Honor explicit model URL overrides for NVIDIA endpoints (#18152)

Some NVIDIA hosted models (e.g. meta/llama-3.2-11b-vision-instruct )
expose a full
endpoint URL per model that does not follow the normal base_url +
url_suffix
assembly. Previously the Go driver always called {base}/chat/completions
, so chat
requests for these vision models hit the wrong endpoint and failed.

This PR adds an optional per-model url field in conf/models/nvidia.json
. When
present, every NVIDIA driver request (chat, streaming chat, embedding,
rerank, model
listing) uses it directly; otherwise the standard assembly is unchanged.

conf/models/nvidia.json

@@ -72,6 +72,7 @@
       "name": "meta/llama-3.2-11b-vision-instruct",
       "content_length": 131072,
       "max_output": 8192,
+      "url":"https://integrate.api.nvidia.com/v1/meta/llama-3.2-11b-vision-instruct",
       "model_types": [
         "chat",
         "vision"
@@ -97,6 +98,7 @@
       "name": "meta/llama-3.2-90b-vision-instruct",
       "content_length": 131072,
       "max_output": 8192,
+      "url":"https://integrate.api.nvidia.com/v1/meta/llama-3.2-90b-vision-instruct",
       "model_types": [
         "chat",
         "vision"

internal/entity/models/model.go

@@ -166,6 +166,7 @@ type Model struct {
 	Thinking      *ModelThinking `json:"thinking"`
 	Tools         *ModelTools    `json:"tools"`
 	Class         *string        `json:"class"`
+	URL           string         `json:"url"`
 	MaxDimension  *int           `json:"max_dimension"`  // used by embedding models
 	MaxBatchSize  *int           `json:"max_batch_size"` // used by embedding models
 	Dimensions    []int          `json:"dimensions"`

internal/entity/models/nvidia.go

@@ -64,6 +64,30 @@ func (n *NvidiaModel) Name() string {
 	return "nvidia"
 }
 
+// resolveEndpoint returns the endpoint URL for modelName. When the
+// model's preset config in conf/models/nvidia.json carries an explicit
+// url, that full endpoint is used as-is because it does not follow the
+// standard baseURL + suffix assembly; otherwise the usual assembly
+// applies. Every endpoint (chat, embedding, rerank, models list) goes
+// through this path so the override is honored uniformly.
+func (n *NvidiaModel) resolveEndpoint(modelName string, apiConfig *APIConfig, suffix string) (string, error) {
+	if modelName != "" {
+		if pm := GetProviderManager(); pm != nil {
+			if provider := pm.FindProvider("NVIDIA"); provider != nil {
+				if model := pm.FindModel(provider, modelName); model != nil && model.URL != "" {
+					return strings.TrimSuffix(model.URL, "/"), nil
+				}
+			}
+		}
+	}
+
+	resolvedBaseURL, err := n.baseModel.GetBaseURL(apiConfig)
+	if err != nil {
+		return "", err
+	}
+	return fmt.Sprintf("%s/%s", strings.TrimRight(resolvedBaseURL, "/"), strings.TrimLeft(suffix, "/")), nil
+}
+
 func (n *NvidiaModel) ChatWithMessages(ctx context.Context, modelName string, messages []Message, apiConfig *APIConfig, chatModelConfig *ChatConfig, modelUsage *common.ModelUsage) (*ChatResponse, error) {
 	if err := n.baseModel.APIConfigCheck(apiConfig); err != nil {
 		return nil, err
@@ -73,11 +97,10 @@ func (n *NvidiaModel) ChatWithMessages(ctx context.Context, modelName string, me
 		return nil, fmt.Errorf("messages is empty")
 	}
 
-	resolvedBaseURL, err := n.baseModel.GetBaseURL(apiConfig)
+	baseURL, err := n.resolveEndpoint(modelName, apiConfig, n.baseModel.URLSuffix.Chat)
 	if err != nil {
 		return nil, err
 	}
-	baseURL := fmt.Sprintf("%s/%s", resolvedBaseURL, n.baseModel.URLSuffix.Chat)
 	reqBody := buildRequestBody(chatModelConfig, modelName, messages, false)
 
 	if chatModelConfig != nil {
@@ -108,11 +131,10 @@ func (n *NvidiaModel) ChatStreamlyWithSender(ctx context.Context, modelName stri
 		return fmt.Errorf("messages is empty")
 	}
 
-	resolvedBaseURL, err := n.baseModel.GetBaseURL(apiConfig)
+	baseURL, err := n.resolveEndpoint(modelName, apiConfig, n.baseModel.URLSuffix.Chat)
 	if err != nil {
 		return err
 	}
-	baseURL := fmt.Sprintf("%s/%s", resolvedBaseURL, n.baseModel.URLSuffix.Chat)
 	reqBody := buildRequestBody(modelConfig, modelName, messages, true)
 
 	if modelConfig != nil {
@@ -152,13 +174,11 @@ func (n *NvidiaModel) Embed(ctx context.Context, modelName *string, texts []stri
 		return nil, fmt.Errorf("model name is required")
 	}
 
