vulkan: GATED_DELTA_NET silently wrong for raw_gates=1 (ERR ~1.0, 3 cases, 15/15 runs) — supports_op claims a path the shader doesn't implement

[bri-prism]

Summary

On the Vulkan backend, GATED_DELTA_NET returns essentially garbage for every raw_gates=1 case it claims to support. test-backend-ops reports ERR between 0.90 and 1.00 against the CPU reference (tolerance 1e-7), i.e. the output is uncorrelated with the expected result rather than slightly off. It is silent — no abort, no warning, no fallback.

This is not caused by any open PR: it reproduces on stock prism @ 422590f5d (current HEAD at time of filing).

Environment

Intel(R) Arc(TM) B390 GPU | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 |
shared memory: 49152 | int dot: 1 | matrix cores: KHR_coopmat

Intel proprietary Windows driver, Windows 11. Build: prism @ 422590f5d, MinGW/ucrt64 g++, Ninja, -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release.

Reproduce

test-backend-ops test -o GATED_DELTA_NET -b Vulkan0

Result

Three cases fail, all head_count=4, head_size=128, raw_gates=1:

GATED_DELTA_NET(type=f32,head_count=4,head_size=128,n_seq_tokens=1, n_seqs=1,v_repeat=1,permuted=0,kda=0,K=1,rows_mode=0,cache_rows=-1,raw_gates=1)
GATED_DELTA_NET(type=f32,head_count=4,head_size=128,n_seq_tokens=1, n_seqs=2,v_repeat=2,permuted=0,kda=0,K=1,rows_mode=0,cache_rows=-1,raw_gates=1)
GATED_DELTA_NET(type=f32,head_count=4,head_size=128,n_seq_tokens=64,n_seqs=1,v_repeat=1,permuted=0,kda=0,K=1,rows_mode=0,cache_rows=-1,raw_gates=1)

ERR over five consecutive runs (the op's test data is unseeded, so the values move; the verdict does not):

case	run1	run2	run3	run4	run5
nst=1, ns=1, vr=1	0.897	0.996	0.951	0.954	0.992
nst=1, ns=2, vr=2	0.988	0.986	0.992	0.958	0.984
nst=64, ns=1, vr=1	1.000	1.000	1.000	1.000	1.000

15/15 failures — fully reproducible.

What isolates it to raw_gates

Within the same runs, on the same device and build:

	cases	OK	FAIL
raw_gates=0	45	36	0
raw_gates=1	4	0	3 (+1 reported not supported)

Every raw_gates=0 case passes. No raw_gates=1 case has ever passed. The fourth (K=2, rows_mode=1) is correctly declined by supports_op.

The SYCL backend on the same machine and same commit is the useful control — it declines all four:

GATED_DELTA_NET(...raw_gates=1): not supported [SYCL0]   x4

So SYCL says "I don't implement this" and is correct to; Vulkan says "I implement this" and then computes the wrong answer. The bug is the combination — supports_op advertising raw_gates=1 support that the shader does not actually deliver.

Why it matters

A wrong-but-silent result is worse than an abort here: anything running a gated-delta-net model down the Vulkan path with raw gates gets corrupted state with no diagnostic. Given the tolerance is 1e-7 and the observed error is ~1.0, this is not a precision/accumulation issue — the code path looks functionally wrong rather than imprecise.

Two acceptable fixes, in preference order: make the shader correct for raw_gates=1, or have supports_op decline it (as SYCL does) until it is, which converts silent corruption into a clean CPU fallback.

Related

PR #187 adds a fourth failing case on top of these three (n_seqs=2, v_repeat=1, ERR ~0.92, 4/4 runs) — reported separately at #187 (comment 5765014167). That is a distinct regression; the three above are independent of it and present on stock prism.

[bri-prism]

Fix up as #239 — declines raw_gates in supports_op so these ops fall back to the CPU instead of silently returning wrong results, matching what the SYCL backend already does.

With it applied on the same Arc B390: test-backend-ops test -o GATED_DELTA_NET -b Vulkan0 reports 36/36 passed, Backend Vulkan0: OK on three consecutive runs, and the four raw_gates=1 cases report not supported instead of failing. The full unfiltered suite goes from FAIL (17187 OK / 3 failing) to OK (17188 OK / 0 failing) — the Vulkan backend passes on this device for the first time.

It does not implement raw gates on Vulkan; that needs two new shader bindings (dt_bias, a) plus the sigmoid/softplus math and care around the kda path. This just stops the silent corruption. Leaving this issue open as the record of the missing feature — worth closing only if someone implements the shader side.