How to install Odysseus with Docker Desktop on Windows and Workaround issues with GPU Visibility by USING lmstudio(windows) and patching the TTS implementation with a cpu-only "torch" version of tts_service.py.

[ScuffedMallow]

If you cannot run an LLM on the container after installing, its most likely because your gpu was not visible to the container.
You will run into issues with server-side TTS Mode (local kokoro).

This is how you can workaround it:

First we need to have an llm service that can provide experimental features that are easy to configure and elevate the value of normal ram configurations.
I chose and recommend to install lmstudio on Windows

Using the Graphical User Interface:

1 Open LM Studio.


2 Configure Settings "Developper Mode" to "On"


3Make sure System /Runtime /Engines and Frameworks has the CPU llama.cpp ROCM(amd) or CUDA(NVIDIA) versions depending on your system.


4 Download good model that fit inside your gpu like gemma or qwen.


5 Go to "My Models" [ CTRL + 3 ].


6 Select the desired llm model.


7 Experimental Settings
	Max Concurrent Prediction = 1
	Keep Model in Memory = OFF
	Flash Attention = ON
	K Cache Quantization = Q4_0
	V Cache Quantization = Q4_0
These settings found in the "Load" tab will allow for more context available for the agent.


8 Navigate to the Developer tab.


9 Toggle the "Start server" switch to activate service

10 (Optional) Toggle the "Require Authentication" setting in the server settings drop down menu of the developer tab and generate an API key from the "Manage Tokens" menu

11 Save the API token and the ip address displayed on the top right after "Reachable at:" (if on the same system as docker just use "http://127.0.0.1/1234")

Second order of business is to follow the instructions at
[https://github.com/ScuffedMallow/CPU-only-Torch-Kokoro-Fix-for-Odysseus-TTS-Mode-on-Docker]
to build an image of Odysseus on Docker Desktop Windows with TTS enabled, deploy the containers and configure the agent for CPU only TTS.
The steps for how to patch the TTS issue are in the READ ME.
You have all you need to get this working without friction if you follow the steps.
Do not forget to restart the container every time it is required by the instructions for the patch.

Im not a very technical person i just caught a case of tism because my install could not see the gpu.

I did not use any LLMs to fix it.

I did it by reading source code, logs and docs to trace the error path down to how the tts_service stops (if no cuda compatible device is found) and by searching the docs for kokoro cpu only configurations (where i found the information to patch the implementation of torch and kokoro) Then i made the necessary changes to call the kpipeline with the right settings and context in the file tts_service.py.

If the "tts_service" calls kokoro with access to a torch version with CUDA, it will always try to use CUDA even if no device is found, then stop because no CUDA device was found.

Enjoy the patch and tips! Now its the best agent i have ever had. Easy to use, Configurable, now with TTS, Deep Research, THANKS PEWD!

[ryslan25500-cloud]

#3255 The yoke is forcibly, turned off, and you want to turn it on, you idiot? Have you ever wondered about that?