From 1618700c5a042ecedb3fe3a93d7c90d3b4bd3001 Mon Sep 17 00:00:00 2001
From: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
Date: Thu, 7 Nov 2024 14:26:31 -0800
Subject: [PATCH] Workaround buggy P2P ROCm copy on windows (#7466)

This enables the workaround code only for windows which should help windows users with muliple AMD GPUs
---
 llama/make/Makefile.rocm | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/llama/make/Makefile.rocm b/llama/make/Makefile.rocm
index 136ce015..4ab176b4 100644
--- a/llama/make/Makefile.rocm
+++ b/llama/make/Makefile.rocm
@@ -92,6 +92,12 @@ GPU_COMPILER_CUFLAGS = \
 	-Wno-unused-result \
 	-I.
 
+# Workaround buggy P2P copy on some windows multi-GPU setups
+# This workaround breaks linux systems with small system RAM, so only enable on windows
+ifeq ($(OS),windows)
+	GPU_COMPILER_CUFLAGS += -DGGML_CUDA_NO_PEER_COPY=1
+endif
+
 include make/gpu.make
 
 # Adjust the rules from gpu.make to handle the ROCm dependencies properly