Added support for serving behind a reverse proxy

- Added support for specifying response headers for each backend
  - Allowing users to set `X-Accel-Buffering: no` to disable buffering for streaming responses in nginx
  - Updated `configuration.md` to document the new configuration options
- Modified Vite config to build with `base: "./"`, making assets be accessed via relative paths
- Updated API_BASE to use `document.baseURI`, allowing API calls to be made relative to the base path

docs/getting-started/configuration.md

@@ -122,6 +122,7 @@ backends:
       image: "ghcr.io/ggml-org/llama.cpp:server"
       args: ["run", "--rm", "--network", "host", "--gpus", "all"]
       environment: {}
+    response_headers: {}               # Additional response headers to send with responses
 
   vllm:
     command: "vllm"
@@ -132,24 +133,29 @@ backends:
       image: "vllm/vllm-openai:latest"
       args: ["run", "--rm", "--network", "host", "--gpus", "all", "--shm-size", "1g"]
       environment: {}
+    response_headers: {}               # Additional response headers to send with responses
 
   mlx:
     command: "mlx_lm.server"
     args: []
     environment: {}                    # Environment variables for the backend process
     # MLX does not support Docker
+    response_headers: {}               # Additional response headers to send with responses
 ```
 
 **Backend Configuration Fields:**
 - `command`: Executable name/path for the backend
 - `args`: Default arguments prepended to all instances
 - `environment`: Environment variables for the backend process (optional)
+- `response_headers`: Additional response headers to send with responses (optional)
 - `docker`: Docker-specific configuration (optional)
   - `enabled`: Boolean flag to enable Docker runtime
   - `image`: Docker image to use
   - `args`: Additional arguments passed to `docker run`
   - `environment`: Environment variables for the container (optional)
 
+> If llamactl is behind an nginx proxy, `X-Accel-Buffering: no` may be required for nginx to properly stream the responses without buffering.
+
 **Environment Variables:**
 
 **LlamaCpp Backend:**
@@ -160,6 +166,7 @@ backends:
 - `LLAMACTL_LLAMACPP_DOCKER_IMAGE` - Docker image to use
 - `LLAMACTL_LLAMACPP_DOCKER_ARGS` - Space-separated Docker arguments
 - `LLAMACTL_LLAMACPP_DOCKER_ENV` - Docker environment variables in format "KEY1=value1,KEY2=value2"
+- `LLAMACTL_LLAMACPP_RESPONSE_HEADERS` - Response headers in format "KEY1=value1,KEY2=value2"
 
 **VLLM Backend:**
 - `LLAMACTL_VLLM_COMMAND` - VLLM executable command

pkg/config/config.go

@@ -17,6 +17,7 @@ type BackendSettings struct {
 	Args            []string          `yaml:"args"`
 	Environment     map[string]string `yaml:"environment,omitempty"`
 	Docker          *DockerSettings   `yaml:"docker,omitempty"`
+	ResponseHeaders map[string]string `yaml:"response_headers,omitempty"`
 }
 
 // DockerSettings contains Docker-specific configuration
@@ -58,6 +59,9 @@ type ServerConfig struct {
 
 	// Enable Swagger UI for API documentation
 	EnableSwagger bool `yaml:"enable_swagger"`
+
+	// Response headers to send with responses
+	ResponseHeaders map[string]string `yaml:"response_headers,omitempty"`
 }
 
 // InstancesConfig contains instance management configuration
@@ -337,6 +341,12 @@ func loadEnvVars(cfg *AppConfig) {
 		}
 		parseEnvVars(llamaDockerEnv, cfg.Backends.LlamaCpp.Docker.Environment)
 	}
+	if llamaEnv := os.Getenv("LLAMACTL_LLAMACPP_RESPONSE_HEADERS"); llamaEnv != "" {
+		if cfg.Backends.LlamaCpp.ResponseHeaders == nil {
+			cfg.Backends.LlamaCpp.ResponseHeaders = make(map[string]string)
+		}
+		parseHeaders(llamaEnv, cfg.Backends.LlamaCpp.ResponseHeaders)
+	}
 
 	// vLLM backend
 	if vllmCmd := os.Getenv("LLAMACTL_VLLM_COMMAND"); vllmCmd != "" {
@@ -380,6 +390,12 @@ func loadEnvVars(cfg *AppConfig) {
 		}
 		parseEnvVars(vllmDockerEnv, cfg.Backends.VLLM.Docker.Environment)
 	}
+	if llamaEnv := os.Getenv("LLAMACTL_VLLM_RESPONSE_HEADERS"); llamaEnv != "" {
+		if cfg.Backends.VLLM.ResponseHeaders == nil {
