[BUG] SM120 blockwise FP8 GEMM: TMA descriptor misalignment in Params struct

Which component has the problem?

CUTLASS C++

Bug Report

Summary

TMA descriptors in the Params struct of sm120_mma_tma_blockwise_scaling.hpp and sm90_epilogue_tma_warpspecialized.hpp are not 64-byte aligned, causing prefetch.tensormap to fail with "misaligned address" error on SM120 (RTX 5090).

Environment

• GPU: NVIDIA GeForce RTX 5090 (SM 12.0)
• Driver: 591.44
• CUDA Toolkit: 13.1
• CUTLASS: 4.3.3 (commit d55f6beeebb6df501a250dc82827db97660f06e0)
• OS: Windows 11
• Location: Bare-metal

Steps to Reproduce

1. Build CUTLASS example 87a with -arch=sm_120a:

nvcc -std=c++17 -O3 -arch=sm_120a --expt-relaxed-constexpr   -I cutlass/include   -I cutlass/tools/util/include   -I cutlass/examples/common   cutlass/examples/87_blackwell_geforce_gemm_blockwise/87a_blackwell_geforce_fp8_bf16_gemm_blockwise.cu   -o example87a.exe

2. Run the example:

./example87a.exe

3. Observe: misaligned address error

Root Cause Analysis

Using printf debugging, we traced the crash to prefetch_tma_descriptors():

// sm120_mma_tma_blockwise_scaling.hpp:359
static void prefetch_tma_descriptors(Params const& mainloop_params) {
    cute::prefetch_tma_descriptor(mainloop_params.tma_load_a.get_tma_descriptor());
    // ^^^ CRASH HERE
}

Printing TMA descriptor addresses revealed misalignment:

[TMA] desc_a=0000000200610398 (mod64=24)  // NOT 64-byte aligned\!
[TMA] desc_b=0000000200610420 (mod64=32)  // NOT 64-byte aligned\!

The prefetch.tensormap PTX instruction requires 64-byte alignment:

prefetch.tensormap [%0];  // requires 64-byte aligned address

The Params struct contains TMA_A tma_load_a and TMA_B tma_load_b members without alignas(64), causing the embedded TMA descriptors to be misaligned.

Expected Behavior

TMA descriptor members in Params struct should be 64-byte aligned to satisfy prefetch.tensormap requirements.

Affected Files

1. cutlass/gemm/collective/sm120_mma_tma_blockwise_scaling.hpp (line ~254)
2. cutlass/epilogue/collective/sm90_epilogue_tma_warpspecialized.hpp (line ~244)

Note: The SM90 epilogue file is also affected, which may impact SM90/SM100 kernels as well, though we only tested on SM120.

---

Appendix: Proposed Fix

Adding alignas(64) to the Params struct and TMA members resolves the issue:

Fix for sm120_mma_tma_blockwise_scaling.hpp:

// Device side kernel params
struct alignas(64) Params {  // Add alignas(64) to struct
    using TMA_A = /* ... */;
    using TMA_B = /* ... */;
    
    alignas(64) TMA_A tma_load_a;  // Add alignas(64) to member
    alignas(64) TMA_B tma_load_b;  // Add alignas(64) to member
    // ...
};

Fix for sm90_epilogue_tma_warpspecialized.hpp:

struct alignas(64) Params {  // Add alignas(64) to struct
    using TMA_C = /* ... */;
    using TMA_D = /* ... */;
    
    typename FusionCallbacks::Params thread{};
    alignas(64) TMA_C tma_load_c;  // Add alignas(64) to member
    alignas(64) TMA_D tma_store_d;  // Add alignas(64) to member
    // ...
};

Results After Fix

[TMA] desc_a=00000019006103C0 (mod64=0)  // ALIGNED\!
[TMA] desc_b=0000001900610480 (mod64=0)  // ALIGNED\!
[KERNEL] After mainloop prefetch, before epilogue prefetch
[KERNEL] After epilogue prefetch_tma_descriptors
[FP8 GEMM SM120] GEMM completed OK

The kernel runs without crashing after the alignment fix.

---

Notes on Pull Request

We have not submitted a PR because:

1. SM90 impact unknown: The epilogue fix touches sm90_epilogue_tma_warpspecialized.hpp, which is used by SM90/SM100 kernels. We do not have access to H100/B100 to verify this does not break existing functionality.

2. Separate correctness issue: After fixing alignment, the kernel runs but produces ~12% relative error. This is likely a separate scale factor issue, not related to alignment.

We are happy to submit a PR if the CUTLASS team confirms the SM90 change is safe.

Related Issues

• #2902 - Different alignment issue (partition_S() dropping LDSM alignment)
• #2820 - Arch conditional MMA assertion (solved by -arch=sm_120a)

Start with the Params definitions and prefetch_tma_descriptors() in cutlass/gemm/collective/sm120_mma_tma_blockwise_scaling.hpp and cutlass/epilogue/collective/sm90_epilogue_tma_warpspecialized.hpp. Build and run example 87a on SM120 to confirm descriptor alignment and the original failure. Done means the affected descriptors satisfy the alignment requirement and the relevant SM90/SM100 paths remain functional; the separate FP8 accuracy issue is out of scope.