Summary

Keeps data on the GPU on the device="gpu" path (closes #14 ) :

• Truncation SVD now runs on the array's own device (xp.linalg.svd) instead of copying R to the host with .get() and calling np.linalg.svd.
• New to_host: bool = True argument on apply / compress. With to_host=False the result stays backed by CuPy arrays and can be fed straight into the next call without a device/host round trip. The default preserves the previous behaviour, so existing callers are unaffected.
• CPU path is unchanged.

Performance

Measured on Google Colab:

• GPU: Tesla T4 (15 GB), driver 580.82.07
• CPU: Intel Xeon @ 2.00 GHz, 2 vCPU
• cupy 14.2.0, CUDA runtime 12.9, numpy 2.4.4, Python 3.14.8
• before = fe1779b (main), after = fbec6ae (this branch)

Method: complex128, seed=0, random MPO × MPS, 20 sites, phys_dim=4, median of 10 runs. Each path is warmed up first, and before/after are timed in separate processes on the same session.

chi_out = 64 (input bond dim 64)

chi_out = 128 (input bond dim 128, after only)

Testing

• Full test suite on the T4: 40 passed in ~70 s.
• Chained apply(..., to_host=False) calls accept CuPy inputs and run without host round trips of the result.

Acceptance criteria

• No .get() / D2H transfer inside the per-site loop
• Chained GPU apply calls do not round-trip the result through host memory
• CPU results and numerics unchanged
• GPU tests pass
• Before/after timings recorded above, with hardware