Hf Cloud Python Env Setup

Python Environment Setup for SageMaker

Most SageMaker deployment failures that look like AWS problems are actually Python environment problems: wrong Python version, broken dependency resolution, stale SDK that doesn’t know about a current API. This skill makes env setup boring and correct.

Core rules

1. Never use the system Python. Always work inside an isolated environment.
2. Pin the Python version, not the package versions. Use 3.10, 3.11, or 3.12. Avoid 3.13+ — ML libraries lag on wheel availability and dependency resolution breaks in confusing ways.
3. Install the latest of each package. Don’t defensively pin boto3 or awscli. Newer ones have current API surfaces and security fixes. Only pin if the user explicitly requires a specific version.
4. Check installed versions correctly. Use importlib.metadata.version("package-name"), never module.__version__. The latter is inconsistent across packages.
5. The bundled scripts use boto3 directly. The SageMaker Python SDK is a valid alternative — see “boto3 vs the SageMaker SDK” below.

boto3 vs the SageMaker SDK

The bundled deploy scripts (deploy.py, deploy_async.py, teardown.py) use boto3 directly and read image URIs from AWS’s published Deep Learning Containers catalog (opens in a new tab). That fits this workflow’s explicit-stages design — each skill produces a concrete value (region, role ARN, image URI) that the next one consumes — and boto3 is the stable underlying API client.

The SageMaker Python SDK (v3) is fine to use when the user prefers it or their project already does. Since PR #5960 (opens in a new tab) (June 2026), ModelBuilder auto-routes HuggingFace models to the current containers (text-generation → HuggingFace vLLM, multimodal → vLLM-Omni, embeddings → TEI). Don’t avoid the SDK over stale-image or wrong-container concerns — that routing is fixed.

Two specific SDK cases that still need care:

• Generative rerankers: the SDK routes the text-ranking task to TEI unconditionally, which is wrong for causal-LM rerankers like Qwen3-Reranker — those need vLLM (see hf-cloud-serving-image-selection). Pass the container explicitly for these models.
• SSO assumed-role credentials: v3 has had credential-resolution regressions in ModelTrainer / FrameworkProcessor under SSO profiles. If SDK calls fail with credential errors while aws sts get-caller-identity succeeds in the same shell, suspect this rather than your AWS config.

If you use the SDK, install it into the isolated env like everything else (.venv/bin/python -m pip install sagemaker). The bundled scripts don’t require it.

How to set up

The fastest path is the bundled script — it’s Python, so it runs the same on Windows, macOS, and Linux:

python3 scripts/setup_env.py        # macOS / Linux
python  scripts/setup_env.py        # Windows (PowerShell / cmd)

This script detects uv and uses it if available (faster), falls back to the stdlib venv module, creates .venv/ with Python 3.12 (override: python3 setup_env.py .venv 3.11), refuses unsupported Python versions, installs from the bundled requirements.txt, and is idempotent. It also prints the correct interpreter path for the host OS (see below).

Manual equivalent:

# Preferred: uv
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python --upgrade boto3 awscli   # Windows: .venv\Scripts\python.exe

# Fallback: stdlib venv
python3.12 -m venv .venv
.venv/bin/python -m pip install --upgrade pip boto3 awscli

After setup, invoke the env’s Python explicitly rather than activating the venv. The interpreter path differs by platform:

.venv/bin/python deploy.py            # macOS / Linux
.venv\Scripts\python.exe deploy.py    # Windows

This works the same in scripts, interactive shells, and agent tool calls. The rest of this skill writes .venv/bin/python for brevity — on Windows substitute .venv\Scripts\python.exe.

Verifying

.venv/bin/python scripts/check_versions.py

Prints versions of boto3, botocore, awscli. Uses importlib.metadata.version() so it works on every package, including ones without __version__. Pass arbitrary names: ... check_versions.py transformers huggingface_hub.

Deployment-specific extras

Default requirements.txt covers SageMaker orchestration. Some deployments need extras (huggingface_hub for model inspection, transformers for tokenizer validation). Add these to a deployment-specific requirements file in the project, install with the env’s Python, don’t pin unless there’s a reason.

Common pitfalls

Mysterious pip install resolution errors Almost always Python 3.13+ trying to install packages without wheels yet, or installing into a polluted system Python. Recreate at 3.12: delete .venv and re-run python3 setup_env.py .venv 3.12 (the script recreates the env when the version doesn’t match, so you can also just re-run it).

pip install succeeded but the script says “module not found” You installed into a different interpreter than the one running the script. Always invoke Python explicitly: .venv/bin/python -m pip install ... and .venv/bin/python deploy.py.

Inline python -c "..." one-liners fail in PowerShell PowerShell’s quoting rules mangle nested/escaped quotes in inline Python. Don’t debug the quoting — write the snippet to a small .py file and run that. (All bundled helpers are files for exactly this reason.)

boto3 call fails with “unknown parameter” Your boto3 is older than the API surface. Upgrade with .venv/bin/python -m pip install --upgrade boto3. Don’t downgrade the script to match an old version.

sagemaker (the SDK) installed but the bundled scripts fail The bundled scripts don’t use the SDK — they only need boto3/awscli from requirements.txt. Installing sagemaker alongside is harmless, but it doesn’t replace the requirements install.