Always filter devices (#12108)

* Always filter devices

Avoid crashing on unsupported AMD iGPUs

* Remove cuda device filtering

This interferes with mixed setups

llm/server.go

@@ -360,23 +360,28 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 
 		s.cmd.Env = append(s.cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ggmlPaths, string(filepath.ListSeparator)))
 
-		envWorkarounds := [][2]string{}
+		envWorkarounds := []string{}
 		for _, gpu := range gpus {
 			envWorkarounds = append(envWorkarounds, gpu.EnvWorkarounds...)
 		}
+		// Always filter down the set of GPUs in case there are any unsupported devices that might crash
+		envWorkarounds = append(envWorkarounds, gpus.GetVisibleDevicesEnv()...)
 		pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))
 
 		// Update or add the path variable with our adjusted version
 		pathNeeded := true
+		envWorkaroundDone := make([]bool, len(envWorkarounds))
 		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 len(envWorkarounds) != 0 {
-				for _, kv := range envWorkarounds {
-					if strings.EqualFold(cmp[0], kv[0]) {
-						s.cmd.Env[i] = kv[0] + "=" + kv[1]
+				for j, kv := range envWorkarounds {
+					tmp := strings.SplitN(kv, "=", 2)
+					if strings.EqualFold(cmp[0], tmp[0]) {
+						s.cmd.Env[i] = kv
+						envWorkaroundDone[j] = true
 					}
 				}
 			}
@@ -384,6 +389,11 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		if pathNeeded {
 			s.cmd.Env = append(s.cmd.Env, pathEnv+"="+pathEnvVal)
 		}
+		for i, done := range envWorkaroundDone {
+			if !done {
+				s.cmd.Env = append(s.cmd.Env, envWorkarounds[i])
+			}
+		}
 
 		slog.Info("starting runner", "cmd", s.cmd)
 		slog.Debug("subprocess", "", filteredEnv(s.cmd.Env))