Working on a local ai workstation for weeks.
Zuckerschnute v2 local (uncensored) ai
![[Linked Image]](https://i.imgur.com/zb4URjo.png)
Hardware:
AMD Ryzen 7 5800X3D
AMD Radeon RX 6950 XT (16 GB VRAM)
32 GB RAM
Ubuntu 24.04 LTS
ROCm 6.4.4 / Vulkan
Water cooled
My local AI ecosystem:
Voice / TTS:
AllTalk running in Docker
GPU-accelerated speech synthesis
Integrated directly into SillyTavern
![[Linked Image]](https://i.imgur.com/HYdhS3R.png)
3D Generation – TRELLIS.2 / ComfyUI:
Running locally on AMD RX 6950 XT using ROCm
Modified CuMesh integration for AMD/ROCm compatibility
Adapted the pipeline to work around CUDA-specific dependencies
GPU-accelerated image-to-3D mesh generation
Successfully generating textured 3D models locally
Optimized generation settings to stay within 16 GB VRAM
![[Linked Image]](https://i.imgur.com/GKnDLGN.png)
Video Generation – LTX-Video 2 (Pic2Vid) / 2nd ComfyUI:
Local AI-powered image-to-video generation
Running on AMD Radeon RX 6950 XT (16 GB VRAM)
Integrated into my local AI workstation
No cloud-based video generation required
![[Linked Image]](https://i.imgur.com/0HYURUi.png)
FaceFusion 3.9.2:
Dockerized ROCm environment
Replaced ONNX Runtime ROCm 1.22.2.post3 with 1.21.0
Fixed missing ROCm dependencies (RCCL, ROCm SMI libraries)
Configured LD_LIBRARY_PATH for GPU provider loading
Successfully enabled ROCMExecutionProvider
Approximately 70 FPS during video face-swapping (900-frame test. feels very fast in action)
Modified Gradio cache management to prevent massive temporary file accumulation
![[Linked Image]](https://i.imgur.com/hwRMyPj.png)
Planed: text2pic ( should be easy to get into pic2vid )
Custom Control Software (WIP):
I also developed my own workstation launcher using Lazarus / Free Pascal.
It handles application startup, Docker containers, GPU VRAM monitoring, RAM monitoring, logs, file transfers and temporary file cleanup.
![[Linked Image]](https://i.imgur.com/rRdbf4H.png)
LLM & Chat:
SillyTavern frontend
KoboldCpp (Vulkan)
Violet Lotus 12B (Mistral Nemo fine-tune)
~12K context window
Custom character expressions and automation
Local vector storage for conversation continuity
[img]
https://i.imgur.com/0DtsVZV.png[/img]
[img]
https://i.imgur.com/AxpJ8nm.png[/img]
<3

...skynet is very sweet.

The result:
A completely local AI environment for chatting, speech synthesis, image-to-3D generation and GPU-accelerated video processing.
No cloud APIs, no subscriptions, no external inference servers.
Everything runs on my own hardware.
The RX 6950 XT might not be the easiest GPU for AI workloads, but with enough patience, ROCm troubleshooting and a few creative workarounds, it can do some pretty impressive things.
Ordered 2tb hd now to make a system image...i will never do that again, it was HELL to get this setup running! Hell!
edit: done...ahhhh...
Trellis2 will help me a lot with my games.i guess

And yes, getting all of this working on AMD/Linux was quite an adventure.
Update Trellis2 +15.000 Tokens on rx 6950 xt amd rocm: The problem:
Complex models generated 15,000–21,000+ geometry tokens.
High-resolution SLat sampling consistently crashed with CUDA OOM.
FlashAttention was unavailable, so Trellis2 fell back to PyTorch SDPA.
Attention memory requirements increased dramatically with token count.
The fix:
Modified full_attn.py, specifically _sdpa_varlen().
Implemented query-chunked attention with 1,024 tokens per chunk.
Each query chunk still attends to all keys and values.
No additional libraries, ROCm changes, or PyTorch modifications required.
Result:
Successfully completed HR geometry sampling with 14,993 tokens.
Texture sampling and mesh generation also completed.
Total processing time: 22 minutes 27 seconds.
Previously impossible models now generate successfully!
Greets