NVIDIA lock-in makes switching costly

Cloud and tool cost2 posts from 2 people2d active+11 post in the last 7 days, 0 the 7 days before (steady)Unverified

Posts per day2 posts, Sep 4 to Sep 25

The posts behind it

2, newest first

Post	Where	Date
[MPS] torch.histc and torch.linalg.solve_triangular are 9-1800x slower than alternatives on MPS## 🐛 Describe the bug Two MPS ops are slow enough to dominate inference of common models in `transformers` on Apple Silicon. 1. `torch.histc` costs about 0.17 ms per bin, independent of input size. An equivalent `scatter_add_` or `bincount` is 40-1800x faster: \| elements \|...pytorch/pytorchnassersala	pytorch/pytorchnassersala1 reactions, 6 replies	Sep 232 days ago
CUDA 13.4 DP libcublasLt.so.13 missing cublasLtGroupedMatrixLayoutCreate declared by its own headers...installed via `.ci/docker/common/install_cuda.sh`) produces a `libtorch_cuda.so` that fails to import at runtime: https://github.com/pytorch/pytorch/actions/runs/34293334892/job/102284542319 ## Repro The error is shown when adding ldd -r "$(python -c "import...pytorch/pytorchtinglvv	pytorch/pytorchtinglvv1 reactions, 5 replies	Sep 43 weeks ago

Companies and products named

Company or product	Posts naming it
NVIDIA	2
PyTorch	2
Claude	1
GitHub	1
Docker	1

About this problem

Evidence

2 posts from 2 people in 1 place, about 1 a week over 20 days. Mostly on pytorch/pytorch. Tools named alongside: PyTorch, Claude, GitHub, Docker.

Frustration Frustration 0 of 3· Seen on GitHub issues

How it was grouped

Posts that state a pain and match the "vendor-lock" rule about NVIDIA. First post Sep 4, 2026, latest Sep 23, 2026. Unverified: fewer than 3 posts, or one person or place only. Method

The same problem elsewhere

Posts per day

The posts behind it

Companies and products named

Rising problems by email