[Upstream ds4 bug] v0.5.6.2 cuda-regression target does not compile

[NotSamClemens]

Summary

The upstream Entrpi/ds4 tag v0.5.6.2 builds successfully for DGX Spark with make cuda-spark, but its committed cuda-regression target does not compile because tests/cuda_long_context_smoke.c still calls two CUDA APIs with their older signatures.

Entrpi/ds4 currently has GitHub Issues disabled, so I am reporting the engine release-test defect here in the associated ds4-on-spark tracker.

Environment

• NVIDIA DGX Spark, GB10 / SM121
• Ubuntu 24.04.4 LTS, aarch64
• CUDA 13.0 (Build cuda_13.0.r13.0/compiler.36424714_0)
• DS4 tag: v0.5.6.2
• DS4 commit: 027714a4c290a756ef3e6ca557426528745f2033
• Build: make cuda-spark -j20 succeeds

Reproduction

git clone --branch v0.5.6.2 --single-branch https://github.com/Entrpi/ds4.git
cd ds4
make cuda-spark -j20
make cuda-regression -j20

The regression target fails while compiling tests/cuda_long_context_smoke.c:

tests/cuda_long_context_smoke.c:44:13: error: too few arguments to function ‘ds4_gpu_indexer_topk_tensor’
tests/cuda_long_context_smoke.c:111:9: error: too few arguments to function ‘ds4_gpu_attention_decode_heads_tensor’

Cause

The release changed the declarations in ds4_gpu.h, but the calls in tests/cuda_long_context_smoke.c were not updated.

1. ds4_gpu_indexer_topk_tensor now also requires n_comp_max and il_for_decode1. The header documents (0, UINT32_MAX) for legacy/non-decode1 callers.
2. ds4_gpu_attention_decode_heads_tensor now includes comp_kv_f16 before n_comp, followed by scalars, il_for_decode1, comp_fp8, and comp_scale. The regression call still uses the previous argument layout.

Independent validation

I left the tagged source unchanged and compiled a temporary copy of only the regression test with the header-documented fallback arguments:

ds4_gpu_indexer_topk_tensor(selected, scores, n_comp, n_tokens, top_k,
                            0, UINT32_MAX)

For the attention call, I added 0 for comp_kv_f16 and the trailing NULL, UINT32_MAX, NULL, NULL fallback arguments.

That temporary compatibility test compiled against the exact v0.5.6.2 CUDA objects and completed successfully on GB10:

ds4: CUDA backend initialized on NVIDIA GB10 (sm_121)
cuda-regression: top-k n_comp=32768 n_tokens=32 elapsed=0.000s
cuda long-context regression: OK

This indicates a stale test call-site problem rather than a failure of the exercised CUDA runtime path.

Expected behavior

The tagged v0.5.6.2 make cuda-regression target should compile and run without requiring local source changes.

Could the two call sites in tests/cuda_long_context_smoke.c be updated to the new signatures and included in the next tag?