-to-speech language model and bidirectional streaming.
 
-[This repo](https://github.com/modal-labs/quillman) is meant to serve as a starting point for your own language model-based apps, as well as a playground for experimentation. Contributions are welcome and encouraged!
+On the backend is Kyutai Lab's [Moshi](https://github.com/kyutai-labs/moshi) model, which will continuously listen, plan, and respond to a the user. It uses the [Mimi](https://huggingface.co/kyutai/mimi) streaming encoder/decoder model to maintain an unbroken stream of audio in and out, and a [speech-text foundation model](https://huggingface.co/kyutai/moshiko-pytorch-bf16) to determine when and how to respond.
 
-OpenAI [Whisper V3](https://huggingface.co/openai/whisper-large-v3) is used to produce a transcript, which is then passed into the [LLaMA 3.1 8B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct) language model to generate a response, which is then synthesized by Coqui's [XTTS](https://github.com/coqui-ai/TTS) text-to-speech model. All together, this produces a voice-to-voice chat experience.
+Thanks to bidirectional websocket streaming and use of the [Opus audio codec](https://opus-codec.org/) for compressing audio across the network, response times on good internet can be nearly instantaneous, closely matching the cadence of human speech.
 
 You can find the demo live [here](https://modal-labs--quillman-web.modal.run/).
 
+![Quillman](https://github.com/user-attachments/assets/afda5874-8509-4f56-9f25-d734b8f1c40a)
+
+This repo is meant to serve as a starting point for your own language model-based apps, as well as a playground for experimentation. Contributions are welcome and encouraged!
+
 [Note: this code is provided for illustration only; please remember to check the license before using any model for commercial purposes.]
 
 ## File structure
 
-1. React frontend ([`src/frontend/`](./src/frontend/))
-2. FastAPI server ([`src/app.py`](./src/app.py))
-3. Whisper transcription module ([`src/whisper.py`](./src/whisper.py))
-4. XTTS text-to-speech module ([`src/xtts.py`](./src/xtts.py))
-5. LLaMA 3.1 text generation module ([`src/llama.py`](./src/llama.py))
+1. React frontend ([`src/frontend/`](./src/frontend/)), served by [`src/app.py`](./src/app.py)
+2. Moshi websocket server ([`src/moshi.py`](./src/moshi.py))
 
 ## Developing locally
 
@@ -26,31 +27,53 @@ You can find the demo live [here](https://modal-labs--quillman-web.modal.run/).
 - A [Modal](http://modal.com/) account (`modal setup`)
 - A Modal token set up in your environment (`modal token new`)
 
-### Developing the inference modules
+### Developing the inference module
 
-Whisper, XTTS, and Llama each have a [`local_entrypoint`](https://modal.com/docs/reference/modal.App#local_entrypoint)
-that is invoked when you execute the module with `modal run`.
-This is useful for testing each module standalone, without needing to run the whole app.
+The Moshi server is a [Modal class](https://modal.com/docs/reference/modal.Cls#modalcls) module to load the models and maintain streaming state, with a [FastAPI](https://fastapi.tiangolo.com/) http server to expose a websocket interface over the internet.
 
-For example, to test the Whisper transcription module, run:
+To run a [development server]((https://modal.com/docs/guide/webhooks#developing-with-modal-serve)) for the Moshi module, run this command from the root of the repo.
 
 ```shell
-modal run -q src.whisper
+modal serve src.moshi
 ```
 
-### Developing the http server and frontend
+In the terminal output, you'll find a URL for creating a websocket connection.
+
+While the `modal serve` process is running, changes to any of the project files will be automatically applied. `Ctrl+C` will stop the app. 
+
+### Testing the websocket connection
+From a seperate terminal, we can test the websocket connection directly from the command line with the `tests/moshi_client.py` client.
+
+It requires non-standard dependencies, which can be installed with:
+```shell
+python -m venv venv
+source venv/bin/activate
+pip install -r requirements/requirements-dev.txt
+```
 
-The HTTP server at `src/app.py` is a [FastAPI](https://fastapi.tiangolo.com/) app that chains the inference modules into a single pipeline.
+With dependencies installed, run the terminal client with:
+```shell
+python tests/moshi_client.py
+```
+
+And begin speaking! Be sure to have your microphone and speakers enabled.
+
+### Developing the http server and frontend
 
-It also serves the frontend as static files.
+The http server at `src/app.py` is a second [FastAPI](https://fastapi.tiangolo.com/) app, for serving the frontend as static files.
 
-To run a [development server](https://modal.com/docs/guide/webhooks#developing-with-modal-serve), execute this command from the root directory of this repo:
+A [development server]((https://modal.com/docs/guide/webhooks#developing-with-modal-serve)) can be run with:
 
 ```shell
 modal serve src.app
 ```
 
-In the terminal output, you'll find a URL that you can visit to use your app. While the `modal serve` process is running, changes to any of the project files will be automatically applied. `Ctrl+C` will stop the app.
+Since `src/app.py` imports the `src/moshi.py` module, this also starts the Moshi websocket server.
+
+In the terminal output, you'll find a URL that you can visit to use your app. 
+While the `modal serve` process is running, changes to any of the project files will be automatically applied. `Ctrl+C` will stop the app. 
+
+Note that for frontend changes, the browser cache may need to be cleared.
 
 ### Deploying to Modal
 
@@ -60,4 +83,6 @@ Once you're happy with your changes, [deploy](https://modal.com/docs/guide/manag
 modal deploy src.app
 ```
 
+This will deploy both the frontend server and the Moshi websocket server.
+
 Note that leaving the app deployed on Modal doesn't cost you anything! Modal apps are serverless and scale to 0 when not in use.