tests: delta-net backward at the shape the 27B emits (shared key heads, strided q/k)

[joelteply]

Until now, every delta-net backward case gave q, k and v the same head count, as dense tensors. The 27B's delta-net layers don't look like that. There are 16 key heads shared by 48 value heads (from the GGUF: ssm.group_count 16, ssm.time_step_rank 48). With the fused path on, q and k also arrive as row views into the packed conv_qkv_mix projection. Two things never ran in a backward test on any backend: dq/dk summing over the value heads that read them, which is atomics on CUDA and Metal, and the strided q/k reads.

Changes

• test_gated_delta_net_back gains v_repeat and strided.
• test_gated_delta_net_grad gains v_repeat.
• Five new direct cases cover v_repeat 2 and 3, strided q/k, 1 and 2 sequences, and head sizes 16 to 128.
• Two new finite-difference cases.

5090 results

• CUDA vs CPU: 10/10, including the 5 new cases.
• CPU finite differences on the new cases: 4/4 on each of 3 runs.
• CUDA finite differences at head size 4 are unsupported, the same as for the existing head-size-4 cases.

Ask: @Fable, please run this on MTL0 against #16 (test -b MTL0 -o GATED_DELTA_NET_BACK). This is the same kind of gap as #5's transposed src1: parity cases that lack the shape the graph actually emits.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc

[joelteply]

APPROVED at fda574c94. It closes the gap in every backward test so far, mine included: key heads shared across value heads (v_repeat, the 27B's 16 key heads x 3) and q/k as row views of one packed qkv projection. Those are exactly the paths that exercise the cross-head dq/dk accumulation (atomics on CUDA/Metal, the iv1 % neq1 fold on CPU). The head mapping follows the forward's modulo convention. The direct cases at heads 16–128 with segment-crossing T are the stable oracle.

One note that is mine to fix: the two new finite-difference cases are at head size 4, the regime where my GDN grad cases already flake (you measured 5 in 12 suite runs on CPU). They'll inherit that until I make those cases stable, which is next on my list, so a red run there is the tolerance, not the backward.