From 0bddd1be19e94cc4ca244ddbce39fdf95cea1899 Mon Sep 17 00:00:00 2001
From: Scott White <whitesscott@comcast.net>
Date: Wed, 1 Jul 2026 14:35:01 -0700
Subject: [PATCH] Use <cmath> instead of <math.h> in context attention FMHA
 headers

This replaces C-style <math.h> includes in the context attention FMHA code
with the C++ <cmath> header.

Including <math.h> from a public header is fragile in CUDA/TensorRT builds,
especially on Jetson/Thor where CUDA include paths are active and the C math
header may resolve through CUDA wrapper headers. Using <cmath> keeps the code
consistent with C++ compilation and avoids leaking C math-header behavior into
translation units that include fmhaParams_v2.h.

No runtime behavior change intended.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 cpp/kernels/contextAttentionKernels/contextFMHARunner.cpp | 2 +-
 cpp/kernels/contextAttentionKernels/fmhaParams_v2.h       | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/cpp/kernels/contextAttentionKernels/contextFMHARunner.cpp b/cpp/kernels/contextAttentionKernels/contextFMHARunner.cpp
index 186e3498..6978d19c 100644
--- a/cpp/kernels/contextAttentionKernels/contextFMHARunner.cpp
+++ b/cpp/kernels/contextAttentionKernels/contextFMHARunner.cpp
@@ -24,7 +24,7 @@
 #include <cstring>
 #include <cuda.h>
 #include <cuda_fp16.h>
-#include <math.h>
+#include <cmath>
 #include <memory>
 #include <mutex>
 #include <unordered_map>
diff --git a/cpp/kernels/contextAttentionKernels/fmhaParams_v2.h b/cpp/kernels/contextAttentionKernels/fmhaParams_v2.h
index 4252197d..72a26bdb 100644
--- a/cpp/kernels/contextAttentionKernels/fmhaParams_v2.h
+++ b/cpp/kernels/contextAttentionKernels/fmhaParams_v2.h
@@ -18,7 +18,7 @@
 #pragma once
 
 #include <limits.h>
-#include <math.h>
+#include <cmath>
 #include <stdint.h>
 
 //! \brief Parameters for ALiBi (Attention with Linear Biases) positional encoding