[AryanRG]

Krum currently sums Euclidean distances, but the paper defines scores using squared Euclidean distances. The rankings can differ: with rows [[3,0],[0,7],[2,5],[5,0],[6,7]] and n_byzantine=1, Krum selects [3,0] instead of [2,5].

Remove the square root while retaining the roundoff clamp. Add regressions for Krum, Multi-Krum, selection boundaries, and an autograd/SGD step through both Gramian computation modes.

Validated the full CPU unit suite, focused float32/float64 aggregation and autograd tests, Ruff, full type checking, and wheel/sdist contents plus wheel-import smoke tests. The unit suite retains its existing skips and expected failures; CUDA and the full CI matrix were not run locally.

[AryanRG]

Could a maintainer add cc: fix? The title check requires it, and I don't have permission to set labels.

[ValerianRey]

/opencode:Plan please review this. In particular:

• find the official implementation of Krum and see what they did
• read the paper and verify that the PR's claim is correct
• if this bug was introduced after the commit adding Krum, find out at which commit

[opencode-agent]

APIError: This API key has expired. Create a new API key in OpenCode Console and replace it in your client or secret store. Signing in again will not renew this key.

github run

[ValerianRey]

APIError: This API key has expired. Create a new API key in OpenCode Console and replace it in your client or secret store. Signing in again will not renew this key.

github run

@PierreQuinton

[AryanRG]

@ValerianRey the authors' Krum repository points to AggregaThor. Its implementation computes squared distances and sums the closest n-f-2, matching section 4 of the paper.

In TorchJD, unsquared scoring was already present in the initial commit: torch.cdist followed by a sum. It predates the later refactors and clamp change. Squaring preserves each row's nearest-neighbor order, but can change which row has the lowest total score, as the PR's regression demonstrates.