[Bug]: M_LOG2E availability depends on build configuration in the FlashMLA extension #53935

Your current environment

Environment

• Windows Server 2025
• Visual Studio Build Tools with MSVC
• CUDA/NVCC 13.2
• Python 3.12
• PyTorch 2.11.0+cu130
• SystemPanic/vllm-windows 0.25-based deployment
• DeepSeek-V4-Flash-0731
• 4 x H200 NVL GPUs

🐛 Describe the bug

Background

I encountered this issue while building the vllm-project/FlashMLA fork through SystemPanic/vllm-windows for a deployment on H200 NVL GPUs.

Problem

The vLLM-specific FlashMLA extension uses M_LOG2E in:

csrc/extension/sm90/dense_fp8/softmax.h
csrc/extension/sm90/dense_fp8/pybind.cpp

For example:

const float max_scaled = max(mi) == -INFINITY ? 0.f
                         : max(mi) * (
                               Scale_max ? scale : float(M_LOG2E));

M_LOG2E represents log2(e) and is commonly available in GNU/Linux build environments. Its availability is nevertheless implementation-specific.

With MSVC, the M_* constants are normally exposed only when _USE_MATH_DEFINES is defined before the relevant math header is included. The current source can therefore depend on build flags and translation-unit include order when compiled with NVCC and MSVC.

I understand that native Windows with NVCC and MSVC may not be part of the supported FlashMLA build matrix. However, removing this dependency would be a small cross-platform portability improvement and would not change the existing Linux result.

Tested source-level fix

The FlashMLA extension built successfully after replacing float(M_LOG2E) with the equivalent float literal:

- float(M_LOG2E)
+ 1.4426950408889634f

For the primary file, the tested change was:

diff --git a/csrc/extension/sm90/dense_fp8/softmax.h b/csrc/extension/sm90/dense_fp8/softmax.h
--- a/csrc/extension/sm90/dense_fp8/softmax.h
+++ b/csrc/extension/sm90/dense_fp8/softmax.h
@@
-const float max_scaled = max(mi) == -INFINITY ? 0.f : max(mi) * (Scale_max ? scale : float(M_LOG2E));
+const float max_scaled = max(mi) == -INFINITY ? 0.f : max(mi) * (Scale_max ? scale : 1.4426950408889634f);

The original expression explicitly converts M_LOG2E to a 32-bit float. The replacement literal is also a float because of the f suffix.

Both expressions round to the same IEEE-754 float32 value, so the tested change preserves the numerical result and does not alter the algorithm.

A project-owned named constant could avoid repeating the literal, for example:

inline constexpr float kLog2E = 1.4426950408889634f;

Alternative build-level workaround

The issue can also be avoided in the vLLM CMake integration by defining _USE_MATH_DEFINES for both FlashMLA extension targets:

target_compile_definitions(
    _flashmla_C PRIVATE _USE_MATH_DEFINES=1
)

target_compile_definitions(
    _flashmla_extension_C PRIVATE _USE_MATH_DEFINES=1
)

This is a smaller integration-specific workaround and keeps the existing source unchanged.

Related shared-source occurrences

The same dependency on M_LOG2E also appears in source inherited from deepseek-ai/FlashMLA, including:

csrc/api/dense_decode.h
csrc/sm90/decode/dense/splitkv_mla.cuh
csrc/sm90/decode/sparse_fp8/splitkv_mla.cuh
csrc/smxx/decode/combine/combine.cu
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.