fix(eora): clamp covariance eigenvalues before inversion to avoid outlier adapters

[Qubitium]

@nbasyl @hutm @cmhungsteve Please review. The following bug was encountered by me doing testing on EoRA. The lower the bpw, the higher the chance this becomes and the more damage it will cause. Most of the time it just manifest it self silently as lower EoRA recovery scores and was very sneaky to catch. It often does not manifest as catastrophic failure which is the sneaky part.

Summary

eora_compute_lora eigendecomposes the activation covariance matrix and forms a pseudoinverse via 1 / sqrt(eigenvalue). Numerical noise in the float32 accumulated covariance can yield tiny or slightly negative eigenvalues; the old code replaced negatives with the smallest positive eigenvalue, then computed 1/sqrt(lambda) for every mode. When lambda is near zero this creates arbitrarily large inverse values that propagate as NaN/Inf outliers in the low-rank A/B tensors and, in the worst case, the branch torch.min(L[L > 0]) crashes when every eigenvalue is non-positive.

Clamp the spectrum to a numerical-rank tolerance: discard eigenvalues at or below max(lambda) * max(1, n) * float32_eps by setting their sqrt/rsqrt contributions to zero, so they do not participate in the pseudoinverse. This produces finite adapters for rank-deficient or numerically indefinite covariances.

What Changed

• gptqmodel/eora/eora.py: In eora_compute_lora, replace the minimum-replacement logic with a relative tolerance cutoff and zero-pad the discarded spectrum before sqrt/rsqrt. Also raise FloatingPointError if the eigensolve returns non-finite eigenvalues.
• tests/test_eora.py: New focused unit tests for covariance spectra containing negative/tiny eigenvalues and the all-negative pathological case.

Tests

• I added a new simple/fast unit test for this change.
• I ran the new targeted test locally before opening this PR.

cd tests && python -m pytest test_eora.py -x -q

Result: 5 passed in 2.55s

Review Requirements

• Every changed file has been reviewed by a human.
• I checked that the code matches existing project structure, APIs, and conventions.
• I avoided unnecessary monkeypatching and used the project's normal extension points where possible.

Notes

This fix keeps the existing eora_compute_lora signature and does not change the public EoRA API. The cutoff follows the same float32 numerical-rank reasoning used in the codebase's other Hessian/covariance paths.

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

Added the original EoRA calibration-data balance note to eora_compute_lora as a docstring:

When expanding the calibration data size for EoRA, I suggest maintaining the balance by allocating 50% to general input (C4) and the remaining 50% to downstream task data.

Branch fix-eora-clamp updated and tests still pass (pytest tests/test_eora.py -q → 5 passed).

[Qubitium]

I will merge this first to gpt-qmodel. If issues arise, we can fix in another pr.