From 8744ef1155ce686b7de50e081668d1d03a04524c Mon Sep 17 00:00:00 2001
From: Mark Saroufim <marksaroufim@gmail.com>
Date: Mon, 20 Jul 2026 11:28:52 -0700
Subject: [PATCH] Document MathDx load_inline flags

[skip ci]
---
 src/runners/modal_runner.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/src/runners/modal_runner.py b/src/runners/modal_runner.py
index 5c07faf3..4520bf62 100644
--- a/src/runners/modal_runner.py
+++ b/src/runners/modal_runner.py
@@ -33,10 +33,11 @@
 #        cd src/runners && modal run test_cutlass_image.py
 #
 # For users writing submissions with torch.utils.cpp_extension.load_inline:
-#   C++ headers installed on the image (like CUTLASS) require explicit include paths:
+#   MathDx is already on CPLUS_INCLUDE_PATH. Include cuBLASDx in cuda_sources,
+#   which load_inline compiles with nvcc, rather than in cpp_sources:
 #     load_inline(
 #         ...
-#         extra_include_paths=["/opt/cutlass/include", "/opt/cutlass/tools/util/include"],
+#         cuda_sources=cuda_source,  # #include <cublasdx.hpp> goes here
 #     )
 #   For raw nvcc compilation, CPLUS_INCLUDE_PATH is set so includes work automatically.
 #