---

name: lora_merge_latency description: Merge LoRA low-rank factors into a dense weight and verify deployment-equivalent inference.

LoRA Merge Latency

Use this skill when deploying or validating LoRA without extra inference latency. It consumes a base weight and trained factors and emits a merged dense matrix plus equivalence checks.

Inputs

W0, A, B, alpha, and test inputs.

Outputs

Merged weight W, unmerged-vs-merged maximum error, and optional restored base weight.

Workflow

Compute W = W0 + (alpha/r) BA, run the ordinary dense layer, compare it with the unmerged LoRA forward pass, and record a tight numerical tolerance.

Validation

Run the included tests or validate the skill tree with tests.

Limitations

Merging one adapter at a time is appropriate for single-task batches; mixed adapters may require unmerged dynamic routing.