[Alssndr0]

Optional patch, off by default, in the same style as qsa-logits-workspace: backport of commit 514102a of vllm-project/vllm#58863 (by Juergen Schmied, suggested by @ArtyomITA; the PR is still open upstream).

What it changes

One file, model_executor/layers/mamba/gdn/recoverssm_gdn.py:

• the verify kernel records the decay it multiplied by, exp(g), instead of g;
• the commit replays the accepted tokens forward with the same update, h = fma(c, k, h * decay), instead of the closed form s0 * decay + corr.

The committed GDN state is then the per-token state bit for bit. The cost is the same: n loads of c, k and the decay.

The v5.2 image has the closed form (the file in sha256:76eda2f6... is the pre-514102a version). On a GB10 the closed form fails the PR's bit-identity checks in tests/kernels/mamba/test_recoverssm_gdn.py. With this change, 62 of 63 of those tests pass; the remaining one is XPU-only.

Tested

• run.sh's patch step against myllmbox/qwen38-flash-next-vllm:v5.2 (digest 76eda2f6...): the file is copied out of the image, the dry run passes, the patch applies, and the result compiles inside the image.
• The patched file is byte-identical to the one I have served since 2026-10-08 on one DGX Spark (hibrid48, and INT4-AutoRound in tests), built into a local image on top of v5.2. Output-quality checks there stayed at each checkpoint's run-to-run KL floor. In that image it ran together with other changes; I'll describe those in a separate issue.

To try it:

server:
  patches: recoverssm-exact-commit

[bilikaz]

Thanks. Our 0.32 image carries #58863 at a head that already includes 514102a, so the exact commit comes in by default there. For the current v5.2 kit, merging this as an opt-in patch.