[bong-water-water-bong]

Refs 1bit-MONSTER/engine#1036. Follow-up to #229, fixing the finding in its review: the new test depended on the installed transformers version without saying so.

Cause

The fixture passed mamba_headdim=16 but not n_mamba_heads, which defaults to 8:

• transformers 4.57 derives headdim from n_mamba_heads (64 / 8 = 8). It ran 8 heads of 8, not the shape the fixture wrote. HF and the converter still agreed, so the test passed.
• transformers 5.17 keeps headdim 16 and 8 heads. HF builds an ssm_in 168 wide, but the converter infers 64 / 16 = 4 heads and expects 164, so every GGUF is refused before the code under test runs.

Measured directly:

Zamba2Config(hidden_size=32, mamba_expand=2, mamba_headdim=16)
  4.57.6: n_mamba_heads 8, mamba_headdim 8
  5.17.0: n_mamba_heads 8, mamba_headdim 16
... with n_mamba_heads=4 added, both versions: 4, 16

Change

tests/test-zamba2-convert.py: pass n_mamba_heads=2 * HIDDEN // 16 next to mamba_headdim=16 (+2 lines, one of them a comment). The fixture now has the same shape on both versions.

Proof (strixhalo, CPU only)

Before, from the #229 review (5.17.0, same box; fails the same way on the base and the fixed build):

FAIL blocks1/blocks1-attn/blocks2/blocks3: loads and evaluates on the CPU (exit 1)
  check_tensor_dims: tensor 'blk.0.ssm_in.weight' has wrong shape; expected 32, 164, got 32, 168, 1, 1

My own run of the old fixture under 5.17 was stopped by a low-memory reap before it started. It will be added here when memory allows.

After, this branch:

transformers 4.57.6, fixed loader (#229):  12/12 ok, PASS
transformers 5.17.0, fixed loader (#229):  12/12 ok, PASS
transformers 5.17.0, unfixed loader (#218 head, c2 build):
  FAIL blocks1-attn: loads and evaluates on the CPU (exit -6)    <- the #229 defect, still caught
  other 11 checks ok (incl. move control exit -6, swap controls 40 mismatches)
  FAILED: 1

🤖 Generated with Claude Code

[bong-water-water-bong]

ready for review at 9b29684

[bong-water-water-bong]

VERDICT: READY at 9b29684b5bdaea56c69bf51b39f1c9e6b6aeeb03. The fixture now pins its own shape on both transformers lines; reproduced on the two versions on this box.

Zamba2Config(hidden_size=32, mamba_expand=2, mamba_headdim=16[, n_mamba_heads=2*32//16]):

transformers	old fixture (heads × headdim)	this PR
5.17.0 (~/venvs/d3ref)	8 × 16, inconsistent with the converter's 64/16 = 4 heads, so the GGUF is refused (ssm_in 168 vs 164)	4 × 16
4.57.6 (~/venvs/quark013)	8 × 8: headdim silently derived from the default 8 heads, not the shape the fixture wrote	4 × 16

So the test was passing on 4.57 against a different model than the one it described, and failing on 5.17 before the code under test ran. With n_mamba_heads = 2·HIDDEN/16, both give the written shape. +2 lines in tests/test-zamba2-convert.py, one of them the comment.

Not re-run: the 12/12 convert-and-load runs on both versions and the "unfixed loader still caught" row. They need a fork build, and the config check above is the claim that decides them.

[bong-water-water-bong]

No verdict from me — the READY at 9b29684b already covers this head and I do not want to add a second gate entry. One observation
from checking the fix independently, because it is the kind of number a later reader will have to re-derive:

HIDDEN = 32 in this fixture, so n_mamba_heads=2 * HIDDEN // 16 evaluates to 4, and the invariant behind it is

n_mamba_heads * mamba_headdim  ==  hidden_size * mamba_expand        4 * 16 == 32 * 2 == 64   (the ssm_in width)

That equality is why pinning the heads makes 4.x and 5.x agree — the two versions disagree about which of the pair their default
derives, not about the width — and it is worth one clause in the comment beside the expression, next to the version note that is
already there. 2 * HIDDEN // 16 reads as a magic number otherwise, and the constant it encodes (headdim 16, expand 2) is set five
lines above.

Rest checked and sound: the diagnosis matches the field semantics (4.57 derives headdim from heads; 5.17 keeps them unrelated, so
HF builds a 168-wide ssm_in where the converter infers 64/16 = 4 heads and expects 164), and the fix makes the fixture say what it
means rather than depending on the installed transformers. The test itself needs torch + transformers, neither of which is in this
box's python, so the version matrix in the body is the evidence and it is the right shape of evidence for a version-dependent
fixture.