fix(dmd): only pass driver="gesvd" to torch.linalg.svd on CUDA

[kv-248]

Problem

With dmd_svd_precision="high", _dmd_svd() calls torch.linalg.svd(X, full_matrices=False, driver="gesvd") on every device. PyTorch only accepts driver= for CUDA inputs:

RuntimeError: torch.linalg.svd: keyword argument `driver=` is only supported on CUDA inputs with cuSOLVER backend.

_dmd_fit_one() catches that exception as a degenerate fit and returns None, so on CPU (and any non-CUDA device) the DMD calibrator never forecasts with "high". It quietly reuses the last snapshot, with no warning, while "low" and "medium" work.

Repro on CPU (rank-3 linear dynamics, which DMD represents exactly; 6 snapshots, forecast step 7):

low    rel_err=7.32e-08 returned_newest_snapshot=False
medium rel_err=7.32e-08 returned_newest_snapshot=False
high   rel_err=1.02e-01 returned_newest_snapshot=True

Change

• dmd.py: use driver="gesvd" only when X.device.type == "cuda", and the default driver otherwise. This is the same guard svdquant/lowrank.py already uses (driver="gesvd" if weight.device.type == "cuda" else None). The _dmd_svd docstring now says "high" is CUDA-only and matches "medium" elsewhere.
• tests/api/test_dmd_svd_precision.py: a CPU test that runs DMDState for all three precision levels on a rank-3 trajectory and checks that the forecast matches step 7 and is not just the newest snapshot.

CUDA behaviour is unchanged.

Validation (CPU only, no GPU available)

Python 3.11, torch 2.14.0+cpu, diffusers 0.40.0, on main at 99fc5a8. Rebased onto 5c2d065 (only README changed since) and reran the new test: 3 passed.

• New test on unpatched main: 1 failed, 2 passed ([high] fails). With the fix: 3 passed.
• pytest tests/api tests/utils with the fix: 112 passed, 25 skipped, 1 failed. The one failure is test_load_configs[api/config.yaml], which uses a path relative to tests/; it passes when run from tests/ (1 passed) and is unrelated to this change.
• flake8 --config=setup.cfg is clean on both files; the test file is formatted with the repo's .style.yapf.
• Not tested: the CUDA path (no GPU here). The CUDA branch is unchanged code.

This patch was prepared with the help of an AI coding assistant (Claude Code).

[Copilot]

Copilot review overview

🟢 Approval recommended

The focused fix follows an existing repository pattern and is covered by an appropriate regression test.

Review effort: Balanced
Findings: None

What changed in this PR

Fixes CPU/non-CUDA DMD forecasting when using high SVD precision.

Changes:

• Applies the gesvd driver only to CUDA tensors.
• Adds CPU regression coverage for every precision level.

File	Description
src/​cache_dit/​caching/​cache_contexts/​calibrators/​dmd.py	Guards the CUDA-only SVD driver.
tests/​api/​test_dmd_svd_precision.py	Verifies accurate CPU forecasts.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[DefTruth]

LGTM~ Thx for this fix!