[BUG] Example 93 GQA reduction races when reusing shared scratch

Which component has the problem?

CUTLASS C++

Bug Report

Describe the bug

Example 93 can overwrite shared maximum-reduction scratch before all four warps finish reading it. Under Compute Sanitizer racecheck, this produces shared-memory hazards and incorrect attention output.

Steps/Code to reproduce bug

In a checkout of 59e3a3338d516ca6ce0e073af8da65289678a35c:

curl -fL https://raw.githubusercontent.com/1sgtpepper/cutlass/923e9400d6901f90f30d18510320d08d838fbcac/test/regression/gqa_max_scratch.cu -o gqa_max_scratch.cu
nvcc -std=c++17 -O3 -lineinfo --expt-relaxed-constexpr \
  -gencode=arch=compute_100a,code=sm_100a \
  -Iinclude -Itools/util/include -Iexamples/93_blackwell_low_latency_gqa \
  gqa_max_scratch.cu -o gqa_max_scratch
compute-sanitizer --tool racecheck --error-exitcode 3 \
  ./gqa_max_scratch --repetitions 3

The fixture calls the production kernel with two CTAs, 512 KV tokens, head dimension 64 and three stages. A double-precision softmax reference checks BF16 output with 1/256 absolute tolerance. One run produced 2,193 wrong values and eight race hazards; counts vary with scheduling.

Expected behavior

Every participating warp finishes reading a reduction's scratch before another reduction reuses it.

Environment details (please complete the following information):

• CUTLASS 59e3a3338d516ca6ce0e073af8da65289678a35c, main as of September 8, 2026. The affected reduction is unchanged in v4.8.0dev and v4.7.1; these releases were not executed.
• B200/SM100, driver 580.95.05; CUDA 13.1.1 and bundled Compute Sanitizer.
• Modal cloud GPU, Ubuntu 24.04 container, Linux x86_64.

Additional context

cta_reduce synchronizes before reading warp partials but not after reading them. Its caller reuses the same scratch for subsequent KV tiles. Ordinary launches passed; the numerical failures above were observed under racecheck.

Start with the production kernel's cta_reduce path in Example 93 and the gqa_max_scratch.cu fixture linked in the report. Build the fixture with nvcc, then run it under Compute Sanitizer racecheck to reproduce the shared-memory hazards. Done means the fixture reports no race hazards and its BF16 output passes the stated reference tolerance.