# Setup Read this once. Each lab's README repeats the parts it needs. --- ## What you need | Lab | Python | GPU | Disk | Install size | |---|---|---|---|---| | A | 3.9+ | no | ~500MB | small | | B | 3.9+ | no | ~4GB | torch (~2GB) | | C, D | 3.9+ | 6GB VRAM, or rent | ~8GB | torch + diffusers | | E | 3.9+ | no | ~15GB models | none (Ollama app) | | F | 3.9+ | yes — free Colab is fine | small locally | none locally | Python 3.9 or newer. Check with `python --version`. On some systems the command is `python3`. --- ## The three routes ### 1. Just run it (recommended, and what most readers should do) Unzip the download, open a lab folder, and double-click **`RUN-THIS-Windows.bat`** or **`RUN-THIS-Mac-Linux.command`**. It handles Python, the environment, and the install for you, and explains what to do if Python is missing. Everything below is for people who would rather drive it themselves. ### 2. Local install by hand **Run these one line at a time.** Do not paste the fence lines. **Windows (PowerShell)** ```powershell cd latent-space-labs\lab-a-embeddings py -m venv .venv .\.venv\Scripts\python.exe -m pip install -r requirements.txt .\.venv\Scripts\python.exe walk_the_space.py ``` **macOS / Linux** ```bash cd latent-space-labs/lab-a-embeddings python3 -m venv .venv ./.venv/bin/python -m pip install -r requirements.txt ./.venv/bin/python walk_the_space.py ``` The labs call the venv's interpreter by path instead of using `activate`. This is not stylistic. `activate` is the step that fails silently: on Windows the bash form `source .venv/bin/activate` errors, and if you keep going, `pip` installs into whatever Python is on your PATH — commonly Anaconda's base environment, where it can break unrelated packages. Calling the interpreter directly makes that failure impossible. **Check which environment you are actually in** at any point: ``` python verify_setup.py ``` Run it from a lab folder with that lab's interpreter. It reports the Python in use, whether it is a venv, what is installed, and whether the lab's models resolve. **Use a virtual environment per lab.** They have conflicting-ish dependency weights and you do not want a 2GB torch install for Lab A, which doesn't need it. Labs C and D can share one. ### 3. Google Colab (nothing installed) Labs A, B, and C ship a self-contained `colab.ipynb`. Each notebook carries the lab's code inside it, so there is nothing to clone and nothing to upload besides the notebook itself. This is the right answer if: - your work laptop is locked down and you can't install things - you don't want 8GB of ML libraries on your machine - you tried route 1 and hit a wall Colab gives you a free GPU (T4, ~15GB VRAM), which is enough for every lab here including the fine-tune. ### 4. Rented GPU (Labs C, D, F) You need this only if you want to run diffusion locally and have no GPU. 1. Make an account at RunPod (or any equivalent — nothing here is RunPod-specific). 2. Deploy a **Pod**, not Serverless. 3. Pick an **RTX 4090** or **RTX A5000**. Roughly $0.35–0.69/hr. **Do not rent an H100.** It costs four times as much and you will not perceive any difference on these workloads. 4. Choose a PyTorch template. It comes with torch and CUDA already set up. 5. Connect via the web terminal or JupyterLab. 6. Upload the lab folder (or re-download the zip on the pod), `pip install -r requirements.txt`, run the lab. 7. **Terminate the pod when you're done.** ### The storage trap — read this bit **A stopped pod still bills for storage.** Network volumes start around $0.07/GB/month and accrue whether or not the pod is running. **Terminate, don't stop.** Copy anything you want off the pod first. This is the single most common way people get a surprise GPU bill, and it is entirely avoidable by knowing it exists. --- ## Windows / PowerShell Every lab works on Windows. The commands differ enough to be worth spelling out, because the failure modes are different too. ### Quickstart ```powershell cd latent-space-labs\lab-a-embeddings py -m venv .venv .