Remote Debugging Workflows for nanoChat

Remote Debugging Workflows for nanoChat | Generated by AI

Home 2026.06

Question: Can you debug nanoGPT/nanoChat on your home workstation (Ubuntu/LAN) remotely, or do you need to physically sit at the machine? How do PyCharm, VSCode, and Zed compare for this workflow?

Answer:

You can absolutely remote debug. The misconception is that PyCharm remote debugging isn’t available — PyCharm Professional supports remote debugging via SSH with full breakpoint/inspection capability. But given your terminal-first philosophy, you probably want something faster anyway.

Three practical workflows

1. CLI-first: `debugpy` over SSH (my recommendation for you)

This is terminal-native, fast, and works anywhere. Install debugpy on your workstation, start your training script with it, and attach from your Mac via VSCode’s debugger (lightweight, no GUI overhead).

# On workstation (192.168.1.36)
pip install debugpy

# Launch script with debugger listening
python -m debugpy.adapter --listen 0.0.0.0:5678 your_training_script.py
# or inline:
python -c "
import debugpy; debugpy.listen(('0.0.0.0', 5678))
debugpy.wait_for_client()
exec(open('train.py').read())
"

Then in VSCode on your Mac, create .vscode/launch.json:

{
  "version": "0.2.0",
  "configurations": [
    {
      "name": "Remote Attach (Workstation)",
      "type": "python",
      "request": "attach",
      "connect": {
        "host": "192.168.1.36",
        "port": 5678
      },
      "pathMapping": {
        "/home/lzw/nanoGPT": "/Users/zhiwei/nanoGPT"  // local path mapping
      }
    }
  ]
}

Hit F5, breakpoint, inspect — all from your Mac. No remote IDE overhead.

2. VSCode/Zed remote SSH development (cleaner UX, slower startup)

Both can mount your remote filesystem and run the debugger directly on the workstation:

VSCode:

# Install Remote - SSH extension
# In VSCode: Cmd+Shift+P → "Remote-SSH: Connect to Host"
# Enter: ssh lzw@192.168.1.36

# Then open folder, terminal is directly on workstation
# Debugger runs there natively

Zed:

# Newer Zed has SSH support (version 0.140+)
# File → Open Remote → SSH
# zed --new-window /home/lzw/nanoGPT

Pro: IDE feels native. Con: Network latency if your WiFi is weak, and Zed’s remote is still evolving.

3. PyCharm Professional remote debugging (overkill for your workflow)

Yes, it works:

Settings → Project → Python Interpreter → Add → SSH Interpreter
Configure SSH to 192.168.1.36
Set breakpoints, debug directly

But it’s GUI-heavy and slower than debugpy + VSCode for quick iterations.

My honest take for your setup

Since you’re terminal-first and doing GPU training with nanoGPT, use debugpy + VSCode or VSCode SSH remote. Why:

debugpy keeps you in code/terminal, fast attach/detach, minimal overhead. Perfect for long-running training jobs where you inspect state occasionally.
VSCode SSH remote if you’re doing active development (editing config, modifying loss calculations) — seamless file sync, integrated terminal.
Zed remote if you want to live there long-term (Zed is snappier than VSCode for large codebases, native).

For actual debugging during training (inspecting tensors, gradients, activations), you probably want pdb/pdbpp in the terminal anyway:

# In your training loop
import pdb

if step % 100 == 0:
    print(f"Loss: {loss}, hidden shape: {h.shape}")
    pdb.set_trace()  # Drop into REPL, inspect h, grads, etc.

Over SSH with Warp:

ssh lzw@192.168.1.36
cd ~/nanoGPT && python train.py
# Hits breakpoint, you're in pdb shell
# Inspect variables, evaluate code
# `c` to continue

No PyCharm needed. No remote IDE. Just SSH + pdb + Warp.

For AMD MI300X cloud workloads: Same approach. debugpy listening on the cloud instance, attach from your Mac. Or just log aggressively (print → wandb/tensorboard) since cloud boxes are stateless and you’re usually iterating batch configs, not stepping through code.

Back Donate