Merge branch 'ollama:main' into main

2025-12-25 16:08:01 +00:00 · 2024-05-07 14:47:50 +08:00
parent 0d107ca31c 7c5330413b
commit 4bd9b97d7b
22 changed files with 610 additions and 163 deletions
--- a/README.md
+++ b/README.md
@@ -284,17 +284,19 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [OllamaGUI](https://github.com/enoch1118/ollamaGUI)
 - [OpenAOE](https://github.com/InternLM/OpenAOE)
 - [Odin Runes](https://github.com/leonid20000/OdinRunes)
- [LLM-X: Progressive Web App](https://github.com/mrdjohnson/llm-x)
+- [LLM-X](https://github.com/mrdjohnson/llm-x) (Progressive Web App)
 - [AnythingLLM (Docker + MacOs/Windows/Linux native app)](https://github.com/Mintplex-Labs/anything-llm)
 - [Ollama Basic Chat: Uses HyperDiv Reactive UI](https://github.com/rapidarchitect/ollama_basic_chat)
 - [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
- [QA-Pilot: Chat with Code Repository](https://github.com/reid41/QA-Pilot)
- [ChatOllama: Open Source Chatbot based on Ollama with Knowledge Bases](https://github.com/sugarforever/chat-ollama)
- [CRAG Ollama Chat: Simple Web Search with Corrective RAG](https://github.com/Nagi-ovo/CRAG-Ollama-Chat)
- [RAGFlow: Open-source Retrieval-Augmented Generation engine based on deep document understanding](https://github.com/infiniflow/ragflow)
- [chat: chat web app for teams](https://github.com/swuecho/chat)
+- [QA-Pilot](https://github.com/reid41/QA-Pilot) (Chat with Code Repository)
+- [ChatOllama](https://github.com/sugarforever/chat-ollama) (Open Source Chatbot based on Ollama with Knowledge Bases)
+- [CRAG Ollama Chat](https://github.com/Nagi-ovo/CRAG-Ollama-Chat) (Simple Web Search with Corrective RAG)
+- [RAGFlow](https://github.com/infiniflow/ragflow) (Open-source Retrieval-Augmented Generation engine based on deep document understanding)
+- [StreamDeploy](https://github.com/StreamDeploy-DevRel/streamdeploy-llm-app-scaffold) (LLM Application Scaffold)
+- [chat](https://github.com/swuecho/chat) (chat web app for teams)
 - [Lobe Chat](https://github.com/lobehub/lobe-chat) with [Integrating Doc](https://lobehub.com/docs/self-hosting/examples/ollama)
- [Ollama RAG Chatbot: Local Chat with multiple PDFs using Ollama and RAG.](https://github.com/datvodinh/rag-chatbot.git)
+- [Ollama RAG Chatbot](https://github.com/datvodinh/rag-chatbot.git) (Local Chat with multiple PDFs using Ollama and RAG)
+- [BrainSoup](https://www.nurgo-software.com/products/brainsoup) (Flexible native client with RAG & multi-agent automation)

 ### Terminal

@@ -351,6 +353,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama-ex for Elixir](https://github.com/lebrunel/ollama-ex)
 - [Ollama Connector for SAP ABAP](https://github.com/b-tocs/abap_btocs_ollama)
 - [Testcontainers](https://testcontainers.com/modules/ollama/)
+- [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)

 ### Mobile

--- a/api/types.go
+++ b/api/types.go
@@ -436,6 +436,13 @@ type Duration struct {
 	time.Duration
 }

+func (d Duration) MarshalJSON() ([]byte, error) {
+	if d.Duration < 0 {
+		return []byte("-1"), nil
+	}
+	return []byte("\"" + d.Duration.String() + "\""), nil
+}
+
 func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 	var v any
 	if err := json.Unmarshal(b, &v); err != nil {
@@ -449,7 +456,7 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 		if t < 0 {
 			d.Duration = time.Duration(math.MaxInt64)
 		} else {
-			d.Duration = time.Duration(t * float64(time.Second))
+			d.Duration = time.Duration(int(t) * int(time.Second))
 		}
 	case string:
 		d.Duration, err = time.ParseDuration(t)
@@ -459,6 +466,8 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 		if d.Duration < 0 {
 			d.Duration = time.Duration(math.MaxInt64)
 		}
+	default:
+		return fmt.Errorf("Unsupported type: '%s'", reflect.TypeOf(v))
 	}

 	return nil
--- a/api/types_test.go
+++ b/api/types_test.go
@@ -21,6 +21,11 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 			req:  `{ "keep_alive": 42 }`,
 			exp:  &Duration{42 * time.Second},
 		},
+		{
+			name: "Positive Float",
+			req:  `{ "keep_alive": 42.5 }`,
+			exp:  &Duration{42 * time.Second},
+		},
 		{
 			name: "Positive Integer String",
 			req:  `{ "keep_alive": "42m" }`,
@@ -31,6 +36,11 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 			req:  `{ "keep_alive": -1 }`,
 			exp:  &Duration{math.MaxInt64},
 		},
+		{
+			name: "Negative Float",
+			req:  `{ "keep_alive": -3.14 }`,
+			exp:  &Duration{math.MaxInt64},
+		},
 		{
 			name: "Negative Integer String",
 			req:  `{ "keep_alive": "-1m" }`,
@@ -48,3 +58,50 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 		})
 	}
 }
+
+func TestDurationMarshalUnmarshal(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    time.Duration
+		expected time.Duration
+	}{
+		{
+			"negative duration",
+			time.Duration(-1),
+			time.Duration(math.MaxInt64),
+		},
+		{
+			"positive duration",
+			time.Duration(42 * time.Second),
+			time.Duration(42 * time.Second),
+		},
+		{
+			"another positive duration",
+			time.Duration(42 * time.Minute),
+			time.Duration(42 * time.Minute),
+		},
+		{
+			"zero duration",
+			time.Duration(0),
+			time.Duration(0),
+		},
+		{
+			"max duration",
+			time.Duration(math.MaxInt64),
+			time.Duration(math.MaxInt64),
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			b, err := json.Marshal(Duration{test.input})
+			require.NoError(t, err)
+
+			var d Duration
+			err = json.Unmarshal(b, &d)
+			require.NoError(t, err)
+
+			assert.Equal(t, test.expected, d.Duration, "input %v, marshalled %v, got %v", test.input, string(b), d.Duration)
+		})
+	}
+}
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -898,7 +898,12 @@ func RunServer(cmd *cobra.Command, _ []string) error {
 		return err
 	}

-	return server.Serve(ln)
+	err = server.Serve(ln)
+	if errors.Is(err, http.ErrServerClosed) {
+		return nil
+	}
+
+	return err
 }

 func initializeKeypair() error {
--- a/docs/api.md
+++ b/docs/api.md
@@ -95,7 +95,7 @@ The final response in the stream also includes additional data about the generat
 - `context`: an encoding of the conversation used in this response, this can be sent in the next request to keep a conversational memory
 - `response`: empty if the response was streamed, if not streamed, this will contain the full response

-To calculate how fast the response is generated in tokens per second (token/s), divide `eval_count` / `eval_duration`.
