[Bugfix][Layout] Add cuda_fp8.h include to fix __nv_fp8_e4m3 undefined

[XFDG]

Summary

mega_mhc.cuh uses __nv_fp8_e4m3 in NormArgs, which is compiled by host g++ via csrc/python_api.cpp. CUTLASS gates cuda_fp8.h behind CUDA_FP8_ENABLED which requires __CUDACC_VER_* — defined only by nvcc, not g++ — so the type is invisible during host compilation.

Fix

Add #include <cuda_fp8.h> directly in mega_mhc.cuh before any other includes, so the type is visible to both nvcc and g++.

Testing

• Built successfully with bash ./develop.sh on B200 (SM100, CUDA 13.1, gcc-13)
• python3 -c "import deep_gemm; print(deep_gemm.__version__)" passes

Fixes #438

[ds-review-bot]

🤖 ds-review-bot Code Review

v6

直接包含 cuda_fp8.h 为主机编译提供了所需的 FP8 类型定义，符合项目 CUDA 版本要求，未发现该变更引入回归。当前环境缺少 CUDA 工具链，未实际验证构建。

v5

该 MR 在 deep_gemm/include/deep_gemm/layout/mega_mhc.cuh 顶部显式添加 #include <cuda_fp8.h>，修复主机端 g++ 编译 csrc/python_api.cpp 时 __nv_fp8_e4m3 未定义的问题。已核实包含链：python_api.cpp → apis/mega_mhc.hpp → jit_kernels/impls/sm100_mega_mhc.hpp → layout/mega_mhc.cuh，其中第 66 行 NormArgs::y_fp8 使用了 __nv_fp8_e4m3*。CUTLASS 的 cutlass/arch/barrier.h 依赖 CUDA_FP8_ENABLED（需 __CUDACC_VER_*，仅 nvcc 定义）才会间接引入 cuda_fp8.h，因此主机编译下该类型确实不可见。修复方式正确、最小、无副作用：cuda_fp8.h 自带 include guard，对 nvcc 路径无影响，且自 CUDA 11.8 起提供，与 SM100/CUDA 13 环境兼容。变更仅 2 行、1 个文件，与描述一致。建议合并。

v4

This MR adds an explicit #include <cuda_fp8.h> at the top of deep_gemm/include/deep_gemm/layout/mega_mhc.cuh, placed directly after #pragma once and before the CUTLASS/DeepGEMM includes. The header is pulled into host (g++) translation units such as csrc/jit_kernels/impls/sm100_mega_mhc.hpp, where NormArgs::y_fp8 is declared as __nv_fp8_e4m3*. CUTLASS only includes <cuda_fp8.h> when CUDA_FP8_ENABLED is defined, and that macro requires __CUDACC_VER_* (nvcc-only), so the type was invisible during host compilation and the build failed with __nv_fp8_e4m3 undefined. The fix is minimal, correct, and sufficient: it makes the type visible to both nvcc and g++, while being harmless for device builds because cuda_fp8.h has its own include guards. The nv_bfloat16 type used in the same struct already remains available unconditionally through cutlass/bfloat16.h -> <cuda_bf16.h>, so no additional include is needed, and placing this include before the CUTLASS headers avoids any include-ordering dependency.

Review outcome: approved. No defects, regressions, or required changes were identified.

Files reviewed: 1
Issues found: 🔵 1 suggestion
Inline comments posted: 1

[XFDG]

Resolved the remaining bot thread after rechecking it: the comment explicitly marks the repository-wide header scan as optional and out of scope for this focused fix. The PR remains the minimal one-header correction, and the bot review found no regression. No code change was needed.