llm: Remove GGML_CUDA_NO_PEER_COPY for ROCm (#7174)

This workaround logic in llama.cpp is causing crashes for users with less system memory than VRAM.
2025-12-15 18:27:08 +00:00 · 2024-10-12 09:56:49 -07:00
parent 7fe3902552
commit c3d321d405
2 changed files with 1 additions and 2 deletions
--- a/llm/generate/gen_windows.ps1
+++ b/llm/generate/gen_windows.ps1
@@ -340,7 +340,6 @@ function build_rocm() {
            "-DCMAKE_C_COMPILER=clang.exe",
            "-DCMAKE_CXX_COMPILER=clang++.exe",
            "-DGGML_HIPBLAS=on",
-            "-DGGML_CUDA_NO_PEER_COPY=on",
            "-DHIP_PLATFORM=amd",
            "-DGGML_AVX=on",
            "-DGGML_AVX2=off",