From 29cad6bea8118b940ca919ebf5452d4ff661d593 Mon Sep 17 00:00:00 2001
From: Gentoo <gentoo@localhost>
Date: Fri, 4 Sep 2026 02:00:00 +0200
Subject: [PATCH] Fix build with NVTX disabled

Several translation units include nvtx.h directly after core.h has already
selected nvtx_stub.h.  When NVTX_NO_IMPL is defined this declares nccl_domain
twice.  Make nvtx.h honor the same switch as core.h, and provide the missing
no-op NCCL_NVTX3_FUNC_RANGE macro in the stub.
---
 src/include/nvtx.h      | 4 ++++
 src/include/nvtx_stub.h | 1 +
 2 files changed, 5 insertions(+)

diff --git a/src/include/nvtx.h b/src/include/nvtx.h
index 36182f5..b34fc0c 100644
--- a/src/include/nvtx.h
+++ b/src/include/nvtx.h
@@ -8,6 +8,9 @@
 #ifndef NCCL_NVTX_H_
 #define NCCL_NVTX_H_
 
+#ifdef NVTX_NO_IMPL
+#include "nvtx_stub.h"
+#else
 #include "nvtx3/nvtx3.hpp"
 #include "roctx.h"
 
@@ -215,4 +218,5 @@ class ncclNvtxRange {
 
 extern void initNvtxRegisteredEnums();
 
+#endif  // NVTX_NO_IMPL
 #endif
diff --git a/src/include/nvtx_stub.h b/src/include/nvtx_stub.h
index 0fe8ae5..c6b0f0f 100644
--- a/src/include/nvtx_stub.h
+++ b/src/include/nvtx_stub.h
@@ -13,6 +13,7 @@ struct nccl_domain {
 };
 
 #define NVTX3_FUNC_RANGE_IN(domain)
+#define NCCL_NVTX3_FUNC_RANGE
 #define nvtxNameOsThreadA(syscall, thread)
 #define NVTX3_FUNC_WITH_PARAMS(N, T, P)
 #define NVTX3_PAYLOAD(...) __VA_ARGS__
-- 
2.51.0