Interdisciplinary · Reference

New ML Desktop — Setup Summary

Built 30 August 2026 — the machine ordered to unblock SmolVLA training and general heavier local compute

Current state: machine built and fully set up — Ubuntu 26.04, GPU driver/CUDA confirmed working, dev tools installed, Wake-on-LAN to the home server tested and working. First real workload (XTTS fine-tuning) surfaced and resolved a PyTorch/Blackwell compatibility gap on the RTX 5060 Ti.

Waiting on: the new camera before lerobot is cloned here and SmolVLA training begins — everything else on this machine is ready to go.

1. Hardware

Component	Details
Case	CSL Sprint 5753 (Ventrix, aRGB)
CPU	AMD Ryzen 9 9900X, 12-core @ 4.4GHz (integrated Radeon "Granite Ridge" graphics also present)
Cooler	be quiet! Pure Rock 3 Black
GPU	GeForce RTX 5060 Ti, 16GB GDDR7 (MSI Shadow 2X OC)
RAM	32GB DDR5-6000 (2×16GB, Patriot Viper Venom)
Motherboard	ASUS Prime B840M-A WIFI (socket AM5)
Boot drive	1TB Western Digital Black SN7100 NVMe
PSU	BoostBoxx 700W, 80+ Gold
Extra storage	2× 1TB Seagate Barracuda ST1000DM003 (from spare stock, previously an Intel RAID pair — metadata wiped, reformatted individually)

2. Software

OS: Ubuntu 26.04 LTS Desktop
GPU driver: Nvidia proprietary driver 595.84 (auto-installed by the Ubuntu installer), CUDA 13.2 confirmed working via nvidia-smi
Network: WiFi working (built-in, antenna already seated), ethernet also available as backup. Reaches the home server at 192.168.2.42 (ping ~6–16ms)
Dev tools installed: Python 3.14.4, pip, venv, git, VS Code (via snap), build-essential
Environment strategy: plain venv per project (no conda) — deliberate choice given past mix-up issues with nested venvs
GitHub access: SSH key generated and added, ssh -T git@github.com confirmed working

3. Storage

Mount point	Drive	Size	Filesystem
`/` (system)	NVMe	~916GB usable	ext4
`/mnt/storage1`	Seagate HDD #1	916GB (870GB free)	ext4
`/mnt/storage2`	Seagate HDD #2	916GB (870GB free)	ext4

Both HDDs SMART-checked: PASSED, 0 reallocated sectors, 0 pending sectors, clean error logs. ~12,095 power-on hours each (~500 days). Extended SMART self-test completed without error.

4. Wake-on-LAN client

Configured: 30 August 2026

This machine can wake the home server (192.168.2.42) from cold poweroff via a magic packet.

# Package installed
sudo apt install wakeonlan

# Alias added to ~/.bashrc
alias wakeserver='wakeonlan 2c:56:dc:74:c4:ff'

Usage:

wakeserver
# Server is reachable via SSH ~60-90 seconds later
ssh r@192.168.2.42

Tested and confirmed: cold poweroff → magic packet → SSH connected ✅

5. Role in the overall system

Pi4 (Brain) + Pi5           Robot arm — lerobot, SmolVLA (Phase 4/5)
      ↕ network
M4 Mac                      Simple tasks, intermediate node
      ↕ network
Home Server (192.168.2.42)  Always-on storage + PostgreSQL (planned)
      ↕ network
This machine (ML Desktop)   Heavy compute:
                              - SmolVLA training (robot arm Phase 6 ext.)
                              - my_db_project Stage 3 ML
                              - XTTS / Piper fine-tuning experiments

Docker: not installed, not needed. This machine points at the server's PostgreSQL directly. One canonical DB, no local copy diverging.

6. GitHub Repository Inventory (as of this session)

Full account: 13 repos (7 public, 6 private)

Active — cloned to `~/Documents/github/active/`

Repo	Visibility	Notes
`iliya-radulov.github.io`	Public	Portfolio site
`arm`	Private	Full documentation
`UselessTools`	Public	Small utility projects
`my_db_project`	Public	Materials science DB — Stage 2 near complete
`so101-pickplace-project`	Private	Robot arm — waiting on new camera for SmolVLA
`iliya-radulov`	Public	Profile repo

To archive/remove later (not cloned here)

Repo	Visibility	Notes
`odenwald-monitor`	Private	Live on home server — do not touch (~1 month remaining)
`LinkedInLearning`	Private	—
`library_sql_python_demo`	Public	—
`lianes_lib_project`	Public	—
`azure_podcast`	Public	—
`Manipulating_data_with_SQL-and_Tableau`	—	—
`FCC`	Private	—

External (not your repo)

lerobot (huggingface/lerobot) — clone when SmolVLA training starts (waiting on new camera)

7. Open Items / Follow-ups

☐ Clone lerobot once ready to start SmolVLA training (need new camera first)
☐ git gc --aggressive on my_db_project to shrink .git history (~783 MB)
☐ Decide fate of the 7 archive/remove repos
☐ odenwald-monitor: revisit train-station/DB matching logic once data collection wraps up (~1 month)
☐ Set up SSH key auth to home server (currently password auth)
✅ XTTS fine-tuning experiment on the Dimitar voice dataset — done, first real workload on this machine. GPU compatibility fix: ML Desktop: Blackwell GPU Fix. The actual training run and its result are filed as a curiosity piece in Fragments.

8. First Real Workload — Blackwell GPU Fix

The GPU's first genuine workload (setting up xtts-finetune-webui) surfaced a real PyTorch/Blackwell architecture compatibility gap — nvidia-smi reported the GPU fine, but the pinned torch version predated Blackwell support entirely. Full diagnosis and fix, worth knowing about for any future CUDA project on this machine: ML Desktop: Blackwell GPU Fix (RTX 5060 Ti).