-	resolvedBaseURL, err := n.baseModel.GetBaseURL(apiConfig)
+	baseURL, err := n.resolveEndpoint(*modelName, apiConfig, n.baseModel.URLSuffix.Embedding)
 	if err != nil {
 		return nil, err
 	}
 
-	baseURL := fmt.Sprintf("%s/%s", strings.TrimSuffix(resolvedBaseURL, "/"), n.baseModel.URLSuffix.Embedding)
-
 	reqBody := map[string]interface{}{
 		"model":           *modelName,
 		"input":           texts,
@@ -264,13 +284,11 @@ func (n *NvidiaModel) Rerank(ctx context.Context, modelName *string, query strin
 		return nil, fmt.Errorf("model name is required")
 	}
 
-	resolvedBaseURL, err := n.baseModel.GetBaseURL(apiConfig)
+	baseURL, err := n.resolveEndpoint(*modelName, apiConfig, n.baseModel.URLSuffix.Rerank)
 	if err != nil {
 		return nil, err
 	}
 
-	baseURL := fmt.Sprintf("%s/%s", strings.TrimSuffix(resolvedBaseURL, "/"), n.baseModel.URLSuffix.Rerank)
-
 	topN := len(documents)
 	if rerankConfig != nil && rerankConfig.TopN > 0 && rerankConfig.TopN < topN {
 		topN = rerankConfig.TopN
@@ -376,11 +394,10 @@ func (n *NvidiaModel) ListModels(ctx context.Context, apiConfig *APIConfig) ([]L
 		return nil, err
 	}
 
-	resolvedBaseURL, err := n.baseModel.GetBaseURL(apiConfig)
+	baseURL, err := n.resolveEndpoint("", apiConfig, n.baseModel.URLSuffix.Models)
 	if err != nil {
 		return nil, err
 	}
-	baseURL := fmt.Sprintf("%s/%s", strings.TrimRight(resolvedBaseURL, "/"), strings.TrimLeft(n.baseModel.URLSuffix.Models, "/"))
 
 	modelListCtx, cancel := context.WithTimeout(ctx, nonStreamCallTimeout)
 	defer cancel()
@@ -422,7 +439,7 @@ func (n *NvidiaModel) ListModels(ctx context.Context, apiConfig *APIConfig) ([]L
 		provider = pm.FindProvider("NVIDIA")
 	}
 	models := parseNvidiaModelList(modelList, provider)
-	if n.usesHostedCatalog(resolvedBaseURL) {
+	if n.usesHostedCatalog(baseURL) {
 		catalogCtx, catalogCancel := context.WithTimeout(ctx, nonStreamCallTimeout)
 		catalog, catalogErr := n.fetchHostedCatalog(catalogCtx)
 		catalogCancel()

internal/entity/models/nvidia_test.go

@@ -229,6 +229,162 @@ func TestNvidiaCatalogResourceRejectsMalformedDeprecation(t *testing.T) {
 	}
 }
 