\.venv\Scripts\Activate.ps1 pip install -r requirements.txt python walk_the_space.py ``` Use `py` rather than `python` to launch. Windows ships a `python.exe` stub that opens the Microsoft Store instead of running anything, and `py` skips it. ### "Running scripts is disabled on this system" The most common Windows wall. PowerShell blocks script execution by default, so `Activate.ps1` refuses to run. Fix it for your user only — no admin needed: ```powershell Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned ``` If your organisation locks execution policy by group policy, you cannot override it. Two ways around: ```powershell .\.venv\Scripts\activate.bat # the .bat works when .ps1 is blocked .\.venv\Scripts\python.exe walk_the_space.py # or skip activation entirely ``` That last line is the reliable escape hatch: call the venv's interpreter directly and never activate anything. ### Line continuation The labs' multi-line commands use `\`, which is bash. In PowerShell the continuation character is a backtick: ```powershell python compare_quants.py --tags llama3.1:8b-instruct-q8_0 ` llama3.1:8b-instruct-q4_K_M ` llama3.1:8b-instruct-q2_K ``` Or just put it on one line, which is what I'd do. ### Paths Forward slashes work fine in Python arguments on Windows — `--csv data/mine.csv` is fine. It's only the shell commands that need backslashes. If a path contains spaces, quote it: ```powershell python cluster_your_data.py --csv "C:\Users\me\My Data\tickets.csv" --column text ``` ### Other Windows-specific gotchas **`ollama serve` says the port is in use** — the Ollama desktop app is already running it in the tray. Skip `ollama serve` entirely; it's already up. Check with: ```powershell curl.exe http://localhost:11434/api/tags ``` Use `curl.exe`, not `curl` — bare `curl` is an alias for `Invoke-WebRequest` in PowerShell and takes different arguments. **Long path errors during pip install** — torch has deeply nested paths. Enable long paths: ```powershell New-ItemProperty -Path "HKLM:\SYSTEM\CurrentControlSet\Control\FileSystem" ` -Name LongPathsEnabled -Value 1 -PropertyType DWORD -Force ``` That one does need admin. If you don't have it, clone the repo somewhere short like `C:\labs` instead of a deep folder under Documents. **Model downloads go to your C: drive and fill it** — Hugging Face caches under `%USERPROFILE%\.cache\huggingface`. Redirect it if C: is small: ```powershell $env:HF_HOME = "D:\hf-cache" ``` Set it in the same session before running the lab, or add it to your user environment variables to make it stick. **CUDA not detected on a machine that has an NVIDIA GPU** — the default pip `torch` is CPU-only on some Windows setups. Reinstall with the CUDA build: ```powershell pip uninstall torch pip install torch --index-url https://download.pytorch.org/whl/cu121 ``` Verify: ```powershell python -c "import torch; print(torch.cuda.is_available())" ``` **Git isn't installed** — download the repo as a ZIP from GitHub and extract it. Nothing in these labs needs git. --- ## Corporate laptop workarounds **Can't install Python.** Use Colab. Labs A, B, C work fully; D can run there too with a runtime change to GPU. **pip fails with SSL errors.** A proxy is intercepting HTTPS. Ask IT for the proxy cert, or just use Colab. Don't burn an afternoon on this — it isn't the lab. **Downloads blocked.** Hugging Face and Ollama both pull from CDNs that some corporate firewalls block. Colab again, or your personal machine. **No admin rights.** `python -m venv` doesn't need admin. `pip install --user` also works. Ollama does need an install, so Lab E is the one that may be genuinely blocked; everything else has a route. --- ## Verifying your setup ```bash python -c "import sys; print(sys.version)" python -c "import torch; print(torch.__version__, torch.cuda.is_available())" ``` For Ollama: ```bash ollama --version curl http://localhost:11434/api/tags ``` If that curl returns JSON, Ollama is running and Lab E will work. --- ## If you get stuck Check `TROUBLESHOOTING.md`, then email me with your OS, Python version, and the full error text. Reported failures are how TROUBLESHOOTING.md grows.