Point the nnterp pin at our fork for the text-tower fix

[can-goodfire]

Unblocks us from waiting on ndif-team/nnterp#52 to be reviewed.

Why

nnsight's LanguageModel refuses any config registered with AutoModelForImageTextToText. qwen3_5_moe — the text tower of Qwen3.6-35B-A3B, and the target of the hookpoint-vocabulary work — is registered there and under AutoModelForCausalLM, so the tower loads cleanly and the refusal is over-broad. Without the fix, nnterp cannot load the model at all.

#52 fixes it upstream and is open for review. This repoints the existing git source at our fork in the meantime.

What changes

One line plus its comment:

-nnterp = { git = "https://github.com/ndif-team/nnterp", rev = "b4a3127…" }
+nnterp = { git = "https://github.com/can-goodfire/nnterp", rev = "45f386b…" }

45f386b is the previous upstream pin b4a3127 plus the two commits in #52, so this is a strict superset of what we had — the packaging fix (#49) that motivated the git source in the first place is still in there. The comment now records both reasons and how each retires:

• when PR3 — round-1 MoE components: the router, the routed output, the shared expert #52 merges → repoint at ndif-team
• when a PyPI wheel ships both fixes → drop the git source entirely

uv lock --upgrade-package nnterp moved only nnterp (1.3.1.dev14+gb4a312746 → 1.3.1.dev16+g45f386b7c); the other 239 packages are unchanged.

Verification

Installed from the fork's public URL rather than a local path, to prove the fork actually unblocks a clean checkout:

uv pip install "nnterp @ git+https://github.com/can-goodfire/nnterp@can/text-tower-loading"

tiny-random/qwen3.5-moe then loads as Qwen3_5MoeForCausalLM with no vision tower attached, and traces to logits.

One caveat worth knowing

This branch still resolves transformers 4.57.1, which carries no qwen3_5_moe. So the fix ships here but cannot be exercised on this lineage until a transformers >= 5.16 bump reaches it. That bump landed downstream on can/protocol-refactor (#46, merged) and whether it should also come to the path to main is a separate call — flagging rather than bundling it.

Also worth noting: can/protocol-refactor has no nnterp dependency at all, so the round-1 hookpoint work is not blocked by #52 — its pytorch_hooks backend loads the tower through AutoModelForCausalLM directly.

🤖 Generated with Claude Code

[can-goodfire]

NNterp implemented fixes upstream