|
| 1 | +package provider |
| 2 | + |
| 3 | +import ( |
| 4 | + "bytes" |
| 5 | + "encoding/json" |
| 6 | + "fmt" |
| 7 | + "net/http" |
| 8 | + "net/http/httptest" |
| 9 | + "strings" |
| 10 | + "testing" |
| 11 | + |
| 12 | + "github.com/coder/aibridge/config" |
| 13 | + "go.opentelemetry.io/otel/trace/noop" |
| 14 | + "golang.org/x/sync/errgroup" |
| 15 | +) |
| 16 | + |
| 17 | +type message struct { |
| 18 | + Role string |
| 19 | + Content string |
| 20 | +} |
| 21 | + |
| 22 | +type providerStrategy interface { |
| 23 | + DefaultModel() string |
| 24 | + formatMessages(messages []message) []any |
| 25 | + buildRequestBody(model string, messages []any, stream bool) map[string]any |
| 26 | +} |
| 27 | +type responsesProvider struct{} |
| 28 | + |
| 29 | +func (*responsesProvider) DefaultModel() string { |
| 30 | + return "gpt-5" |
| 31 | +} |
| 32 | + |
| 33 | +func (*responsesProvider) formatMessages(messages []message) []any { |
| 34 | + formatted := make([]any, 0, len(messages)) |
| 35 | + for _, msg := range messages { |
| 36 | + formatted = append(formatted, map[string]any{ |
| 37 | + "type": "message", |
| 38 | + "role": msg.Role, |
| 39 | + "content": msg.Content, |
| 40 | + }) |
| 41 | + } |
| 42 | + return formatted |
| 43 | +} |
| 44 | + |
| 45 | +func (*responsesProvider) buildRequestBody(model string, messages []any, stream bool) map[string]any { |
| 46 | + return map[string]any{ |
| 47 | + "model": model, |
| 48 | + "input": messages, |
| 49 | + "stream": stream, |
| 50 | + } |
| 51 | +} |
| 52 | + |
| 53 | +type chatCompletionsProvider struct{} |
| 54 | + |
| 55 | +func (*chatCompletionsProvider) DefaultModel() string { |
| 56 | + return "gpt-4" |
| 57 | +} |
| 58 | + |
| 59 | +func (*chatCompletionsProvider) formatMessages(messages []message) []any { |
| 60 | + formatted := make([]any, 0, len(messages)) |
| 61 | + for _, msg := range messages { |
| 62 | + formatted = append(formatted, map[string]string{ |
| 63 | + "role": msg.Role, |
| 64 | + "content": msg.Content, |
| 65 | + }) |
| 66 | + } |
| 67 | + return formatted |
| 68 | +} |
| 69 | + |
| 70 | +func (*chatCompletionsProvider) buildRequestBody(model string, messages []any, stream bool) map[string]any { |
| 71 | + return map[string]any{ |
| 72 | + "model": model, |
| 73 | + "messages": messages, |
| 74 | + "stream": stream, |
| 75 | + } |
| 76 | +} |
| 77 | + |
| 78 | +func generateConversation(provider providerStrategy, targetSize int, numMessages int) []any { |
| 79 | + if targetSize <= 0 { |
| 80 | + return nil |
| 81 | + } |
| 82 | + if numMessages < 1 { |
| 83 | + numMessages = 1 |
| 84 | + } |
| 85 | + |
| 86 | + roles := []string{"user", "assistant"} |
| 87 | + messages := make([]message, numMessages) |
| 88 | + for i := range messages { |
| 89 | + messages[i].Role = roles[i%2] |
| 90 | + } |
| 91 | + // Ensure last message is from user (required for LLM APIs). |
| 92 | + if messages[len(messages)-1].Role != "user" { |
| 93 | + messages[len(messages)-1].Role = "user" |
| 94 | + } |
| 95 | + |
| 96 | + overhead := measureJSONSize(provider.formatMessages(messages)) |
| 97 | + |
| 98 | + bytesPerMessage := targetSize - overhead |
| 99 | + if bytesPerMessage < 0 { |
| 100 | + bytesPerMessage = 0 |
| 101 | + } |
| 102 | + |
| 103 | + perMessage := bytesPerMessage / len(messages) |
| 104 | + remainder := bytesPerMessage % len(messages) |
| 105 | + |
| 106 | + for i := range messages { |
| 107 | + size := perMessage |
| 108 | + if i == len(messages)-1 { |
| 109 | + size += remainder |
| 110 | + } |
| 111 | + messages[i].Content = strings.Repeat("x", size) |
| 112 | + } |
| 113 | + |
| 114 | + return provider.formatMessages(messages) |
| 115 | +} |
| 116 | + |
| 117 | +func measureJSONSize(v any) int { |
| 118 | + data, err := json.Marshal(v) |
| 119 | + if err != nil { |
| 120 | + return 0 |
| 121 | + } |
| 122 | + return len(data) |
| 123 | +} |
| 124 | + |
| 125 | +// generateChatCompletionsPayload creates a JSON payload with the specified number of messages. |
| 126 | +// Messages alternate between user and assistant roles to simulate a conversation. |
| 127 | +func generateChatCompletionsPayload(payloadSize int, messageCount int, stream bool) []byte { |
