---
title: "2026-08-31 ve2e removal of legacy inference adapter"
description: "Removed old checkpoint translation after adopting current-format models as the inference contract."
image: "https://docs.virdx.dev/img/virdx-social-card.png"
---

> Documentation Index
> Fetch the complete documentation index at: https://docs.virdx.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# 2026-08-31 ve2e removal of legacy inference adapter

## Work Done

- Deleted `src/ve2e/inference/legacy/` and its old config and DinoFlex translation code.
- Replaced the compatibility fallback with a direct `torch.load` of current self-contained checkpoints.
- Reduced the inference package to `checkpoint.py`, `classification.py`, and its public `__init__.py`; the largest file is 90 lines.
- Passed Ruff, formatting, compilation, mocked benchmark orchestration, a direct current-checkpoint loader smoke test, complexity checks, and calldiff inspection.
- Appended and normally pushed commit `aa065752` to `feat/isup-inference`.

## Lessons Learned: Pitfalls

- Temporary compatibility code can become most of a small package. Keeping it after the supported model contract changes obscures the main inference path.

## Lessons Learned: Improvements

- State checkpoint-format support explicitly. When only current self-contained checkpoints are supported, load them directly and let incompatible checkpoints fail normally.

Source: https://docs.virdx.dev/knowledge/inbox/2026-08-31-clinical-ve2e-drop-legacy-inference-adapter/index.mdx
