Repository navigation
GPU based quadratic solver #591
Description
- qpth (locuslab/qpth)
- Pure PyTorch, GPU-native interior-point solver
- Accepts Q as (m, m) and h as (batch, m) — it auto-broadcasts Q, which directly exploits your shared-Gramian structure
- Your constraint maps cleanly: G_ineq = -I, h_i = -u_i
- Caveat: minimally maintained, and numerically sensitive — recommend float64
- lqp_py (ipo-lab/lqp_py)
- PyTorch + GPU, designed for batched QPs
- Only supports box constraints (l <= x <= u), but your u_i <= v is a half-bounded box (l = u_i, u = +inf), so it maps cleanly
- Differentiable
- Minimally maintained, simpler API
- cvxpylayers
- Declare once in CVXPY, run as a PyTorch layer; GPU via CuClarabel backend
- Well-maintained, general, but more overhead than specialized solvers, possibly not batched
Activity
ValerianRey commentedon May 3, 2026
See #664
The main problem is that it seems too slow, but maybe it's still faster than quadprog for large m. It's also not maintained enough.
Not sure we even need cvxpylayers, maybe cvxpy is enough to run CuClarabel. However, it seems that CuClarabel requires Julia, which would be a big new dependency. It also seems slow according to what people say here. But in my opinion it's worth trying: it might be decently fast in our use-case.