Skip to content

fix(vllm): apply every steer op at a site, as eager does - #29

Merged
hijohnnylin merged 1 commit into
mainfrom
fix/vllm-steer-ops-per-site
Sep 28, 2026
Merged

hijohnnylin merged 1 commit into
mainfrom
fix/vllm-steer-ops-per-site

Conversation

@hijohnnylin

Copy link
Copy Markdown
Contributor

vLLM kept only the last steer op per site (hooked and static). Ops at a site now compose in order, matching eager.

Tests: CPU parity in tests/test_steer_math_parity.py; GPU check against eager in tests/test_steer_ops_per_site_gpu.py (fails on main at ~0.83 cosine).

Made with Cursor

The hooked and static writes kept one op per site, so two ops on one layer steered with the last
alone. The ops at a site now compose in order; a static site of plain adds takes their sum.

Co-authored-by: Cursor <cursoragent@cursor.com>
@hijohnnylin
hijohnnylin merged commit 13e898c into main Sep 28, 2026
8 checks passed
@hijohnnylin
hijohnnylin deleted the fix/vllm-steer-ops-per-site branch September 28, 2026 18:17
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant