diff --git a/app/ollama.iss b/app/ollama.iss
index 7362eeeb..d5940be6 100644
--- a/app/ollama.iss
+++ b/app/ollama.iss
@@ -136,7 +136,7 @@ Type: filesandordirs; Name: "{%TEMP}\ollama*"
 Type: filesandordirs; Name: "{%LOCALAPPDATA}\Programs\Ollama"
 
 [Messages]
-WizardReady=Ollama Windows Preview
+WizardReady=Ollama
 ReadyLabel1=%nLet's get you up and running with your own large language models.
 SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or finish the other installer, then click OK to continue with this install, or Cancel to exit.
 
diff --git a/discover/gpu_info_nvcuda.c b/discover/gpu_info_nvcuda.c
index a1a38bfc..466e1ac2 100644
--- a/discover/gpu_info_nvcuda.c
+++ b/discover/gpu_info_nvcuda.c
@@ -4,6 +4,7 @@
 #include "gpu_info_nvcuda.h"
 
 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
+  LOG(resp->ch.verbose, "initializing %s\n", nvcuda_lib_path);
   CUresult ret;
   resp->err = NULL;
   resp->num_devices = 0;
@@ -57,8 +58,10 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
       resp->cudaErr = -1;
       return;
     }
+    LOG(resp->ch.verbose, "dlsym: %s - %p\n", l[i].s, *l[i].p);
   }
 
+  LOG(resp->ch.verbose, "calling cuInit\n");
   ret = (*resp->ch.cuInit)(0);
   if (ret != CUDA_SUCCESS) {
     LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
@@ -75,15 +78,18 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
   resp->ch.driver_minor = 0;
 
   // Report driver version if we're in verbose mode, ignore errors
+  LOG(resp->ch.verbose, "calling cuDriverGetVersion\n");
   ret = (*resp->ch.cuDriverGetVersion)(&version);
   if (ret != CUDA_SUCCESS) {
     LOG(resp->ch.verbose, "cuDriverGetVersion failed: %d\n", ret);
   } else {
+    LOG(resp->ch.verbose, "raw version 0x%x\n", version);
     resp->ch.driver_major = version / 1000;
     resp->ch.driver_minor = (version - (resp->ch.driver_major * 1000)) / 10;
     LOG(resp->ch.verbose, "CUDA driver version: %d.%d\n", resp->ch.driver_major, resp->ch.driver_minor);
   }
 
+  LOG(resp->ch.verbose, "calling cuDeviceGetCount\n");
   ret = (*resp->ch.cuDeviceGetCount)(&resp->num_devices);
   if (ret != CUDA_SUCCESS) {
     LOG(resp->ch.verbose, "cuDeviceGetCount err: %d\n", ret);
@@ -94,6 +100,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     resp->cudaErr = ret;
     return;
   }
+  LOG(resp->ch.verbose, "device count %d\n", resp->num_devices);
 }
 
 const int buflen = 256;
diff --git a/docs/development.md b/docs/development.md
index f2039a08..13457ae3 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -108,7 +108,7 @@ Custom CPU settings are not currently supported in the new Go server build but w
 
 #### Containerized Linux Build
 
-If you have Docker available, you can build linux binaries with `OLLAMA_NEW_RUNNERS=1 ./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
+If you have Docker available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
 
 ### Windows
 
diff --git a/llama/make/Makefile.rocm b/llama/make/Makefile.rocm
index 947c43a6..4ab176b4 100644
--- a/llama/make/Makefile.rocm
+++ b/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
diff --git a/llama/make/gpu.make b/llama/make/gpu.make
index 939fa41a..fbd8dbca 100644
--- a/llama/make/gpu.make
+++ b/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)/%