Merge branch 'ollama:main' into main

llama/make/Makefile.rocm

@@ -58,6 +58,8 @@ endif
 GPU_COMPILER_CUFLAGS = \
 	$(GPU_COMPILER_FPIC) \
 	$(addprefix -m,$(GPU_RUNNER_CPU_FLAGS)) \
+	-mf16c \
+	-mfma \
 	-parallel-jobs=2 \
 	-c \
 	-O3 \
@@ -77,6 +79,9 @@ GPU_COMPILER_CUFLAGS = \
 	-D_CRT_SECURE_NO_WARNINGS \
 	-D_GNU_SOURCE \
 	-D_XOPEN_SOURCE=600 \
+	-DUSE_PROF_API=1 \
+	-std=gnu++14 \
+	-x hip \
 	-mllvm=-amdgpu-early-inline-all=true \
 	-mllvm=-amdgpu-function-calls=false \
 	-Wno-expansion-to-defined \
@@ -87,6 +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

llama/make/gpu.make

@@ -85,7 +85,7 @@ $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(RUNNERS
 	GOARCH=$(ARCH) CGO_LDFLAGS="$(TARGET_CGO_LDFLAGS)" go build -buildmode=pie  $(GPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(GPU_RUNNER_CPU_FLAGS) $(GPU_RUNNER_GO_TAGS)) -o $@ ./runner
 $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(GPU_RUNNER_OBJS) $(DIST_GPU_RUNNER_LIB_DEPS) $(COMMON_HDRS) $(GPU_RUNNER_HDRS)
 	@-mkdir -p $(dir $@)
-	$(CCACHE) $(GPU_COMPILER) --shared $(GPU_RUNNER_DRIVER_LIB_LINK) -L${DIST_GPU_RUNNER_DEPS_DIR} $(foreach lib, $(GPU_RUNNER_LIBS_SHORT), -l$(lib)) $(GPU_RUNNER_OBJS) -o $@
+	$(CCACHE) $(GPU_COMPILER) --shared -L$(GPU_LIB_DIR) $(GPU_RUNNER_DRIVER_LIB_LINK) -L${DIST_GPU_RUNNER_DEPS_DIR} $(foreach lib, $(GPU_RUNNER_LIBS_SHORT), -l$(lib)) $(GPU_RUNNER_OBJS) -o $@
 
 # Distribution targets
 $(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%