cuda: fused FWHT quantizer for 64-wide warps

[bri-prism]

Overview

The Prism release build fails on ubuntu-22-rocm (run 37038350494). The fused FWHT + q8_1 quantizer in quantize.cu asserts QK8_1 == warp_size, so every instantiation fails to compile for the wave64 targets in that job (gfx908, gfx90a, gfx942). windows-hip passes because it only builds 32-wide RDNA targets.

A 64-wide warp holds two complete 32-blocks of the transform output. The quantizer already reduces with warp_reduce_max<QK8_1> / warp_reduce_sum<QK8_1>, which on HIP shuffle with width 32 and so reduce each half on its own. The only 32-wide assumption left is the block index iqs = lane. This changes it to lane % QK8_1 and relaxes the assert to warp_size % QK8_1 == 0. The shared butterfly (ggml_cuda_fwht_block_butterfly) is already generic in the warp size.

On 32-wide warps lane % 32 == lane, so NVIDIA and RDNA behave exactly as before.

Testing

• ROCm 7.2.1 (rocm/dev-ubuntu-22.04:7.2.1-complete, same version as the release job) with the release job's HIP flags (-DHIP_PLATFORM=amd -DGGML_HIP_ROCWMMA_FATTN=ON):
  • prism at a14c7de, GPU_TARGETS=gfx90a: fails in quantize.cu with the same 32 == warp_size static assert as the release run.
  • This branch, GPU_TARGETS=gfx908;gfx90a;gfx942: ggml-hip builds and links, 0 errors.
• Not run on wave64 hardware. The change is checked by reading the reduction width and the store indices.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Claude Code helped find the cause and write the change.

[khosravipasha]

🤖 Reviewing on behalf of the maintainers

The wave64 indexing fix looks correct and preserves wave32 behavior. Please update the PR with the ROCm compile results when available; wave64 hardware validation remains outstanding.