+To calculate how fast the response is generated in tokens per second (token/s), divide `eval_count` / `eval_duration` * `10^9`.

 ```json
 {
--- a/docs/faq.md
+++ b/docs/faq.md
@@ -140,7 +140,7 @@ Refer to the section [above](#how-do-i-configure-ollama-server) for how to set e

 - macOS: `~/.ollama/models`
 - Linux: `/usr/share/ollama/.ollama/models`
- Windows: `C:\Users\<username>\.ollama\models`
+- Windows: `C:\Users\%username%\.ollama\models`

 ### How do I set them to a different location?

--- a/docs/import.md
+++ b/docs/import.md
@@ -125,7 +125,7 @@ Publishing models is in early alpha. If you'd like to publish your model to shar

 1. Create [an account](https://ollama.com/signup)
 2. Copy your Ollama public key:
-  - macOS: `cat ~/.ollama/id_ed25519.pub`
+  - macOS: `cat ~/.ollama/id_ed25519.pub | pbcopy`
  - Windows: `type %USERPROFILE%\.ollama\id_ed25519.pub`
  - Linux: `cat /usr/share/ollama/.ollama/id_ed25519.pub`
 3. Add your public key to your [Ollama account](https://ollama.com/settings/keys)
@@ -136,6 +136,8 @@ Next, copy your model to your username's namespace:
 ollama cp example <your username>/example
 ```

+> Note: model names may only contain lowercase letters, digits, and the characters `.`, `-`, and `_`.
+
 Then push the model:

 ```
--- a/docs/linux.md
+++ b/docs/linux.md
@@ -105,7 +105,7 @@ sudo chmod +x /usr/bin/ollama
 To view logs of Ollama running as a startup service, run:

 ```bash
-journalctl -u ollama
+journalctl -e -u ollama
 ```

 ## Uninstall
--- a/docs/windows.md
+++ b/docs/windows.md
@@ -1,4 +1,3 @@
-<<<<<<< HEAD
 # Ollama Windows Preview

 Welcome to the Ollama Windows preview.
@@ -59,4 +58,4 @@ If you'd like to install or integrate Ollama as a service, a standalone
 `ollama-windows-amd64.zip` zip file is available containing only the Ollama CLI
 and GPU library dependencies for Nvidia and AMD. This allows for embedding
 Ollama in existing applications, or running it as a system service via `ollama
-serve` with tools such as [NSSM](https://nssm.cc/).
+serve` with tools such as [NSSM](https://nssm.cc/).
--- a/gpu/amd_common.go
+++ b/gpu/amd_common.go
@@ -81,8 +81,10 @@ func commonAMDValidateLibDir() (string, error) {
 	}

 	// Well known location(s)
-	if rocmLibUsable(RocmStandardLocation) {
-		return RocmStandardLocation, nil
+	for _, path := range RocmStandardLocations {
+		if rocmLibUsable(path) {
+			return path, nil
+		}
 	}

 	// Installer payload location if we're running the installed binary
--- a/gpu/amd_linux.go
+++ b/gpu/amd_linux.go
@@ -25,12 +25,12 @@ const (
 	// Prefix with the node dir
 	GPUTotalMemoryFileGlob = "mem_banks/*/properties" // size_in_bytes line
 	GPUUsedMemoryFileGlob  = "mem_banks/*/used_memory"
-	RocmStandardLocation   = "/opt/rocm/lib"
 )

 var (
 	// Used to validate if the given ROCm lib is usable
-	ROCmLibGlobs = []string{"libhipblas.so.2*", "rocblas"} // TODO - probably include more coverage of files here...
+	ROCmLibGlobs          = []string{"libhipblas.so.2*", "rocblas"} // TODO - probably include more coverage of files here...
+	RocmStandardLocations = []string{"/opt/rocm/lib", "/usr/lib64"}
 )

 // Gather GPU information from the amdgpu driver if any supported GPUs are detected
--- a/gpu/amd_windows.go
+++ b/gpu/amd_windows.go
@@ -14,7 +14,6 @@ import (
 )

 const (
-	RocmStandardLocation = "C:\\Program Files\\AMD\\ROCm\\5.7\\bin" // TODO glob?

 	// TODO  We're lookinng for this exact name to detect iGPUs since hipGetDeviceProperties never reports integrated==true
 	iGPUName = "AMD Radeon(TM) Graphics"
@@ -22,7 +21,8 @@ const (

 var (
 	// Used to validate if the given ROCm lib is usable
-	ROCmLibGlobs = []string{"hipblas.dll", "rocblas"} // TODO - probably include more coverage of files here...
