fix(cuda): increase argmax topk kernel limit from 32 to 64

[chimpera]

Summary

• Increases the custom topk_f32 CUDA kernel limit from K ≤ 32 to K ≤ 64 by expanding the fixed-size register arrays
• Raises the CUB TopK auto-fallback threshold from K > 32 to K > 64 so the custom path covers the full top_k=64 range

Problem

On CUDA 12.x / pre-CCCL 3.2 Linux builds, the CUB TopK fallback (GGML_CUDA_DFLASH_CUB_TOP_K_AVAILABLE) is unavailable because it requires CCCL ≥ 3.2. The custom topk_f32 kernel is the only path, and it has a hardcoded K ≤ 32 limit.

The Gemma 4 GGUF bakes general.sampling.top_k = 64 into its metadata. The DFlash reduced verifier picks this up and passes K=64 to the target model's verification logits, which hits:

argmax.cu:557: GGML_ASSERT(K <= 32) failed

This crashes the server on the first inference request with any Gemma 4 model + DFlash on CUDA 12.x builds.

Fix

• heap_val[32] / heap_idx[32] → heap_val[64] / heap_idx[64] (register arrays, +256 bytes per thread)
• GGML_ASSERT(K <= 32) → GGML_ASSERT(K <= 64)
• CUB auto-threshold: K > 32 → K > 64

Shared memory at K=64, 32 warps: ~16 KB — well within the 48 KB default limit. Register pressure increase is modest and causes no issues on tested hardware (RTX 5090, RTX 3090).

Test plan

• Build succeeds on CUDA 12.x (GCC, Linux) with GGML_CUDA_FA=ON, GGML_CUDA_FA_ALL_QUANTS=ON
• Launch Gemma 4 31B + DFlash draft model with default sampling (top_k=64 from metadata) and verify first inference completes without GGML_ASSERT
• Run Qwen 3.6 27B + DFlash (top_k=20) to verify no regression on K ≤ 32 path

🤖 Generated with Claude Code

[Anbeeld]

Seems legit. Thanks!