Spark Environment Setup

DGX Spark ships a GB10 Grace Blackwell chip: aarch64 CPU, SM121 GPU, 128GB unified memory, CUDA 13. This is a narrower and younger platform than a standard x86 CUDA 12 box, so package selection and ABI matching matter more than usual — the wheel ecosystem for aarch64 + CUDA 13 is still filling in.

When to Use This Skill

• Setting up a fresh Spark box for training or inference.
• Hitting an import error mentioning libcudart, a missing symbol, or a wheel that “installed fine but won’t load.”
• A framework install (PyTorch, Unsloth, TRL, vLLM, xformers) fails, hangs, or silently falls back to CPU.
• Deciding whether to use an NGC container or bare pip.
• Restoring a working setup after an OS reinstall or a base-image update, needing to re-verify from scratch.

Each of these accepts the same general fix: match the container/wheel combination to CUDA 13 and SM121, don’t fight the ABI.

Container-First Rule

Quick decision, before the detail below:

• Standard training/inference work → NGC PyTorch container.
• Unsloth-centric fine-tuning → Unsloth container (it ships the pinned Triton/xformers/transformers combination already validated for that path).
• Neither fits (custom system package, local IDE interpreter) → bare pip, following the exact sequence further down.

Default to a container. Use nvcr.io/nvidia/pytorch:25.09-py3 as the base for general work — the newest tag confirmed working on this hardware; pull a newer blessed tag if locally available rather than hard-blocking on 25.11-py3. NGC’s tag is dated, so running it directly is fine:

docker run --runtime=nvidia --gpus all -it --rm \
  nvcr.io/nvidia/pytorch:25.09-py3

unsloth/unsloth:dgxspark-latest is a moving tag by contrast — resolve and pin its digest before running it for anything reproducible; the bare tag is a discovery step only, not the default invocation. Full pull-inspect-pin sequence and flag rationale/volume mounts for finetuning/ run dirs: references/container-workflow.md. Treat bare pip as the exception.

The reason for the container-first stance is pinning, not convenience. Triton, xformers, and transformers versions interact narrowly with GB10’s SM121 target and CUDA 13; a container locks all of them together against a combination already validated on this hardware. Bare pip leaves that resolution to you, one broken import at a time.

When bare pip is warranted, follow the NVIDIA playbook’s install sequence verbatim and in order:

pip install "transformers==5.13.1" "peft==0.19.1" "hf_transfer==0.1.9" "datasets==4.3.0" "trl==1.8.0"
pip install --no-deps "unsloth==2026.7.2" "unsloth_zoo==2026.7.2" "bitsandbytes==0.49.2"
pip install -U "torchao==0.17.0"

The second command’s --no-deps flag is not optional — letting pip re-resolve Unsloth’s dependency tree on aarch64 is a common way to pull in an incompatible torch or triton build. The third line is not optional either: the NGC base image’s bundled torchao is too old for current peft‘s LoRA-attach path (ImportError: ... torchao ... only versions above 0.16.0 are supported) — a hard blocker, not a warning. Every == pin above is load-bearing, taken from the dated known-good version matrix in references/stack-matrix.md (its Last verified date governs staleness) — an unpinned install resolves current PyPI versions well outside what this Unsloth release supports.

Pull a fresh tag when a new blessed release is announced. Rebuild locally from one of the two bases only when a project needs an extra system package layered in — not to “upgrade” a component the image already pins. Details on both paths: references/container-workflow.md.

One more preflight: official DGX Spark playbooks have shipped broken before. Check recent issues on github.com/NVIDIA/dgx-spark-playbooks (and the other resources in references/stack-matrix.md) before trusting a recipe verbatim for a long run.

The ABI Rule

The single most common failure on Spark is a CUDA 12/13 ABI mismatch: a wheel built against libcudart.so.12 loaded on a system that only has libcudart.so.13. The install usually succeeds; the failure surfaces later as a missing-symbol error or a segfault that doesn’t obviously point at CUDA.

Fix: pull wheels from download.pytorch.org/whl/cu130 (the cu130-tagged aarch64 builds), or use one of the containers above, which already carry a matched build. Before chasing a stack trace that mentions a CUDA symbol, check which CUDA tag the installed wheel was built against:

python3 -c "import torch; print(torch.version.cuda)"

If that output doesn’t start with 13, the ABI mismatch is the first thing to fix. NGC container builds (e.g. nvcr.io/nvidia/pytorch:25.09-py3) build torch internally against CUDA 13 with no +cu130 wheel tag — pip show torch won’t say cu130 there, and that absence alone is not a failure.

Typical symptoms:

• ImportError: undefined symbol referencing a CUDA runtime function.
• A segfault on the first .cuda() call, no useful traceback.
• A wheel that installs cleanly, then fails at import time — pip’s resolver doesn’t check CUDA ABI, only version constraints.
• Two “identical” environments behaving differently — usually one has a cu130 wheel, the other a cu121/cu124 leftover.

The fix is the same regardless of symptom: match the wheel’s CUDA tag to the system, or use a container that already does.

Component Quick Table

Condensed status for the components most likely to come up. Full table with wheel URLs, build flags, the sm_121 vs sm_121a distinction, and the dated known-good version matrix: references/stack-matrix.md.

Everything else — Unsloth, Axolotl, TRL, PEFT — installs cleanly through the container-first path above. LLaMA-Factory and NeMo are fragile on Spark; check upstream issues first.

Verification Commands

Confirm the environment can actually see the GPU before running anything expensive:

import torch
print(torch.cuda.is_available(), torch.version.cuda)

This call returns two values; the exact output format is one line, <bool> <cuda-version>:

True 13.0

If it prints False instead, don’t jump straight to a wheel reinstall — ABI mismatch is one cause among several:

Check nvidia-smi first — if it doesn’t show the GPU, it’s one of the first three, not ABI. Reinstall a wheel only once ABI is confirmed. Per-hypothesis detail: references/stack-matrix.md. Run right after the container starts, before installing project-specific packages.

One more check: if Triton kernel compilation fails once training starts, set TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas and retry — see references/stack-matrix.md for the full workaround list.

Next Steps

A verified environment is only the starting point. See also: spark-training-gotchas for failure preflights before a training run, and spark-memory-thermal-ops for unified-memory OOMs and thermal throttling during long ones.