+func withNvidiaProviderManager(t *testing.T, models []*Model) {
+	t.Helper()
+	saved := providerManager
+	providerManager = &ProviderManager{
+		Providers: []Provider{{
+			Name:   "NVIDIA",
+			URL:    map[string]string{"default": "https://integrate.api.nvidia.com/v1"},
+			Models: models,
+		}},
+	}
+	t.Cleanup(func() { providerManager = saved })
+}
+
+func TestNvidiaResolveEndpointUsesModelURL(t *testing.T) {
+	const modelURL = "https://integrate.api.nvidia.com/v1/meta/llama-3.2-11b-vision-instruct"
+	withNvidiaProviderManager(t, []*Model{{
+		Name: "meta/llama-3.2-11b-vision-instruct",
+		URL:  modelURL,
+	}})
+
+	driver := NewNvidiaModel(
+		map[string]string{"default": "https://integrate.api.nvidia.com/v1"},
+		URLSuffix{Chat: "chat/completions"},
+	)
+	got, err := driver.resolveEndpoint("meta/llama-3.2-11b-vision-instruct", &APIConfig{}, "chat/completions")
+	if err != nil {
+		t.Fatalf("resolveEndpoint() error = %v", err)
+	}
+	if got != modelURL {
+		t.Fatalf("resolveEndpoint() = %q, want %q", got, modelURL)
+	}
+}
+
+func TestNvidiaResolveEndpointFallsBackToAssembly(t *testing.T) {
+	withNvidiaProviderManager(t, []*Model{{Name: "meta/llama-3.1-8b-instruct"}})
+
+	driver := NewNvidiaModel(
+		map[string]string{"default": "https://integrate.api.nvidia.com/v1/"},
+		URLSuffix{Chat: "chat/completions", Embedding: "embeddings"},
+	)
+	got, err := driver.resolveEndpoint("meta/llama-3.1-8b-instruct", &APIConfig{}, "chat/completions")
+	if err != nil {
+		t.Fatalf("resolveEndpoint() error = %v", err)
+	}
+	if want := "https://integrate.api.nvidia.com/v1/chat/completions"; got != want {
+		t.Fatalf("resolveEndpoint() = %q, want %q", got, want)
+	}
+
+	got, err = driver.resolveEndpoint("meta/llama-3.1-8b-instruct", &APIConfig{}, "embeddings")
+	if err != nil {
+		t.Fatalf("resolveEndpoint() error = %v", err)
+	}
+	if want := "https://integrate.api.nvidia.com/v1/embeddings"; got != want {
+		t.Fatalf("resolveEndpoint() = %q, want %q", got, want)
+	}
+}
+
+func TestNvidiaResolveEndpointIgnoresManagerWhenNil(t *testing.T) {
+	saved := providerManager
+	providerManager = nil
+	defer func() { providerManager = saved }()
+
+	driver := NewNvidiaModel(
+		map[string]string{"default": "https://integrate.api.nvidia.com/v1"},
+		URLSuffix{},
+	)
+	got, err := driver.resolveEndpoint("meta/llama-3.2-11b-vision-instruct", &APIConfig{}, "chat/completions")
+	if err != nil {
+		t.Fatalf("resolveEndpoint() error = %v", err)
+	}
+	if want := "https://integrate.api.nvidia.com/v1/chat/completions"; got != want {
+		t.Fatalf("resolveEndpoint() = %q, want %q", got, want)
+	}
+}
+
+func TestNvidiaChatUsesModelSpecificURL(t *testing.T) {
+	withSSRFBypass(t)
+	const apiKey = "nvapi-test"
+	const modelName = "meta/llama-3.2-11b-vision-instruct"
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/meta/llama-3.2-11b-vision-instruct" {
+			t.Fatalf("path = %s, want model-specific chat endpoint", r.URL.Path)
+		}
+		_ = json.NewEncoder(w).Encode(map[string]interface{}{
+			"choices": []map[string]interface{}{{
+				"message":       map[string]interface{}{"role": "assistant", "content": "pong"},
+				"finish_reason": "stop",
+			}},
+			"usage": map[string]interface{}{"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+		})
+	}))
+	defer server.Close()
+
+	withNvidiaProviderManager(t, []*Model{{
+		Name: modelName,
+		URL:  server.URL + "/v1/meta/llama-3.2-11b-vision-instruct",
+	}})
+	driver := NewNvidiaModel(
+		map[string]string{"default": server.URL + "/v1"},
+		URLSuffix{Chat: "chat/completions"},
+	)
+	resp, err := driver.ChatWithMessages(
+		context.Background(),
+		modelName,
+		[]Message{{Role: "user", Content: "hi"}},
+		&APIConfig{ApiKey: ptr(apiKey)},
+		nil,
+		nil,
+	)
+	if err != nil {
+		t.Fatalf("ChatWithMessages() error = %v", err)
+	}
+	if resp.Answer == nil || *resp.Answer != "pong" {
+		t.Fatalf("answer = %v, want pong", resp.Answer)
+	}
+}
+
+func TestNvidiaEmbedUsesModelSpecificURL(t *testing.T) {
+	withSSRFBypass(t)
+	const apiKey = "nvapi-test"
+	const modelName = "nvidia/nv-embed-v1"
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/embeddings/nvidia/nv-embed-v1" {
+			t.Fatalf("path = %s, want model-specific embed endpoint", r.URL.Path)
+		}
+		_ = json.NewEncoder(w).Encode(map[string]interface{}{
+			"data": []map[string]interface{}{{"index": 0, "embedding": []float64{0.1, 0.2}}},
+		})
+	}))
+	defer server.Close()
+
+	withNvidiaProviderManager(t, []*Model{{
+		Name: modelName,
+		URL:  server.URL + "/v1/embeddings/nvidia/nv-embed-v1",
+	}})
+	driver := NewNvidiaModel(
+		map[string]string{"default": server.URL + "/v1"},
+		URLSuffix{Embedding: "embeddings"},
+	)
+	namePtr := modelName
+	got, err := driver.Embed(
+		context.Background(),
+		&namePtr,
+		[]string{"hello"},
+		&APIConfig{ApiKey: ptr(apiKey)},
+		nil,
+		nil,
+	)
+	if err != nil {
+		t.Fatalf("Embed() error = %v", err)
+	}
+	if len(got) != 1 || len(got[0].Embedding) != 2 {
+		t.Fatalf("embedding = %#v, want 1 vector of length 2", got)
+	}
+}
+
 func ptr[T any](value T) *T {
 	return &value
 }