| 128 | + provider := &chatCompletionsProvider{} |
| 129 | + messages := generateConversation(provider, payloadSize, messageCount) |
| 130 | + |
| 131 | + body := provider.buildRequestBody(provider.DefaultModel(), messages, stream) |
| 132 | + bodyBytes, err := json.Marshal(body) |
| 133 | + if err != nil { |
| 134 | + panic(err) |
| 135 | + } |
| 136 | + return bodyBytes |
| 137 | +} |
| 138 | + |
| 139 | +// generateResponsesPayload creates a JSON payload for the responses API with the specified number of input items. |
| 140 | +// Input items alternate between user and assistant roles to simulate a conversation. |
| 141 | +func generateResponsesPayload(payloadSize int, inputCount int, stream bool) []byte { |
| 142 | + provider := &responsesProvider{} |
| 143 | + inputs := generateConversation(provider, payloadSize, inputCount) |
| 144 | + |
| 145 | + body := provider.buildRequestBody(provider.DefaultModel(), inputs, stream) |
| 146 | + bodyBytes, err := json.Marshal(body) |
| 147 | + if err != nil { |
| 148 | + panic(err) |
| 149 | + } |
| 150 | + return bodyBytes |
| 151 | +} |
| 152 | + |
| 153 | +func BenchmarkOpenAI_CreateInterceptor_ChatCompletions(b *testing.B) { |
| 154 | + provider := NewOpenAI(config.OpenAI{ |
| 155 | + BaseURL: "https://api.openai.com/v1/", |
| 156 | + Key: "test-key", |
| 157 | + }) |
| 158 | + |
| 159 | + tracer := noop.NewTracerProvider().Tracer("test") |
| 160 | + messagesPerRequest := 50 |
| 161 | + requestCount := 100 |
| 162 | + maxConcurrentRequests := 10 |
| 163 | + payloadSizes := []int{2000, 10000, 50000, 100000, 2000000} |
| 164 | + for _, payloadSize := range payloadSizes { |
| 165 | + for _, stream := range []bool{true, false} { |
| 166 | + payload := generateChatCompletionsPayload(payloadSize, messagesPerRequest, stream) |
| 167 | + name := fmt.Sprintf("stream=%t/payloadSize=%d/requests=%d", stream, payloadSize, requestCount) |
| 168 | + |
| 169 | + b.Run(name, func(b *testing.B) { |
| 170 | + b.ResetTimer() |
| 171 | + for range b.N { |
| 172 | + eg := errgroup.Group{} |
| 173 | + eg.SetLimit(maxConcurrentRequests) |
| 174 | + for i := 0; i < requestCount; i++ { |
| 175 | + eg.Go(func() error { |
| 176 | + req := httptest.NewRequest(http.MethodPost, routeChatCompletions, bytes.NewReader(payload)) |
| 177 | + w := httptest.NewRecorder() |
| 178 | + _, err := provider.CreateInterceptor(w, req, tracer) |
| 179 | + if err != nil { |
| 180 | + return err |
| 181 | + } |
| 182 | + return nil |
| 183 | + }) |
| 184 | + } |
| 185 | + } |
| 186 | + }) |
| 187 | + } |
| 188 | + } |
| 189 | +} |
| 190 | + |
| 191 | +func BenchmarkOpenAI_CreateInterceptor_Responses(b *testing.B) { |
| 192 | + provider := NewOpenAI(config.OpenAI{ |
| 193 | + BaseURL: "https://api.openai.com/v1/", |
| 194 | + Key: "test-key", |
| 195 | + }) |
| 196 | + |
| 197 | + tracer := noop.NewTracerProvider().Tracer("test") |
| 198 | + messagesPerRequest := 50 |
| 199 | + requestCount := 100 |
| 200 | + maxConcurrentRequests := 10 |
| 201 | + // payloadSizes := []int{2000, 10000, 50000, 100000, 2000000} |
| 202 | + payloadSizes := []int{2000000} |
| 203 | + for _, payloadSize := range payloadSizes { |
| 204 | + for _, stream := range []bool{true, false} { |
| 205 | + payload := generateResponsesPayload(payloadSize, messagesPerRequest, stream) |
| 206 | + name := fmt.Sprintf("stream=%t/payloadSize=%d/requests=%d", stream, payloadSize, requestCount) |
| 207 | + |
| 208 | + b.Run(name, func(b *testing.B) { |
| 209 | + b.ResetTimer() |
| 210 | + for range b.N { |
| 211 | + eg := errgroup.Group{} |
| 212 | + eg.SetLimit(maxConcurrentRequests) |
| 213 | + for i := 0; i < requestCount; i++ { |
| 214 | + eg.Go(func() error { |
| 215 | + req := httptest.NewRequest(http.MethodPost, routeResponses, bytes.NewReader(payload)) |
| 216 | + w := httptest.NewRecorder() |
| 217 | + interceptor, err := provider.CreateInterceptor(w, req, tracer) |
| 218 | + if err != nil { |
| 219 | + return err |
| 220 | + } |
| 221 | + err = interceptor.ProcessRequest(w, req) |
| 222 | + if err != nil { |
| 223 | + return err |
| 224 | + } |
| 225 | + return nil |
| 226 | + }) |
| 227 | + } |
| 228 | + } |
| 229 | + }) |
| 230 | + } |
| 231 | + } |
| 232 | +} |
0 commit comments