+	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // TODO - probably include more coverage of files here...
+	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\5.7\\bin"} // TODO glob?
 )

 func AMDGetGPUInfo() []GpuInfo {
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -27,6 +27,7 @@ import (
 type handles struct {
 	deviceCount int
 	cudart      *C.cudart_handle_t
+	nvcuda      *C.nvcuda_handle_t
 }

 const (
@@ -63,6 +64,22 @@ var CudartWindowsGlobs = []string{
 	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
 }

+var NvcudaLinuxGlobs = []string{
+	"/usr/local/cuda*/targets/*/lib/libcuda.so*",
+	"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
+	"/usr/lib/*-linux-gnu/libcuda.so*",
+	"/usr/lib/wsl/lib/libcuda.so*",
+	"/usr/lib/wsl/drivers/*/libcuda.so*",
+	"/opt/cuda/lib*/libcuda.so*",
+	"/usr/local/cuda/lib*/libcuda.so*",
+	"/usr/lib*/libcuda.so*",
+	"/usr/local/lib*/libcuda.so*",
+}
+
+var NvcudaWindowsGlobs = []string{
+	"c:\\windows\\system*\\nvcuda.dll",
+}
+
 // Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
@@ -75,6 +92,8 @@ func initGPUHandles() *handles {
 	gpuHandles := &handles{}
 	var cudartMgmtName string
 	var cudartMgmtPatterns []string
+	var nvcudaMgmtName string
+	var nvcudaMgmtPatterns []string

 	tmpDir, _ := PayloadsDir()
 	switch runtime.GOOS {
@@ -83,6 +102,9 @@ func initGPUHandles() *handles {
 		localAppData := os.Getenv("LOCALAPPDATA")
 		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", cudartMgmtName)}
 		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartWindowsGlobs...)
+		// Aligned with driver, we can't carry as payloads
+		nvcudaMgmtName = "nvcuda.dll"
+		nvcudaMgmtPatterns = NvcudaWindowsGlobs
 	case "linux":
 		cudartMgmtName = "libcudart.so*"
 		if tmpDir != "" {
@@ -90,11 +112,25 @@ func initGPUHandles() *handles {
 			cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", cudartMgmtName)}
 		}
 		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartLinuxGlobs...)
+		// Aligned with driver, we can't carry as payloads
+		nvcudaMgmtName = "libcuda.so*"
+		nvcudaMgmtPatterns = NvcudaLinuxGlobs
 	default:
 		return gpuHandles
 	}

 	slog.Info("Detecting GPUs")
+	nvcudaLibPaths := FindGPULibs(nvcudaMgmtName, nvcudaMgmtPatterns)
+	if len(nvcudaLibPaths) > 0 {
+		deviceCount, nvcuda, libPath := LoadNVCUDAMgmt(nvcudaLibPaths)
+		if nvcuda != nil {
+			slog.Info("detected GPUs", "count", deviceCount, "library", libPath)
+			gpuHandles.nvcuda = nvcuda
+			gpuHandles.deviceCount = deviceCount
+			return gpuHandles
+		}
+	}
+
 	cudartLibPaths := FindGPULibs(cudartMgmtName, cudartMgmtPatterns)
 	if len(cudartLibPaths) > 0 {
 		deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
@@ -119,6 +155,9 @@ func GetGPUInfo() GpuInfoList {
 		if gpuHandles.cudart != nil {
 			C.cudart_release(*gpuHandles.cudart)
 		}
+		if gpuHandles.nvcuda != nil {
+			C.nvcuda_release(*gpuHandles.nvcuda)
+		}
 	}()

 	// All our GPU builds on x86 have AVX enabled, so fallback to CPU if we don't detect at least AVX
@@ -127,6 +166,12 @@ func GetGPUInfo() GpuInfoList {
 		slog.Warn("CPU does not have AVX or AVX2, disabling GPU support.")
 	}

+	// On windows we bundle the nvidia library one level above the runner dir
+	depPath := ""
+	if runtime.GOOS == "windows" && envconfig.RunnersDir != "" {
+		depPath = filepath.Dir(envconfig.RunnersDir)
+	}
+
 	var memInfo C.mem_info_t
 	resp := []GpuInfo{}

@@ -139,7 +184,11 @@ func GetGPUInfo() GpuInfoList {
 		gpuInfo := GpuInfo{
 			Library: "cuda",
 		}
-		C.cudart_check_vram(*gpuHandles.cudart, C.int(i), &memInfo)
+		if gpuHandles.cudart != nil {
+			C.cudart_check_vram(*gpuHandles.cudart, C.int(i), &memInfo)
+		} else {
+			C.nvcuda_check_vram(*gpuHandles.nvcuda, C.int(i), &memInfo)
+		}
 		if memInfo.err != nil {
 			slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
 			C.free(unsafe.Pointer(memInfo.err))
@@ -155,6 +204,7 @@ func GetGPUInfo() GpuInfoList {
 		gpuInfo.Major = int(memInfo.major)
 		gpuInfo.Minor = int(memInfo.minor)
 		gpuInfo.MinimumMemory = cudaMinimumMemory
+		gpuInfo.DependencyPath = depPath

 		// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
 		resp = append(resp, gpuInfo)
@@ -197,9 +247,10 @@ func GetCPUMem() (memInfo, error) {
 	return ret, nil
 }

-func FindGPULibs(baseLibName string, patterns []string) []string {
+func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
 	var ldPaths []string
+	var patterns []string
 	gpuLibPaths := []string{}
 	slog.Debug("Searching for GPU library", "name", baseLibName)

@@ -219,8 +270,14 @@ func FindGPULibs(baseLibName string, patterns []string) []string {
 		}
 		patterns = append(patterns, filepath.Join(d, baseLibName+"*"))
 	}
+	patterns = append(patterns, defaultPatterns...)
 	slog.Debug("gpu library search", "globs", patterns)
 	for _, pattern := range patterns {
+
+		// Nvidia PhysX known to return bogus results
+		if strings.Contains(pattern, "PhysX") {
+			slog.Debug("skipping PhysX cuda library path", "path", pattern)
+		}
 		// Ignore glob discovery errors
 		matches, _ := filepath.Glob(pattern)
 		for _, match := range matches {
@@ -268,6 +325,23 @@ func LoadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string) {
 	return 0, nil, ""
 }

+func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
+	var resp C.nvcuda_init_resp_t
+	resp.ch.verbose = getVerboseState()
+	for _, libPath := range nvcudaLibPaths {
+		lib := C.CString(libPath)
+		defer C.free(unsafe.Pointer(lib))
+		C.nvcuda_init(lib, &resp)
+		if resp.err != nil {
+			slog.Debug("Unable to load nvcuda", "library", libPath, "error", C.GoString(resp.err))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			return int(resp.num_devices), &resp.ch, libPath
+		}
+	}
+	return 0, nil, ""
+}
+
 func getVerboseState() C.uint16_t {
 	if envconfig.Debug {
 		return C.uint16_t(1)
--- a/gpu/gpu_info.h
+++ b/gpu/gpu_info.h
@@ -58,6 +58,7 @@ void cpu_check_ram(mem_info_t *resp);
 #endif

 #include "gpu_info_cudart.h"
+#include "gpu_info_nvcuda.h"

 #endif  // __GPU_INFO_H__
 #endif  // __APPLE__
--- a/gpu/gpu_info_cudart.h
+++ b/gpu/gpu_info_cudart.h
@@ -6,9 +6,9 @@
 // Just enough typedef's to dlopen/dlsym for memory information
 typedef enum cudartReturn_enum {
  CUDART_SUCCESS = 0,
-  CUDA_ERROR_INVALID_VALUE = 1,
-  CUDA_ERROR_MEMORY_ALLOCATION = 2,
-  CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
+  CUDART_ERROR_INVALID_VALUE = 1,
+  CUDART_ERROR_MEMORY_ALLOCATION = 2,
+  CUDART_ERROR_INSUFFICIENT_DRIVER = 35,
  // Other values omitted for now...
 } cudartReturn_t;

--- a/gpu/gpu_info_nvcuda.c
+++ b/gpu/gpu_info_nvcuda.c
@@ -0,0 +1,203 @@
+#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
+
+#include <string.h>
+#include "gpu_info_nvcuda.h"
+
+void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
+  CUresult ret;
+  resp->err = NULL;
+  resp->num_devices = 0;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  struct lookup {
+    char *s;
+    void **p;
+  } l[] = {
+   
+      {"cuInit", (void *)&resp->ch.cuInit},
+      {"cuDriverGetVersion", (void *)&resp->ch.cuDriverGetVersion},
+      {"cuDeviceGetCount", (void *)&resp->ch.cuDeviceGetCount},
+      {"cuDeviceGet", (void *)&resp->ch.cuDeviceGet},
+      {"cuDeviceGetAttribute", (void *)&resp->ch.cuDeviceGetAttribute},
+      {"cuDeviceGetUuid", (void *)&resp->ch.cuDeviceGetUuid},
+      {"cuCtxCreate_v3", (void *)&resp->ch.cuCtxCreate_v3},
+      {"cuMemGetInfo_v2", (void *)&resp->ch.cuMemGetInfo_v2},
+      {"cuCtxDestroy", (void *)&resp->ch.cuCtxDestroy},
+      {NULL, NULL},
+  };
+
+  resp->ch.handle = LOAD_LIBRARY(nvcuda_lib_path, RTLD_LAZY);
+  if (!resp->ch.handle) {
+    char *msg = LOAD_ERR();
+    LOG(resp->ch.verbose, "library %s load err: %s\n", nvcuda_lib_path, msg);
+    snprintf(buf, buflen,
+            "Unable to load %s library to query for Nvidia GPUs: %s",
+            nvcuda_lib_path, msg);
+    free(msg);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  for (i = 0; l[i].s != NULL; i++) {
+    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
+    if (!*l[i].p) {
+      char *msg = LOAD_ERR();
+      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
+      UNLOAD_LIBRARY(resp->ch.handle);
+      resp->ch.handle = NULL;
+      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
+              msg);
+      free(msg);
+      resp->err = strdup(buf);
+      return;
+    }
+  }
+
+  ret = (*resp->ch.cuInit)(0);
+  if (ret != CUDA_SUCCESS) {
+    LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->ch.handle);
+    resp->ch.handle = NULL;
+    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
+      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
+      return;
+    }
+    snprintf(buf, buflen, "nvcuda init failure: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  int version = 0;
+  nvcudaDriverVersion_t driverVersion;
+  driverVersion.major = 0;
+  driverVersion.minor = 0;
+
+  // Report driver version if we're in verbose mode, ignore errors
+  ret = (*resp->ch.cuDriverGetVersion)(&version);
+  if (ret != CUDA_SUCCESS) {
+    LOG(resp->ch.verbose, "cuDriverGetVersion failed: %d\n", ret);
+  } else {
+    driverVersion.major = version / 1000;
+    driverVersion.minor = (version - (driverVersion.major * 1000)) / 10;
+    LOG(resp->ch.verbose, "CUDA driver version: %d-%d\n", driverVersion.major, driverVersion.minor);
+  }
+
+  ret = (*resp->ch.cuDeviceGetCount)(&resp->num_devices);
+  if (ret != CUDA_SUCCESS) {
+    LOG(resp->ch.verbose, "cuDeviceGetCount err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->ch.handle);
+    resp->ch.handle = NULL;
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+}
+
+const int buflen = 256;
+void nvcuda_check_vram(nvcuda_handle_t h, int i, mem_info_t *resp) {
+  resp->err = NULL;
+  nvcudaMemory_t memInfo = {0,0};
+  CUresult ret;
+  CUdevice device = -1;
+  CUcontext ctx = NULL;
+  char buf[buflen + 1];
+  CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+
+  if (h.handle == NULL) {
+    resp->err = strdup("nvcuda handle isn't initialized");
+    return;
+  }
+
+  ret = (*h.cuDeviceGet)(&device, i);
+  if (ret != CUDA_SUCCESS) {
+    snprintf(buf, buflen, "nvcuda device failed to initialize");
+    resp->err = strdup(buf);
+    return;
+  }
+
+  resp->major = 0;
+  resp->minor = 0;
+  int major = 0;
+  int minor = 0;
+  ret = (*h.cuDeviceGetAttribute)(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device);
+  if (ret != CUDA_SUCCESS) {
+    LOG(h.verbose, "[%d] device major lookup failure: %d\n", i, ret);
+  } else {
+    ret = (*h.cuDeviceGetAttribute)(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device);
+    if (ret != CUDA_SUCCESS) {
+      LOG(h.verbose, "[%d] device minor lookup failure: %d\n", i, ret);
+    } else {
+      resp->minor = minor;  
+      resp->major = major;  
+    }
+  }
+
+  ret = (*h.cuDeviceGetUuid)(&uuid, device);
+  if (ret != CUDA_SUCCESS) {
+    LOG(h.verbose, "[%d] device uuid lookup failure: %d\n", i, ret);
+    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
+  } else {
+    // GPU-d110a105-ac29-1d54-7b49-9c90440f215b
+    snprintf(&resp->gpu_id[0], GPU_ID_LEN,
+        "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
+        uuid.bytes[0],
+        uuid.bytes[1],
+        uuid.bytes[2],
+        uuid.bytes[3],
+        uuid.bytes[4],
+        uuid.bytes[5],
+        uuid.bytes[6],
+        uuid.bytes[7],
+        uuid.bytes[8],
+        uuid.bytes[9],
+        uuid.bytes[10],
+        uuid.bytes[11],
+        uuid.bytes[12],
+        uuid.bytes[13],
+        uuid.bytes[14],
+        uuid.bytes[15]
+      );
+  }
+
+  // To get memory we have to set (and release) a context
+  ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
+  if (ret != CUDA_SUCCESS) {
+    snprintf(buf, buflen, "nvcuda failed to get primary device context %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
+  if (ret != CUDA_SUCCESS) {
+    snprintf(buf, buflen, "nvcuda device memory info lookup failure %d", ret);
+    resp->err = strdup(buf);
+    // Best effort on failure...
+    (*h.cuCtxDestroy)(ctx);
+    return;
+  }
+
+  resp->total = memInfo.total;
+  resp->free = memInfo.free;
+
+  LOG(h.verbose, "[%s] CUDA totalMem %lu mb\n", resp->gpu_id, resp->total / 1024 / 1024);
+  LOG(h.verbose, "[%s] CUDA freeMem %lu mb\n", resp->gpu_id, resp->free / 1024 / 1024);
+  LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
+
+  
+
+  ret = (*h.cuCtxDestroy)(ctx);
+  if (ret != CUDA_SUCCESS) {
+    LOG(1, "nvcuda failed to release primary device context %d", ret);
+  }
+}
+
+void nvcuda_release(nvcuda_handle_t h) {
+  LOG(h.verbose, "releasing nvcuda library\n");
+  UNLOAD_LIBRARY(h.handle);
+  // TODO and other context release logic?
+  h.handle = NULL;
+}
+
+#endif  // __APPLE__
--- a/gpu/gpu_info_nvcuda.h
+++ b/gpu/gpu_info_nvcuda.h
@@ -0,0 +1,71 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_NVCUDA_H__
+#define __GPU_INFO_NVCUDA_H__
+#include "gpu_info.h"
+
+// Just enough typedef's to dlopen/dlsym for memory information
+typedef enum cudaError_enum {
+  CUDA_SUCCESS = 0,
+  CUDA_ERROR_INVALID_VALUE = 1,
+  CUDA_ERROR_MEMORY_ALLOCATION = 2,
+  CUDA_ERROR_NOT_INITIALIZED = 3,
+  CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
+  // Other values omitted for now...
+} CUresult;
+
+typedef enum CUdevice_attribute_enum {
+  CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75,
+  CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR = 76,
+
+  // TODO - not yet wired up but may be useful for Jetson or other
+  // integrated GPU scenarios with shared memory
+  CU_DEVICE_ATTRIBUTE_INTEGRATED = 18
+
+} CUdevice_attribute;
+
+typedef void *nvcudaDevice_t;  // Opaque is sufficient
+typedef struct nvcudaMemory_st {
+  uint64_t total;
+  uint64_t free;
+} nvcudaMemory_t;
+
+typedef struct nvcudaDriverVersion {
+  int major;
+  int minor;
+} nvcudaDriverVersion_t;
+
+typedef struct CUuuid_st {
+    unsigned char bytes[16];
+} CUuuid;
+
+typedef int CUdevice;
+typedef void* CUcontext;
+
+typedef struct nvcuda_handle {
+  void *handle;
+  uint16_t verbose;
+  CUresult (*cuInit)(unsigned int Flags);
+  CUresult (*cuDriverGetVersion)(int *driverVersion);
+  CUresult (*cuDeviceGetCount)(int *);
+  CUresult (*cuDeviceGet)(CUdevice* device, int ordinal);
+  CUresult (*cuDeviceGetAttribute)(int* pi, CUdevice_attribute attrib, CUdevice dev);
+  CUresult (*cuDeviceGetUuid)(CUuuid* uuid, CUdevice dev); // signature compatible with cuDeviceGetUuid_v2
+
+  // Context specific aspects
+  CUresult (*cuCtxCreate_v3)(CUcontext* pctx, void *params, int len, unsigned int flags, CUdevice dev);
+  CUresult (*cuMemGetInfo_v2)(uint64_t* free, uint64_t* total);
+  CUresult (*cuCtxDestroy)(CUcontext ctx);
+} nvcuda_handle_t;
+
+typedef struct nvcuda_init_resp {
+  char *err;  // If err is non-null handle is invalid
+  nvcuda_handle_t ch;
+  int num_devices;
+} nvcuda_init_resp_t;
+
+void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);
+void nvcuda_check_vram(nvcuda_handle_t ch, int device_id, mem_info_t *resp);
+void nvcuda_release(nvcuda_handle_t ch);
+
+#endif  // __GPU_INFO_NVCUDA_H__
+#endif  // __APPLE__
--- a/llm/server.go
+++ b/llm/server.go
@@ -233,13 +233,13 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if runtime.GOOS == "windows" {
 			pathEnv = "PATH"
 		}
-		// append the server directory to LD_LIBRARY_PATH/PATH
+		// prepend the server directory to LD_LIBRARY_PATH/PATH
 		libraryPaths := []string{dir}

 		if libraryPath, ok := os.LookupEnv(pathEnv); ok {
 			// Append our runner directory to the path
 			// This will favor system libraries over our bundled library dependencies
-			libraryPaths = append(filepath.SplitList(libraryPath), libraryPaths...)
+			libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
 		}

 		// Note: we always put the dependency path first
@@ -275,15 +275,31 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			sem:           semaphore.NewWeighted(int64(numParallel)),
 		}

-		libEnv := fmt.Sprintf("%s=%s", pathEnv, strings.Join(libraryPaths, string(filepath.ListSeparator)))
-		s.cmd.Env = append(os.Environ(), libEnv)
+		s.cmd.Env = os.Environ()
 		s.cmd.Stdout = os.Stdout
 		s.cmd.Stderr = s.status

-		// TODO - multiple GPU selection logic...
-		key, val := gpu.GpuInfoList(gpus).GetVisibleDevicesEnv()
-		if key != "" {
-			s.cmd.Env = append(s.cmd.Env, key+"="+val)
+		visibleDevicesEnv, visibleDevicesEnvVal := gpu.GpuInfoList(gpus).GetVisibleDevicesEnv()
+		pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))
+
+		// Update or add the path and visible devices variable with our adjusted version
+		pathNeeded := true
+		devicesNeeded := visibleDevicesEnv != ""
+		for i := range s.cmd.Env {
+			cmp := strings.SplitN(s.cmd.Env[i], "=", 2)
+			if strings.EqualFold(cmp[0], pathEnv) {
+				s.cmd.Env[i] = pathEnv + "=" + pathEnvVal
+				pathNeeded = false
+			} else if devicesNeeded && strings.EqualFold(cmp[0], visibleDevicesEnv) {
+				s.cmd.Env[i] = visibleDevicesEnv + "=" + visibleDevicesEnvVal
+				devicesNeeded = false
+			}
+		}
+		if pathNeeded {
+			s.cmd.Env = append(s.cmd.Env, pathEnv+"="+pathEnvVal)
+		}
+		if devicesNeeded {
+			s.cmd.Env = append(s.cmd.Env, visibleDevicesEnv+"="+visibleDevicesEnvVal)
 		}

 		slog.Info("starting llama server", "cmd", s.cmd.String())
@@ -300,13 +316,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			continue
 		}

-		// TODO - make sure this is all wired up correctly
-		// if err = s.WaitUntilRunning(); err != nil {
-		// 	slog.Error("error starting llama server", "server", servers[i], "error", err)
-		// 	s.Close()
-		// 	finalErr = err
-		// 	continue
-		// }
 		return s, nil
 	}

@@ -338,7 +347,7 @@ type ServerStatus int

 const ( // iota is reset to 0
 	ServerStatusReady ServerStatus = iota
-	ServerStatusNoSlotsAvaialble
+	ServerStatusNoSlotsAvailable
 	ServerStatusLoadingModel
 	ServerStatusNotResponding
 	ServerStatusError
@@ -348,7 +357,7 @@ func (s ServerStatus) ToString() string {
 	switch s {
 	case ServerStatusReady:
 		return "llm server ready"
-	case ServerStatusNoSlotsAvaialble:
+	case ServerStatusNoSlotsAvailable:
 		return "llm busy - no slots available"
 	case ServerStatusLoadingModel:
 		return "llm server loading model"
@@ -405,7 +414,7 @@ func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
 	case "ok":
 		return ServerStatusReady, nil
 	case "no slot available":
-		return ServerStatusNoSlotsAvaialble, nil
+		return ServerStatusNoSlotsAvailable, nil
 	case "loading model":
 		return ServerStatusLoadingModel, nil
 	default:
@@ -413,6 +422,29 @@ func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
 	}
 }

+// getServerStatusRetry will retry if ServerStatusNoSlotsAvailable is received
+func (s *llmServer) getServerStatusRetry(ctx context.Context) (ServerStatus, error) {
+	var retries int
+	for {
+		status, err := s.getServerStatus(ctx)
+		if err != nil {
+			return status, err
+		}
+
+		if status == ServerStatusNoSlotsAvailable {
+			if retries >= 10 {
+				return status, fmt.Errorf("no slots available after %d retries", retries)
+			}
+
+			time.Sleep(5 * time.Millisecond)
+			retries++
+			continue
+		}
+
+		return status, nil
+	}
+}
+
 func (s *llmServer) Ping(ctx context.Context) error {
 	_, err := s.getServerStatus(ctx)
 	if err != nil {
@@ -510,7 +542,6 @@ ws ::= ([ \t\n] ws)?
 `

 const maxBufferSize = 512 * format.KiloByte
-const maxRetries = 3

 type ImageData struct {
 	Data []byte `json:"data"`
@@ -586,7 +617,7 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	}

 	// Make sure the server is ready
-	status, err := s.getServerStatus(ctx)
+	status, err := s.getServerStatusRetry(ctx)
 	if err != nil {
 		return err
 	} else if status != ServerStatusReady {
@@ -600,133 +631,113 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 		}
 	}

-	retryDelay := 100 * time.Microsecond
-	for retries := 0; retries < maxRetries; retries++ {
-		if retries > 0 {
-			time.Sleep(retryDelay) // wait before retrying
-			retryDelay *= 2        // exponential backoff
-		}
+	// Handling JSON marshaling with special characters unescaped.
+	buffer := &bytes.Buffer{}
+	enc := json.NewEncoder(buffer)
+	enc.SetEscapeHTML(false)

-		// Handling JSON marshaling with special characters unescaped.
-		buffer := &bytes.Buffer{}
-		enc := json.NewEncoder(buffer)
-		enc.SetEscapeHTML(false)
+	if err := enc.Encode(request); err != nil {
+		return fmt.Errorf("failed to marshal data: %v", err)
+	}

-		if err := enc.Encode(request); err != nil {
-			return fmt.Errorf("failed to marshal data: %v", err)
-		}
+	endpoint := fmt.Sprintf("http://127.0.0.1:%d/completion", s.port)
+	serverReq, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, buffer)
+	if err != nil {
+		return fmt.Errorf("error creating POST request: %v", err)
+	}
+	serverReq.Header.Set("Content-Type", "application/json")

-		endpoint := fmt.Sprintf("http://127.0.0.1:%d/completion", s.port)
-		req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, buffer)
+	res, err := http.DefaultClient.Do(serverReq)
+	if err != nil {
+		return fmt.Errorf("POST predict: %v", err)
+	}
+	defer res.Body.Close()
+
+	if res.StatusCode >= 400 {
+		bodyBytes, err := io.ReadAll(res.Body)
 		if err != nil {
-			return fmt.Errorf("error creating POST request: %v", err)
+			return fmt.Errorf("failed reading llm error response: %w", err)
 		}
-		req.Header.Set("Content-Type", "application/json")
+		log.Printf("llm predict error: %s", bodyBytes)
+		return fmt.Errorf("%s", bodyBytes)
+	}

-		resp, err := http.DefaultClient.Do(req)
-		if err != nil {
-			return fmt.Errorf("POST predict: %v", err)
-		}
-		defer resp.Body.Close()
+	scanner := bufio.NewScanner(res.Body)
+	buf := make([]byte, 0, maxBufferSize)
+	scanner.Buffer(buf, maxBufferSize)

-		if resp.StatusCode >= 400 {
-			bodyBytes, err := io.ReadAll(resp.Body)
-			if err != nil {
-				return fmt.Errorf("failed reading llm error response: %w", err)
+	// keep track of the last token generated, this is used to abort if the model starts looping
+	var lastToken string
+	var tokenRepeat int
+
+	for scanner.Scan() {
+		select {
+		case <-ctx.Done():
+			// This handles the request cancellation
+			return ctx.Err()
+		default:
+			line := scanner.Bytes()
+			if len(line) == 0 {
+				continue
 			}
-			log.Printf("llm predict error: %s", bodyBytes)
-			return fmt.Errorf("%s", bodyBytes)
-		}

-		scanner := bufio.NewScanner(resp.Body)
-		buf := make([]byte, 0, maxBufferSize)
-		scanner.Buffer(buf, maxBufferSize)
+			evt, ok := bytes.CutPrefix(line, []byte("data: "))
+			if !ok {
+				return fmt.Errorf("error parsing llm response stream: %s", line)
+			}

-		retryNeeded := false
-		// keep track of the last token generated, this is used to abort if the model starts looping
-		var lastToken string
-		var tokenRepeat int
+			var c completion
+			if err := json.Unmarshal(evt, &c); err != nil {
+				return fmt.Errorf("error unmarshaling llm prediction response: %v", err)
+			}

-		for scanner.Scan() {
-			select {
-			case <-ctx.Done():
-				// This handles the request cancellation
-				return ctx.Err()
+			switch {
+			case strings.TrimSpace(c.Content) == lastToken:
+				tokenRepeat++
 			default:
-				line := scanner.Bytes()
-				if len(line) == 0 {
-					continue
-				}
-
-				// try again on slot unavailable
-				if bytes.Contains(line, []byte("slot unavailable")) {
-					retryNeeded = true
-					break
-				}
-
-				evt, ok := bytes.CutPrefix(line, []byte("data: "))
-				if !ok {
-					return fmt.Errorf("error parsing llm response stream: %s", line)
-				}
-
-				var c completion
-				if err := json.Unmarshal(evt, &c); err != nil {
-					return fmt.Errorf("error unmarshaling llm prediction response: %v", err)
-				}
-
-				switch {
-				case strings.TrimSpace(c.Content) == lastToken:
-					tokenRepeat++
-				default:
-					lastToken = strings.TrimSpace(c.Content)
-					tokenRepeat = 0
-				}
-
-				// 30 picked as an arbitrary max token repeat limit, modify as needed
-				if tokenRepeat > 30 {
-					slog.Debug("prediction aborted, token repeat limit reached")
-					return ctx.Err()
-				}
-
-				if c.Content != "" {
-					fn(CompletionResponse{
-						Content: c.Content,
-					})
-				}
-
-				if c.Stop {
-					fn(CompletionResponse{
-						Done:               true,
-						PromptEvalCount:    c.Timings.PromptN,
-						PromptEvalDuration: parseDurationMs(c.Timings.PromptMS),
-						EvalCount:          c.Timings.PredictedN,
-						EvalDuration:       parseDurationMs(c.Timings.PredictedMS),
-					})
-					return nil
-				}
+				lastToken = strings.TrimSpace(c.Content)
+				tokenRepeat = 0
 			}
-		}

-		if err := scanner.Err(); err != nil {
-			if strings.Contains(err.Error(), "unexpected EOF") {
-				s.Close()
-				msg := ""
-				if s.status != nil && s.status.LastErrMsg != "" {
-					msg = s.status.LastErrMsg
-				}
-
-				return fmt.Errorf("an unknown error was encountered while running the model %s", msg)
+			// 30 picked as an arbitrary max token repeat limit, modify as needed
+			if tokenRepeat > 30 {
+				slog.Debug("prediction aborted, token repeat limit reached")
+				return ctx.Err()
 			}
-			return fmt.Errorf("error reading llm response: %v", err)
-		}

-		if !retryNeeded {
-			return nil // success
+			if c.Content != "" {
+				fn(CompletionResponse{
+					Content: c.Content,
+				})
+			}
+
+			if c.Stop {
+				fn(CompletionResponse{
+					Done:               true,
+					PromptEvalCount:    c.Timings.PromptN,
+					PromptEvalDuration: parseDurationMs(c.Timings.PromptMS),
+					EvalCount:          c.Timings.PredictedN,
+					EvalDuration:       parseDurationMs(c.Timings.PredictedMS),
+				})
+				return nil
+			}
 		}
 	}

-	// should never reach here ideally
-	return fmt.Errorf("max retries exceeded")
+	if err := scanner.Err(); err != nil {
+		if strings.Contains(err.Error(), "unexpected EOF") {
+			s.Close()
+			msg := ""
+			if s.status != nil && s.status.LastErrMsg != "" {
+				msg = s.status.LastErrMsg
+			}
+			return fmt.Errorf("an unknown error was encountered while running the model %s", msg)
+		}
+
+		return fmt.Errorf("error reading llm response: %v", err)
+	}
+
+	return nil
 }

 type EmbeddingRequest struct {
@@ -743,8 +754,9 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
 		return nil, err
 	}
 	defer s.sem.Release(1)
+
 	// Make sure the server is ready
-	status, err := s.getServerStatus(ctx)
+	status, err := s.getServerStatusRetry(ctx)
 	if err != nil {
 		return nil, err
 	} else if status != ServerStatusReady {
@@ -799,7 +811,7 @@ func (s *llmServer) Tokenize(ctx context.Context, content string) ([]int, error)
 	status, err := s.getServerStatus(ctx)
 	if err != nil {
 		return nil, err
-	} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvaialble {
+	} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvailable {
 		return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
 	}

@@ -851,7 +863,7 @@ func (s *llmServer) Detokenize(ctx context.Context, tokens []int) (string, error
 	status, err := s.getServerStatus(ctx)
 	if err != nil {
 		return "", err
-	} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvaialble {
+	} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvailable {
 		return "", fmt.Errorf("unexpected server status: %s", status.ToString())
 	}

--- a/scripts/install.sh
+++ b/scripts/install.sh
@@ -166,8 +166,8 @@ fi

 if check_gpu lspci amdgpu || check_gpu lshw amdgpu; then
    # Look for pre-existing ROCm v6 before downloading the dependencies
-    for search in "${HIP_PATH:-''}" "${ROCM_PATH:-''}" "/opt/rocm"; do
-        if [ -n "${search}" ] && [ -e "${search}/lib/libhipblas.so.2" ]; then
+    for search in "${HIP_PATH:-''}" "${ROCM_PATH:-''}" "/opt/rocm" "/usr/lib64"; do
+        if [ -n "${search}" ] && [ -e "${search}/libhipblas.so.2" -o -e "${search}/lib/libhipblas.so.2" ]; then
            status "Compatible AMD GPU ROCm library detected at ${search}"
            install_success
            exit 0
--- a/server/routes.go
+++ b/server/routes.go
@@ -1041,6 +1041,7 @@ func Serve(ln net.Listener) error {
 	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
 	go func() {
 		<-signals
+		srvr.Close()
 		done()
 		sched.unloadAllRunners()
 		gpu.Cleanup()
--- a/server/sched.go
+++ b/server/sched.go
@@ -100,6 +100,12 @@ func (s *Scheduler) processPending(ctx context.Context) {
 			return
 		case pending := <-s.pendingReqCh:
 			// Block other requests until we get this pending request running
+
+			if pending.ctx.Err() != nil {
+				slog.Debug("pending request cancelled or timed out, skipping scheduling")
+				continue
+			}
+
 			for {
 				var runnerToExpire *runnerRef
 				s.loadedMu.Lock()
@@ -435,6 +441,10 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 		timeout = 2 * time.Minute // Initial load can take a long time for big models on slow systems...
 	}

+	if runner.Options == nil {
+		return true
+	}
+
 	// Don't reload runner if num_gpu=-1 was provided
 	optsExisting := runner.Options.Runner
 	optsNew := req.opts.Runner
--- a/server/sched_test.go
+++ b/server/sched_test.go
@@ -352,11 +352,9 @@ func TestGetRunner(t *testing.T) {
 	scenario1c.req.model.ModelPath = "bad path"
 	slog.Info("scenario1c")
 	successCh1c, errCh1c := s.GetRunner(scenario1c.ctx, scenario1c.req.model, scenario1c.req.opts, scenario1c.req.sessionDuration)
-	require.Len(t, s.pendingReqCh, 0)
-	require.Len(t, successCh1c, 0)
-	require.Len(t, errCh1c, 0)
-
+	// Starts in pending channel, then should be quickly processsed to return an error
 	time.Sleep(5 * time.Millisecond)
+	require.Len(t, successCh1c, 0)
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 0)
 	s.loadedMu.Unlock()