fix(gpu): run on Blackwell, and refuse a GPU this PyTorch build cannot use - #120

Merged

ncoevoet merged 13 commits into

fix/issue-119-blackwell-cuda-image

Sep 1, 2026

Commits on Sep 1, 2026

fix(docker): publish a Blackwell-capable CUDA image and keep a legacy tag
ncoevoet
committed
43aec7b
fix(device): refuse a GPU this PyTorch build ships no kernels for
ncoevoet
committed
eee97a9
fix(install): route a CUDA 12.6 driver to the cu126 wheel index
ncoevoet
committed
40f1bf3
ci: gate the split locks and the dependabot ignore list
ncoevoet
committed
27c3265
docs: document the three published images and the architecture split
ncoevoet
committed
22838e7
fix(device): ask the arch gate everywhere, not just where it was added
ncoevoet
committed
acd20aa
fix(install): pick the wheel from the weakest card, not just the driver
ncoevoet
committed
41a3cdd
ci(docker): keep the provenance attestation while gating on the smoke test
ncoevoet
committed
410d041
ci: close three blind spots in the gates added last commit
ncoevoet
committed
bcc5cfb
refactor(diagnostics): say the wheel-index advice and the GPU label once
ncoevoet
committed
a97f58e
test(device): cover the arch parse and the probe phase CI could not see
ncoevoet
committed
6f11a93
docs(changelog): record the wheel floor and the widened device gate
ncoevoet
committed
d138724
fix(client): patch the undici CVEs via an override, not an Angular bump
ncoevoet
committed
404d297