Add Dockerfile.blackwell for NVIDIA Blackwell (sm_120) GPUs

CUDA 12.8-based image variant that can build for and run on sm_120:
- nvidia/cuda:12.8.1 base, PyTorch 2.8.0+cu128 (ships sm_120 kernels)
- CUDA_ARCHITECTURES defaults to 120; arch-list awk handles 3-digit archs
- COLMAP 3.12.6 / GLOMAP 1.2.0 (CUDA 12.8-compatible releases)
- tiny-cuda-nn pinned to a Blackwell-capable master commit
- tcnn/gsplat built with --no-build-isolation
- nerfacc JIT kernels pre-compiled at build time (runtime stage has no nvcc)
- apt sandbox workaround for seccomp-related GPG verification failures
- docs: mention the variant in the docker install section

Validated on RTX 5060: image builds clean; splatfacto and nerfacto both
train to completion inside the container.
This commit is contained in:
yesMohanHere committed 2026-07-18 18:55:55 +05:30
1 parent 4c8c27c11a
commit 88ee118755
2 files changed
+189

No files matched your search

+8
View File
@@ -267,6 +267,14 @@ docker build \
--file Dockerfile .
```
For NVIDIA Blackwell GPUs (RTX 50-series, compute capability sm_120), use the CUDA 12.8-based
`Dockerfile.blackwell` variant instead — the stock image targets CUDA 11.8, which does not
support sm_120:
```bash
docker build --tag nerfstudio-blackwell -f Dockerfile.blackwell .
```
### Using an interactive container
The docker container can be launched with an interactive terminal where nerfstudio commands can be entered as usual. Some parameters are required and some are strongly recommended for usage as following: