tests : GET_ROWS -> RESHAPE -> GATED_DELTA_NET case, eager + CUDA graph capture/replay

[professorpalmer]

Overview

The test owed from #220 (Copilot review there, and my note on #220's close): the GET_ROWS -> RESHAPE -> GATED_DELTA_NET graph that ggml_cuda_try_gdn_gather_skip folds into the GDN kernel, with a nonzero cache row, run eagerly and through CUDA graph capture and replay.

• test_case::n_eval() (default 1): evaluate the same graph that many times, with initialize_tensors between runs. On CUDA the first run is eager, the second captures a graph, the rest replay it.
• test_gated_delta_net_gathered_state: one sequence, state gathered from a cache of 5 (or 3) rows, n_eval() = 4, and ids points at a different nonzero row on every run. A replay that kept the captured row would fail on run 3.

No new test file; this is in test-backend-ops.

Additional information

RTX 4070 (sm_89), CUDA 13, Windows, -o GATED_DELTA_NET -b CUDA0: 41/41 with the fusion on (default), with GGML_CUDA_GDN_GATHER_FUSION=0, and with GGML_CUDA_DISABLE_GRAPHS=1.

test-backend-ops cannot see whether the gather was actually skipped or a graph actually captured, so I checked that once with a local program (not in this PR) on the same graph: with the fusion on the gathered tensor stays untouched (NaN-filled before each run) and the backend logs CUDA graph warmup complete; with the fusion off it holds the selected row; with graphs disabled no capture happens. All four runs match the CPU in every mode (NMSE ~1e-14).

Negative control: with the fused kernel hard-wired to cache row 1 (gated_delta_net.cu, the s_ids[sequence] read), both new cases fail (ERR 1.9e-5 and 1.02 against 1e-7) and pass again with GGML_CUDA_GDN_GATHER_FUSION=0. The first evaluation reads row 1, so a single-evaluation test would have passed; the failure comes from the later runs.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - the test case, the n_eval hook and the checks above were written and run with Claude Code (Claude Opus 5.5).

🤖 Generated with Claude Code