Workaround broken ROCm p2p copy

Enable the build flag for llama.cpp to use CPU copy for multi-GPU scenarios.

llm/generate/gen_linux.sh

@@ -254,7 +254,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
     fi
     init_vars
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DLLAMA_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""

llm/generate/gen_windows.ps1

@@ -366,6 +366,7 @@ function build_rocm() {
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
             "-DGGML_HIPBLAS=on",
+            "-DLLAMA_CUDA_NO_PEER_COPY=on",
             "-DHIP_PLATFORM=amd",
             "-DGGML_AVX=on",
             "-DGGML_AVX2=off",