llm: New memory management

This changes the memory allocation strategy from upfront estimation to
tracking actual allocations done by the engine and reacting to that. The
goal is avoid issues caused by both under-estimation (crashing) and
over-estimation (low performance due to under-utilized GPUs).

It is currently opt-in and can be enabled for models running on the
Ollama engine by setting OLLAMA_NEW_ESTIMATES=1. Behavior in other
cases is unchanged and will continue to use the existing estimates.

llama/llama.go

@@ -62,6 +62,22 @@ func BackendInit() {
 	C.llama_backend_init()
 }
 
+func EnumerateGPUs() []string {
+	var ids []string
+
+	for i := range C.ggml_backend_dev_count() {
+		device := C.ggml_backend_dev_get(i)
+
+		if C.ggml_backend_dev_type(device) == C.GGML_BACKEND_DEVICE_TYPE_GPU {
+			var props C.struct_ggml_backend_dev_props
+			C.ggml_backend_dev_get_props(device, &props)
+			ids = append(ids, C.GoString(props.id))
+		}
+	}
+
+	return ids
+}
+
 func GetModelArch(modelPath string) (string, error) {
 	mp := C.CString(modelPath)
 	defer C.free(unsafe.Pointer(mp))

llama/patches/0016-temporary-prevent-rocm-cuda-mixed-loading.patch

@@ -1,32 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Daniel Hiltgen <daniel@ollama.com>
-Date: Sun, 22 Jun 2025 09:22:05 -0700
-Subject: [PATCH] temporary prevent rocm+cuda mixed loading
-
----
- ggml/src/ggml-backend-reg.cpp | 12 ++++++++++--
- 1 file changed, 10 insertions(+), 2 deletions(-)
-
-diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
-index 3040b2aa..f1e9c180 100644
---- a/ggml/src/ggml-backend-reg.cpp
-+++ b/ggml/src/ggml-backend-reg.cpp
-@@ -581,8 +581,16 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
- 
-     ggml_backend_load_best("blas", silent, dir_path);
-     ggml_backend_load_best("cann", silent, dir_path);
--    ggml_backend_load_best("cuda", silent, dir_path);
--    ggml_backend_load_best("hip", silent, dir_path);
-+
-+    // Avoid mixed hip+cuda configurations
-+    const char * hip_devices = std::getenv("HIP_VISIBLE_DEVICES");
-+    const char * rocr_devices = std::getenv("ROCR_VISIBLE_DEVICES");
-+    if (!hip_devices && !rocr_devices) {
-+        ggml_backend_load_best("cuda", silent, dir_path);
-+    } else {
-+        ggml_backend_load_best("hip", silent, dir_path);
-+    }
-+
-     ggml_backend_load_best("metal", silent, dir_path);
-     ggml_backend_load_best("rpc", silent, dir_path);
-     ggml_backend_load_best("sycl", silent, dir_path);