Misc. bug: Qwen4exp, top_k_radix_cuda not compiled/used in Qwen4exp #29326

Name and Version

All versions

Operating systems

Windows

Which llama.cpp modules do you know to be affected?

llama-server

Command line

llama-server affected. Anything that compiles/builds top-k.cu.

Problem description & steps to reproduce

top-k.cu:
Line 51 replacement:
#if !defined(GGML_CUDA_USE_CUB) && defined(GGML_USE_HIP)
-->
#if defined(GGML_USE_HIP) || defined(GGML_CUDA_USE_CUB)

Line 211 replacement:
#endif // !defined(GGML_CUDA_USE_CUB) && defined(GGML_USE_HIP)
-->
#endif // defined(GGML_USE_HIP) || defined(GGML_CUDA_USE_CUB)

Line 227 insert:
#if defined(GGML_CUDA_USE_CUB)
// Serial top-k launches one pass per row and leaves the GPU idle. For big multi-row inputs (indexer top-k
// at prefill) the parallel radix kernel wins despite reading the data more times.
if (ncols > 1024 && nrows >= 16) {
top_k_radix_cuda(pool, src0_d, dst_d, ncols, nrows, k, stream);
return;
}
#endif

Enables the top-k radix to be used in CUDA builds.

Qwen4exp (Qwen38FlashNext) long term context Prompt Processing gains of +20% at all ranges. Tetsted up to 148k context, gains held throughout.

Tests pass locally, tested on Windows SM75

First Bad Commit

top-k.cu only targets rocM.

Relevant log output