[model-config] feat: add environment variables support for backends

backend/backend.proto

@@ -445,6 +445,9 @@ message ModelOptions {
   // Unknown keys produce an error at LoadModel time.
   string EngineArgs = 73;
 
+  // EnvVars carries environment variables to be passed to the backend process.
+  map<string, string> EnvVars = 75;
+
   // Proxy carries the cloud-proxy backend's per-model configuration.
   // Empty for non-proxy backends.
   ProxyOptions Proxy = 74;

core/backend/options.go

@@ -187,6 +187,10 @@ func ModelOptions(c config.ModelConfig, so *config.ApplicationConfig, opts ...mo
 		defOpts = append(defOpts, model.WithModelSizeBytes(sizeBytes))
 	}
 
+	if c.Environment != nil && len(c.Environment) > 0 {
+		defOpts = append(defOpts, model.WithEnvVars(c.Environment))
+	}
+
 	return append(defOpts, opts...)
 }
 
@@ -421,6 +425,14 @@ func grpcModelOpts(c config.ModelConfig, modelPath string) *pb.ModelOptions {
 		opts.DraftModel = filepath.Join(modelPath, c.DraftModel)
 	}
 
+	// Add environment variables from model configuration
+	if c.Environment != nil && len(c.Environment) > 0 {
+		opts.EnvVars = make(map[string]string)
+		for k, v := range c.Environment {
+			opts.EnvVars[k] = v
+		}
+	}
+
 	return opts
 }
 

core/config/meta/registry.go

@@ -66,6 +66,13 @@ func DefaultRegistry() map[string]FieldMetaOverride {
 			Options:     UsecaseOptions,
 			Order:       6,
 		},
+		"env": {
+			Section:     "general",
+			Label:       "Environment Variables",
+			Description: "Environment variables to be applied to the backend process",
+			Component:   "map-editor",
+			Order:       7,
+		},
 
 		// --- LLM ---
 		"context_size": {

core/config/model_config.go

@@ -134,6 +134,9 @@ type ModelConfig struct {
 	Proxy        ProxyConfig        `yaml:"proxy,omitempty" json:"proxy,omitempty"`
 	MITM         MITMModelConfig    `yaml:"mitm,omitempty" json:"mitm,omitempty"`
 	Limits       LimitsConfig       `yaml:"limits,omitempty" json:"limits,omitempty"`
+
+	// Environment variables to set when starting the backend process
+	Environment map[string]string `yaml:"env,omitempty" json:"env,omitempty"`
 }
 
 // @Description Admission-control limits applied per request. The

core/config/model_config_test.go

@@ -98,6 +98,12 @@ parameters:
 			Expect(valid).To(BeTrue())
 			Expect(err).NotTo(HaveOccurred())
 
+			// Test environment variables configuration parsing from YAML
+			envYAML := "name: env-test-model\nbackend: test-backend\nenv:\n  TEST_ENV_VAR: test_value\n  ANOTHER_VAR: another_value\n"
+			var envConfig ModelConfig
+			Expect(yaml.Unmarshal([]byte(envYAML), &envConfig)).To(Succeed())
+			Expect(envConfig.Environment).To(HaveKeyWithValue("TEST_ENV_VAR", "test_value"))
+			Expect(envConfig.Environment).To(HaveKeyWithValue("ANOTHER_VAR", "another_value"))
 			tcAndChat := FLAG_TOKEN_CLASSIFY | FLAG_CHAT
 			tcCombined := ModelConfig{
 				Name:          "ner-and-chat",

core/services/worker/supervisor.go

@@ -81,7 +81,7 @@ func (s *backendSupervisor) startBackend(backend, backendPath string) (string, e
 	bindAddr := fmt.Sprintf("0.0.0.0:%d", port)
 	clientAddr := fmt.Sprintf("127.0.0.1:%d", port)
 
-	proc, err := s.ml.StartProcess(backendPath, backend, bindAddr)
+	proc, err := s.ml.StartProcess(backendPath, backend, bindAddr, nil)
 	if err != nil {
 		s.mu.Unlock()
 		return "", fmt.Errorf("starting backend process: %w", err)

docs/content/advanced/model-configuration.md

@@ -909,6 +909,24 @@ Multiple detectors union their detections; overlapping spans resolve to the stro
 
 > The earlier regex pattern tier (`pii.patterns`, the global pattern catalogue, `--pii-config`, and the `/api/pii/patterns` admin endpoints) has been removed, along with response/streaming-side redaction. Those keys now no-op with a startup warning; migrate to `pii.detectors` + a detector's `pii_detection` block.
 
+## Environment Variables Configuration
+
+Model configurations can specify environment variables passed to the backend process:
+
+```yaml
+name: vllm-model
+backend: vllm
+parameters:
+  model: my-vllm-model
+
+env:
+  VLLM_WORKER_MULTIPROC_METHOD: "spawn"
+  VLLM_CACHE_DIR: "/tmp/vllm_cache"
+  CUDA_VISIBLE_DEVICES: "0,1"
+```
+
+Environment variables are appended to the system environment variables and will override any conflicting system variables with the same name.
+
 ## Complete Example
 
 Here's a comprehensive example combining many options:

pkg/model/initializers.go

@@ -119,7 +119,7 @@ func (ml *ModelLoader) spawnGRPCModel(backend, uri string, o *Options, modelID,
 				return nil, fmt.Errorf("failed allocating free ports: %s", err.Error())
 			}
 			// Make sure the process is executable
-			process, err := ml.startProcess(uri, modelID, serverAddress)
+			process, err := ml.StartProcess(uri, modelID, serverAddress, o.envVars)
 			if err != nil {
 				xlog.Error("failed to launch", "error", err, "path", uri)
 				return nil, err

pkg/model/loader_options.go

@@ -24,6 +24,9 @@ type Options struct {
 	// by the caller using the vram estimation scaffolding.  When non-zero it is
 	// registered with the watchdog so size-aware eviction can rank models.
 	modelSizeBytes int64
+
+	// envVars contains model-specific environment variables to pass to the backend process
+	envVars map[string]string
 }
 
 type Option func(*Options)
@@ -97,6 +100,12 @@ func WithModelSizeBytes(bytes int64) Option {
 	}
 }
 
+func WithEnvVars(envVars map[string]string) Option {
+	return func(o *Options) {
+		o.envVars = envVars
+	}
+}
+
 func NewOptions(opts ...Option) *Options {
 	o := &Options{
 		gRPCOptions:       &pb.ModelOptions{},

pkg/model/process.go

@@ -141,11 +141,11 @@ func (ml *ModelLoader) GetGRPCPID(id string) (int, error) {
 // StartProcess starts a gRPC backend process and returns its process handle.
 // This is the public wrapper for the internal startProcess method, used by
 // the serve-backend CLI subcommand to start a backend on a specified address.
-func (ml *ModelLoader) StartProcess(grpcProcess, id string, serverAddress string, args ...string) (*process.Process, error) {
-	return ml.startProcess(grpcProcess, id, serverAddress, args...)
+func (ml *ModelLoader) StartProcess(grpcProcess, id string, serverAddress string, envVars map[string]string, args ...string) (*process.Process, error) {
+	return ml.startProcess(grpcProcess, id, serverAddress, envVars, args...)
 }
 
-func (ml *ModelLoader) startProcess(grpcProcess, id string, serverAddress string, args ...string) (*process.Process, error) {
+func (ml *ModelLoader) startProcess(grpcProcess, id string, serverAddress string, envVars map[string]string, args ...string) (*process.Process, error) {
 	// Make sure the process is executable
 	// Check first if it has executable permissions
 	if fi, err := os.Stat(grpcProcess); err == nil {
@@ -175,6 +175,13 @@ func (ml *ModelLoader) startProcess(grpcProcess, id string, serverAddress string
 	// and the GPU would silently fall back to CPU). No-op for other backends.
 	env = append(env, vulkanICDEnv(workDir)...)
 
+	// Add model-specific environment variables
+	if envVars != nil {
+		for key, value := range envVars {
+			env = append(env, fmt.Sprintf("%s=%s", key, value))
+		}
+	}
+
 	grpcControlProcess := process.New(
 		process.WithTemporaryStateDir(),
 		process.WithName(filepath.Base(grpcProcess)),