Dose Cosmos Curate run on Nvidia Spark(arm64 platform)?

[ZhangYong19800721]

I try to run cosmos curate on Nvidia Spark(arm 64 platform). but I encounted many problems, cli infomation show that it dosen't support linux-aarch64. So, i want to make sure that if is no possible to run cosmos curate on spart(arm64).
ps: I can successfully run cosmos curate on x86 amd64 computer with nvidia GPU 5070 (only the hello world example).

[BulletsHo]

not with the ready-made container, but yes if you build it from source — this has been the case since a few months after you posted.

I dug into the actual repo (not just docs) to check this properly, and there's a specific reason you hit that error.

Your timing is the actual explanation

Your post is dated Mar 2, 2026. Two days later, v1.2.0 shipped with this changelog line:

ARM64 CLI and container build support

(https://github.com/NVIDIA/cosmos-curator/blob/main/CHANGELOG.md#L505 — same release also added GB200 support). Before that, linux-aarch64 wasn't a target platform anywhere in the build tooling, so the CLI was correct at the time: it genuinely didn't support aarch64 yet.

Two different things are both called "Cosmos Curator" — check which one you're using

• NVIDIA's prebuilt image (nvcr.io/nvidia/cosmos/cosmos-curator on NGC) is still x86_64 only. The NGC listing says CPU: x86_64 host CPU and shows "Multi-Arch Support: No" on the latest tag (2.3.0, updated Aug 6, 2026): https://catalog.ngc.nvidia.com/orgs/nvidia/cosmos/containers/cosmos-curator/2.3.0. If you're pulling that image, this is why — it's simply not built for arm64, not a bug.
• The GitHub source (NVIDIA/cosmos-curator, mirrored at nvidia-cosmos/cosmos-curate — identical commits) is currently at v2.4.0 (2026-07-30), and pixi.toml now genuinely targets aarch64: the workspace declares linux-aarch64 plus a dedicated linux-aarch64-cuda platform (CUDA 13.0.2), and feature.runtime has an explicit aarch64 target block adding the CUDA/Triton wheels needed on ARM: https://github.com/NVIDIA/cosmos-curator/blob/main/pixi.toml#L147-L211. This covers core, runtime, cuml, legacy-transformers, seedvr, and sam3 — the environments the default image build actually uses.

So: build the image natively on the Spark instead of pulling NVIDIA's image, and it should pick up the aarch64 targets.

One known gap

The paddle-ocr feature (artificial-text/overlay detection) is still pinned to linux-64 only: https://github.com/NVIDIA/cosmos-curator/blob/main/pixi.toml#L281-L285. If your pipeline uses that stage it won't build on arm64 yet — everything else in the default build set should.

What I'd try

git clone --recurse-submodules https://github.com/NVIDIA/cosmos-curator.git
cd cosmos-curator

# build natively on the Spark itself — do NOT pull the NGC image, it's x86_64-only
pixi run -e tools cosmos-curator image build \
  --image-name cosmos-curator --image-tag local

# launch against your own locally-built image, not nvcr.io's
pixi run -e tools cosmos-curator local launch \
  --image-name cosmos-curator --image-tag local \
  -- pixi run -e default hello-world

If this still fails, the most useful thing you can share is the exact pixi/docker-build error rather than "unsupported platform" — that tells us whether it's actually solving for aarch64 now. My best guess for a remaining friction point is vllm==0.24.0/vllm-omni (a PyPI-only dependency under runtime): Blackwell + aarch64 (sm_121) wheel availability has been a rough edge across the whole CUDA ecosystem this year, not just here (e.g. vllm-project/vllm#36821).

Wider context

DGX Spark (GB10, ARM Grace-Blackwell) support has landed unevenly across the whole Cosmos family, not just Curator. NVIDIA's own docs currently list Cosmos-Reason2 as validated on DGX Spark for inference (https://docs.nvidia.com/cosmos/latest/prerequisites.html), while Cosmos-Predict2.5 users were still hitting hard aarch64 blockers (Transformer Engine / flash-attn not built for ARM) as recently as Feb 2026 (nvidia-cosmos/cosmos-predict2.5#120). Curator quietly joined the "supported at the source level" side in March — the published container just hasn't caught up.

TL;DR: the CLI was right when you tried it (arm64 landed 2 days later, in v1.2.0). It's buildable from source for aarch64 as of v1.2.0+ (latest is v2.4.0) — build locally on the Spark instead of pulling NGC's x86_64-only image. One stage (paddle-ocr) is still x86_64-only. If it still breaks, file a focused issue with the exact error on the actual project repo (https://github.com/NVIDIA/cosmos-curator/issues) — this discussion lives in the general NVIDIA/cosmos mono-repo, which the Curator maintainers may not watch as closely.