+			cfg.Backends.VLLM.ResponseHeaders = make(map[string]string)
+		}
+		parseHeaders(llamaEnv, cfg.Backends.VLLM.ResponseHeaders)
+	}
 
 	// MLX backend
 	if mlxCmd := os.Getenv("LLAMACTL_MLX_COMMAND"); mlxCmd != "" {
@@ -394,6 +410,12 @@ func loadEnvVars(cfg *AppConfig) {
 		}
 		parseEnvVars(mlxEnv, cfg.Backends.MLX.Environment)
 	}
+	if llamaEnv := os.Getenv("LLAMACTL_MLX_RESPONSE_HEADERS"); llamaEnv != "" {
+		if cfg.Backends.MLX.ResponseHeaders == nil {
+			cfg.Backends.MLX.ResponseHeaders = make(map[string]string)
+		}
+		parseHeaders(llamaEnv, cfg.Backends.MLX.ResponseHeaders)
+	}
 
 	// Instance defaults
 	if autoRestart := os.Getenv("LLAMACTL_DEFAULT_AUTO_RESTART"); autoRestart != "" {
@@ -481,6 +503,19 @@ func parseEnvVars(envString string, envMap map[string]string) {
 	}
 }
 
+// parseHeaders parses HTTP headers in format "KEY1=value1;KEY2=value2"
+// and populates the provided environment map
+func parseHeaders(envString string, envMap map[string]string) {
+	if envString == "" {
+		return
+	}
+	for _, envPair := range strings.Split(envString, ";") {
+		if parts := strings.SplitN(strings.TrimSpace(envPair), "=", 2); len(parts) == 2 {
+			envMap[parts[0]] = parts[1]
+		}
+	}
+}
+
 // getDefaultDataDirectory returns platform-specific default data directory
 func getDefaultDataDirectory() string {
 	switch runtime.GOOS {

pkg/instance/instance.go

@@ -198,6 +198,15 @@ func (i *Process) GetProxy() (*httputil.ReverseProxy, error) {
 
 	proxy := httputil.NewSingleHostReverseProxy(targetURL)
 
+	var responseHeaders map[string]string
+	switch i.options.BackendType {
+	case backends.BackendTypeLlamaCpp:
+		responseHeaders = i.globalBackendSettings.LlamaCpp.ResponseHeaders
+	case backends.BackendTypeVllm:
+		responseHeaders = i.globalBackendSettings.VLLM.ResponseHeaders
+	case backends.BackendTypeMlxLm:
+		responseHeaders = i.globalBackendSettings.MLX.ResponseHeaders
+	}
 	proxy.ModifyResponse = func(resp *http.Response) error {
 		// Remove CORS headers from llama-server response to avoid conflicts
 		// llamactl will add its own CORS headers
@@ -207,6 +216,10 @@ func (i *Process) GetProxy() (*httputil.ReverseProxy, error) {
 		resp.Header.Del("Access-Control-Allow-Credentials")
 		resp.Header.Del("Access-Control-Max-Age")
 		resp.Header.Del("Access-Control-Expose-Headers")
+
+		for key, value := range responseHeaders {
+			resp.Header.Set(key, value)
+		}
 		return nil
 	}
 

webui/src/contexts/AuthContext.tsx

@@ -1,4 +1,5 @@
-import { type ReactNode, createContext, useContext, useState, useEffect, useCallback } from 'react'
+import { API_BASE } from '@/lib/api'
+import { type ReactNode, createContext, useCallback, useContext, useEffect, useState } from 'react'
 
 interface AuthContextState {
   isAuthenticated: boolean
@@ -62,7 +63,7 @@ export const AuthProvider = ({ children }: AuthProviderProps) => {
   // Validate API key by making a test request
   const validateApiKey = async (key: string): Promise<boolean> => {
     try {
-      const response = await fetch('/api/v1/instances', {
+      const response = await fetch(API_BASE + '/instances', {
         headers: {
           'Authorization': `Bearer ${key}`,
           'Content-Type': 'application/json'

webui/src/lib/api.ts

@@ -1,7 +1,10 @@
 import type { CreateInstanceOptions, Instance } from "@/types/instance";
 import { handleApiError } from "./errorUtils";
 
-const API_BASE = "/api/v1";
+// Adding baseURI as a prefix to support being served behind a subpath
+// e.g. when llmamctl's `/` is served behind a reverse proxy at `/proxy/...`
+// the baseURI will be `/proxy/` and the API calls will be made to `/proxy/api/v1/<endpoint>`
+export const API_BASE = document.baseURI + "api/v1";
 
 // Base API call function with error handling
 async function apiCall<T>(

webui/vite.config.ts

@@ -21,4 +21,6 @@ export default defineConfig({
     setupFiles: ['./src/test/setup.ts'],
     css: true,
   },
+  // ensures relative asset paths to support being served behind a subpath
+  base: "./"
 })