vulkan: Get FilterID from Backend for Vulkan (#12655)

* vulkan: Get FilterID from Backend for Vulkan

* Fixing patch

discover/runner.go

@@ -550,13 +550,6 @@ func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []s
 	}
 	logutil.Trace("runner enumerated devices", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "devices", devices)
 
-	// Enumerate returned devices starting at 0 per library and assign the per-library index as FilteredID
-	libCounts := make(map[string]int)
-	for i := range devices {
-		lib := devices[i].Library
-		devices[i].FilteredID = strconv.Itoa(libCounts[lib])
-		libCounts[lib]++
-	}
 	return devices
 }
 

llama/patches/0026-GPU-discovery-enhancements.patch

@@ -22,7 +22,7 @@ diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
 index ba181d09..09ff75f9 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
-@@ -169,6 +169,15 @@ extern "C" {
+@@ -169,6 +169,17 @@ extern "C" {
          const char * device_id;
          // device capabilities
          struct ggml_backend_dev_caps caps;
@@ -35,6 +35,8 @@ index ba181d09..09ff75f9 100644
 +        int pci_device_id;
 +        int pci_domain_id;
 +        const char *library;
++        // number with which the devices are accessed (Vulkan)
++        const char *numeric_id;
      };
  
      GGML_API const char *                  ggml_backend_dev_name(ggml_backend_dev_t device);

llama/patches/0028-vulkan-pci-and-memory.patch

@@ -175,7 +175,7 @@ index adea7783..fb7204ce 100644
      ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
      props->caps = {
          /* .async                 = */ false,
-@@ -12564,6 +12633,16 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
+@@ -12564,6 +12633,17 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
          /* .buffer_from_host_ptr  = */ false,
          /* .events                = */ false,
      };
@@ -189,6 +189,7 @@ index adea7783..fb7204ce 100644
 +    props->pci_device_id = ctx->pci_device_id;
 +    props->pci_domain_id = ctx->pci_domain_id;
 +    props->library = GGML_VK_NAME;
++    props->numeric_id = ctx->id.empty() ? nullptr : ctx->id.c_str();
  }
  
  static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {

ml/backend/ggml/ggml.go

@@ -726,6 +726,9 @@ func (b *Backend) BackendDevices() []ml.DeviceInfo {
 		}
 		info.PCIID = fmt.Sprintf("%02x:%02x.%x", props.pci_bus_id, props.pci_device_id, props.pci_domain_id)
 		info.LibraryPath = ggml.LibPaths()
+		if props.numeric_id != nil {
+			info.FilteredID = C.GoString(props.numeric_id)
+		}
 
 		C.ggml_backend_dev_memory(dev, &props.memory_free, &props.memory_total)
 		info.TotalMemory = (uint64)(props.memory_total)

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

@@ -178,6 +178,8 @@ extern "C" {
         int pci_device_id;
         int pci_domain_id;
         const char *library;
+        // number with which the devices are accessed (Vulkan)
+        const char *numeric_id;
     };
 
     GGML_API const char *                  ggml_backend_dev_name(ggml_backend_dev_t device);

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

@@ -12640,6 +12640,7 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
     props->pci_device_id = ctx->pci_device_id;
     props->pci_domain_id = ctx->pci_domain_id;
     props->library = GGML_VK_NAME;
+    props->numeric_id = ctx->id.empty() ? nullptr : ctx->id.c_str();
 }
 static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {