ggml: Avoid allocating CUDA primary context on unused GPUs

The recent memory management changes caused all GPUs to be visible
to the runner, regardless of whether they are ultimately used. This
caused CUDA devices to allocate a primary context (~300 MB VRAM) on
each GPU, for each model. This is unnecessary, so we can both avoid
touching GPUs that we exclude in the early stage of allocation and
freeing the memory for any that we touch but don't use.

The issue will continue to exist for the old engine, since it touches
all devices during initialization.

llama/patches/0024-ggml-Enable-resetting-backend-devices.patch

+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Jesse Gross <jesse@ollama.com>
+Date: Wed, 27 Aug 2025 14:39:48 -0700
+Subject: [PATCH] ggml: Enable resetting backend devices
+
+Touching a CUDA device causes the allocation of a primary context
+with CUDA data structures (~300 MB of VRAM). If a device is
+unused then it can be reset to free these data structures.
+---
+ ggml/include/ggml-backend.h      |  1 +
+ ggml/src/ggml-backend-impl.h     |  4 ++++
+ ggml/src/ggml-backend.cpp        |  8 ++++++++
+ ggml/src/ggml-cuda/ggml-cuda.cu  | 17 +++++++++++++++--
+ ggml/src/ggml-cuda/vendors/hip.h |  1 +
+ 5 files changed, 29 insertions(+), 2 deletions(-)
+
+diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
+index b602a7c78..fda5ceb24 100644
+--- a/ggml/include/ggml-backend.h
++++ b/ggml/include/ggml-backend.h
+@@ -167,6 +167,7 @@ extern "C" {
+     GGML_API void                          ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props);
+     GGML_API ggml_backend_reg_t            ggml_backend_dev_backend_reg(ggml_backend_dev_t device);
+     GGML_API ggml_backend_t                ggml_backend_dev_init(ggml_backend_dev_t device, const char * params);
++    GGML_API void                          ggml_backend_dev_reset(ggml_backend_dev_t device);
+     GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_buffer_type(ggml_backend_dev_t device);
+     GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device);
+     GGML_API ggml_backend_buffer_t         ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size);
+diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h
+index 81749a5a3..6f10c353b 100644
+--- a/ggml/src/ggml-backend-impl.h
++++ b/ggml/src/ggml-backend-impl.h
+@@ -178,6 +178,10 @@ extern "C" {
+         ggml_backend_event_t (*event_new)         (ggml_backend_dev_t dev);
+         void                 (*event_free)        (ggml_backend_dev_t dev, ggml_backend_event_t event);
+         void                 (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
++
++        // (optional) reset device, clearing existing allocations and context
++        // the caller must ensure that there are no outstanding buffers, as these will become invalid
++        void (*reset)(ggml_backend_dev_t dev);
+     };
+ 
+     struct ggml_backend_device {
+diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
+index 05a842ed5..6556943b0 100644
+--- a/ggml/src/ggml-backend.cpp
++++ b/ggml/src/ggml-backend.cpp
+@@ -477,6 +477,14 @@ ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * par
+     return device->iface.init_backend(device, params);
+ }
+ 
++void ggml_backend_dev_reset(ggml_backend_dev_t device) {
++    if (device->iface.reset == NULL) {
++        return;
++    }
++
++    device->iface.reset(device);
++}
++
+ ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) {
+     return device->iface.get_buffer_type(device);
+ }
+diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
+index c7f9dc3a5..e43fde523 100644
+--- a/ggml/src/ggml-cuda/ggml-cuda.cu
++++ b/ggml/src/ggml-cuda/ggml-cuda.cu
+@@ -103,6 +103,11 @@ int ggml_cuda_get_device() {
+     return id;
+ }
+ 
++void ggml_cuda_reset_device(int device) {
++    ggml_cuda_set_device(device);
++    CUDA_CHECK(cudaDeviceReset());
++}
++
+ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
+     ggml_cuda_set_device(device);
+     cudaError_t err;
+@@ -3243,7 +3248,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
+     props->description = ggml_backend_cuda_device_get_description(dev);
+     props->id          = ggml_backend_cuda_device_get_id(dev);
+     props->type        = ggml_backend_cuda_device_get_type(dev);
+-    ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
++
++    // Memory reporting is disabled to avoid allocation of a CUDA primary context (~300 MB per device).
++    // If you need the memory data, call ggml_backend_dev_memory() explicitly.
++    props->memory_total = props->memory_free = 0;
+ 
+     bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
+ #ifdef GGML_CUDA_NO_PEER_COPY
+@@ -3700,6 +3708,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
+     CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
+ }
+ 
++static void ggml_backend_cuda_device_reset(ggml_backend_dev_t dev) {
++    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
++    ggml_cuda_reset_device(ctx->device);
++}
++
+ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
+     /* .get_name                = */ ggml_backend_cuda_device_get_name,
+     /* .get_description         = */ ggml_backend_cuda_device_get_description,
+@@ -3716,6 +3729,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
+     /* .event_new               = */ ggml_backend_cuda_device_event_new,
+     /* .event_free              = */ ggml_backend_cuda_device_event_free,
+     /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
++    /* .reset                   = */ ggml_backend_cuda_device_reset,
+ };
+ 
+ // backend reg
+@@ -3835,7 +3849,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+                 dev_ctx->device = i;
+                 dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
+ 
+-                ggml_cuda_set_device(i);
+                 cudaDeviceProp prop;
+                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
+                 dev_ctx->description = prop.name;
+diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
+index c31f31923..cf22e60d2 100644
+--- a/ggml/src/ggml-cuda/vendors/hip.h
++++ b/ggml/src/ggml-cuda/vendors/hip.h
+@@ -40,6 +40,7 @@
+ #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
+ #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
+ #define cudaDeviceProp hipDeviceProp_t
++#define cudaDeviceReset hipDeviceReset
+ #define cudaDeviceSynchronize hipDeviceSynchronize
+ #define cudaError_t hipError_t
+ #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled

ml/backend/ggml/ggml.go

@@ -629,6 +629,18 @@ func (b *Backend) Load(ctx context.Context, progress func(float32)) error {
 		})
 	}
 
+	// Cleanup any backend state from devices that we didn't end up using
+nextDevice:
+	for _, d := range append(gpus, append(accels, cpus...)...) {
+		for _, backend := range b.schedBackends {
+			if d == C.ggml_backend_get_device(backend) {
+				continue nextDevice
+			}
+		}
+
+		C.ggml_backend_dev_reset(d)
+	}
+
 	if err := g.Wait(); err != nil {
 		return err
 	}

ml/backend/ggml/ggml/include/ggml-backend.h

@@ -167,6 +167,7 @@ extern "C" {
     GGML_API void                          ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props);
     GGML_API ggml_backend_reg_t            ggml_backend_dev_backend_reg(ggml_backend_dev_t device);
     GGML_API ggml_backend_t                ggml_backend_dev_init(ggml_backend_dev_t device, const char * params);
+    GGML_API void                          ggml_backend_dev_reset(ggml_backend_dev_t device);
     GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_buffer_type(ggml_backend_dev_t device);
     GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device);
     GGML_API ggml_backend_buffer_t         ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size);

ml/backend/ggml/ggml/src/ggml-backend-impl.h

@@ -178,6 +178,10 @@ extern "C" {
         ggml_backend_event_t (*event_new)         (ggml_backend_dev_t dev);
         void                 (*event_free)        (ggml_backend_dev_t dev, ggml_backend_event_t event);
         void                 (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
+
+        // (optional) reset device, clearing existing allocations and context
+        // the caller must ensure that there are no outstanding buffers, as these will become invalid
+        void (*reset)(ggml_backend_dev_t dev);
     };
 
     struct ggml_backend_device {

ml/backend/ggml/ggml/src/ggml-backend.cpp

@@ -477,6 +477,14 @@ ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * par
     return device->iface.init_backend(device, params);
 }
 
+void ggml_backend_dev_reset(ggml_backend_dev_t device) {
+    if (device->iface.reset == NULL) {
+        return;
+    }
+
+    device->iface.reset(device);
+}
+
 ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) {
     return device->iface.get_buffer_type(device);
 }

ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -103,6 +103,11 @@ int ggml_cuda_get_device() {
     return id;
 }
 
+void ggml_cuda_reset_device(int device) {
+    ggml_cuda_set_device(device);
+    CUDA_CHECK(cudaDeviceReset());
+}
+
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
     cudaError_t err;
@@ -3243,7 +3248,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
     props->description = ggml_backend_cuda_device_get_description(dev);
     props->id          = ggml_backend_cuda_device_get_id(dev);
     props->type        = ggml_backend_cuda_device_get_type(dev);
-    ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    // Memory reporting is disabled to avoid allocation of a CUDA primary context (~300 MB per device).
+    // If you need the memory data, call ggml_backend_dev_memory() explicitly.
+    props->memory_total = props->memory_free = 0;
 
     bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
 #ifdef GGML_CUDA_NO_PEER_COPY
@@ -3700,6 +3708,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
     CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
 }
 
+static void ggml_backend_cuda_device_reset(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    ggml_cuda_reset_device(ctx->device);
+}
+
 static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
     /* .get_name                = */ ggml_backend_cuda_device_get_name,
     /* .get_description         = */ ggml_backend_cuda_device_get_description,
@@ -3716,6 +3729,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
     /* .event_new               = */ ggml_backend_cuda_device_event_new,
     /* .event_free              = */ ggml_backend_cuda_device_event_free,
     /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
+    /* .reset                   = */ ggml_backend_cuda_device_reset,
 };
 
 // backend reg
@@ -3835,7 +3849,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 dev_ctx->device = i;
                 dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
 
-                ggml_cuda_set_device(i);
                 cudaDeviceProp prop;
                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                 dev_ctx->description = prop.name;

ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h

@@ -40,6 +40,7 @@
 #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
 #define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceReset hipDeviceReset
 #define cudaDeviceSynchronize hipDeviceSynchronize
 #define cudaError_t hipError_t
 #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled