routes.go 8.1 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374
  1. package server
  2. import (
  3. "encoding/json"
  4. "errors"
  5. "fmt"
  6. "io"
  7. "log"
  8. "net"
  9. "net/http"
  10. "os"
  11. "path/filepath"
  12. "reflect"
  13. "strings"
  14. "sync"
  15. "time"
  16. "github.com/gin-contrib/cors"
  17. "github.com/gin-gonic/gin"
  18. "github.com/jmorganca/ollama/api"
  19. "github.com/jmorganca/ollama/llama"
  20. "github.com/jmorganca/ollama/vector"
  21. )
  22. var loaded struct {
  23. mu sync.Mutex
  24. llm *llama.LLM
  25. Embeddings []vector.Embedding
  26. expireAt time.Time
  27. expireTimer *time.Timer
  28. digest string
  29. options api.Options
  30. }
  31. func GenerateHandler(c *gin.Context) {
  32. loaded.mu.Lock()
  33. defer loaded.mu.Unlock()
  34. checkpointStart := time.Now()
  35. var req api.GenerateRequest
  36. if err := c.ShouldBindJSON(&req); err != nil {
  37. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  38. return
  39. }
  40. model, err := GetModel(req.Model)
  41. if err != nil {
  42. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  43. return
  44. }
  45. opts := api.DefaultOptions()
  46. if err := opts.FromMap(model.Options); err != nil {
  47. log.Printf("could not load model options: %v", err)
  48. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  49. return
  50. }
  51. if err := opts.FromMap(req.Options); err != nil {
  52. log.Printf("could not merge model options: %v", err)
  53. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  54. return
  55. }
  56. if model.Digest != loaded.digest || !reflect.DeepEqual(loaded.options, opts) {
  57. if loaded.llm != nil {
  58. loaded.llm.Close()
  59. loaded.llm = nil
  60. loaded.digest = ""
  61. }
  62. if model.Embeddings != nil && len(model.Embeddings) > 0 {
  63. opts.EmbeddingOnly = true // this is requried to generate embeddings, completions will still work
  64. loaded.Embeddings = model.Embeddings
  65. }
  66. llm, err := llama.New(model.ModelPath, opts)
  67. if err != nil {
  68. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  69. return
  70. }
  71. loaded.llm = llm
  72. loaded.digest = model.Digest
  73. loaded.options = opts
  74. }
  75. sessionDuration := 5 * time.Minute
  76. loaded.expireAt = time.Now().Add(sessionDuration)
  77. if loaded.expireTimer == nil {
  78. loaded.expireTimer = time.AfterFunc(sessionDuration, func() {
  79. loaded.mu.Lock()
  80. defer loaded.mu.Unlock()
  81. if time.Now().Before(loaded.expireAt) {
  82. return
  83. }
  84. if loaded.llm == nil {
  85. return
  86. }
  87. loaded.llm.Close()
  88. loaded.llm = nil
  89. loaded.digest = ""
  90. })
  91. }
  92. loaded.expireTimer.Reset(sessionDuration)
  93. checkpointLoaded := time.Now()
  94. prompt, err := model.Prompt(req)
  95. if err != nil {
  96. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  97. return
  98. }
  99. ch := make(chan any)
  100. go func() {
  101. defer close(ch)
  102. fn := func(r api.GenerateResponse) {
  103. loaded.expireAt = time.Now().Add(sessionDuration)
  104. loaded.expireTimer.Reset(sessionDuration)
  105. r.Model = req.Model
  106. r.CreatedAt = time.Now().UTC()
  107. if r.Done {
  108. r.TotalDuration = time.Since(checkpointStart)
  109. r.LoadDuration = checkpointLoaded.Sub(checkpointStart)
  110. }
  111. ch <- r
  112. }
  113. if err := loaded.llm.Predict(req.Context, prompt, fn); err != nil {
  114. ch <- gin.H{"error": err.Error()}
  115. }
  116. }()
  117. streamResponse(c, ch)
  118. }
  119. func PullModelHandler(c *gin.Context) {
  120. var req api.PullRequest
  121. if err := c.ShouldBindJSON(&req); err != nil {
  122. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  123. return
  124. }
  125. ch := make(chan any)
  126. go func() {
  127. defer close(ch)
  128. fn := func(r api.ProgressResponse) {
  129. ch <- r
  130. }
  131. regOpts := &RegistryOptions{
  132. Insecure: req.Insecure,
  133. Username: req.Username,
  134. Password: req.Password,
  135. }
  136. if err := PullModel(req.Name, regOpts, fn); err != nil {
  137. ch <- gin.H{"error": err.Error()}
  138. }
  139. }()
  140. streamResponse(c, ch)
  141. }
  142. func PushModelHandler(c *gin.Context) {
  143. var req api.PushRequest
  144. if err := c.ShouldBindJSON(&req); err != nil {
  145. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  146. return
  147. }
  148. ch := make(chan any)
  149. go func() {
  150. defer close(ch)
  151. fn := func(r api.ProgressResponse) {
  152. ch <- r
  153. }
  154. regOpts := &RegistryOptions{
  155. Insecure: req.Insecure,
  156. Username: req.Username,
  157. Password: req.Password,
  158. }
  159. if err := PushModel(req.Name, regOpts, fn); err != nil {
  160. ch <- gin.H{"error": err.Error()}
  161. }
  162. }()
  163. streamResponse(c, ch)
  164. }
  165. func CreateModelHandler(c *gin.Context) {
  166. var req api.CreateRequest
  167. if err := c.ShouldBindJSON(&req); err != nil {
  168. c.JSON(http.StatusBadRequest, gin.H{"message": err.Error()})
  169. return
  170. }
  171. ch := make(chan any)
  172. go func() {
  173. defer close(ch)
  174. fn := func(resp api.ProgressResponse) {
  175. ch <- resp
  176. }
  177. if err := CreateModel(req.Name, req.Path, fn); err != nil {
  178. ch <- gin.H{"error": err.Error()}
  179. }
  180. }()
  181. streamResponse(c, ch)
  182. }
  183. func DeleteModelHandler(c *gin.Context) {
  184. var req api.DeleteRequest
  185. if err := c.ShouldBindJSON(&req); err != nil {
  186. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  187. return
  188. }
  189. if err := DeleteModel(req.Name); err != nil {
  190. if os.IsNotExist(err) {
  191. c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Name)})
  192. } else {
  193. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  194. }
  195. return
  196. }
  197. }
  198. func ListModelsHandler(c *gin.Context) {
  199. var models []api.ListResponseModel
  200. fp, err := GetManifestPath()
  201. if err != nil {
  202. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  203. return
  204. }
  205. err = filepath.Walk(fp, func(path string, info os.FileInfo, err error) error {
  206. if err != nil {
  207. if errors.Is(err, os.ErrNotExist) {
  208. log.Printf("manifest file does not exist: %s", fp)
  209. return nil
  210. }
  211. return err
  212. }
  213. if !info.IsDir() {
  214. fi, err := os.Stat(path)
  215. if err != nil {
  216. log.Printf("skipping file: %s", fp)
  217. return nil
  218. }
  219. path := path[len(fp)+1:]
  220. slashIndex := strings.LastIndex(path, "/")
  221. if slashIndex == -1 {
  222. return nil
  223. }
  224. tag := path[:slashIndex] + ":" + path[slashIndex+1:]
  225. mp := ParseModelPath(tag)
  226. manifest, err := GetManifest(mp)
  227. if err != nil {
  228. log.Printf("skipping file: %s", fp)
  229. return nil
  230. }
  231. model := api.ListResponseModel{
  232. Name: mp.GetShortTagname(),
  233. Size: manifest.GetTotalSize(),
  234. ModifiedAt: fi.ModTime(),
  235. }
  236. models = append(models, model)
  237. }
  238. return nil
  239. })
  240. if err != nil {
  241. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  242. return
  243. }
  244. c.JSON(http.StatusOK, api.ListResponse{Models: models})
  245. }
  246. func CopyModelHandler(c *gin.Context) {
  247. var req api.CopyRequest
  248. if err := c.ShouldBindJSON(&req); err != nil {
  249. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  250. return
  251. }
  252. if err := CopyModel(req.Source, req.Destination); err != nil {
  253. if os.IsNotExist(err) {
  254. c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Source)})
  255. } else {
  256. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  257. }
  258. return
  259. }
  260. }
  261. func Serve(ln net.Listener, extraOrigins []string) error {
  262. config := cors.DefaultConfig()
  263. config.AllowWildcard = true
  264. // only allow http/https from localhost
  265. allowedOrigins := []string{
  266. "http://localhost",
  267. "http://localhost:*",
  268. "https://localhost",
  269. "https://localhost:*",
  270. "http://127.0.0.1",
  271. "http://127.0.0.1:*",
  272. "https://127.0.0.1",
  273. "https://127.0.0.1:*",
  274. }
  275. allowedOrigins = append(allowedOrigins, extraOrigins...)
  276. config.AllowOrigins = allowedOrigins
  277. r := gin.Default()
  278. r.Use(cors.New(config))
  279. r.GET("/", func(c *gin.Context) {
  280. c.String(http.StatusOK, "Ollama is running")
  281. })
  282. r.HEAD("/", func(c *gin.Context) {
  283. c.Status(http.StatusOK)
  284. })
  285. r.POST("/api/pull", PullModelHandler)
  286. r.POST("/api/generate", GenerateHandler)
  287. r.POST("/api/create", CreateModelHandler)
  288. r.POST("/api/push", PushModelHandler)
  289. r.POST("/api/copy", CopyModelHandler)
  290. r.GET("/api/tags", ListModelsHandler)
  291. r.DELETE("/api/delete", DeleteModelHandler)
  292. log.Printf("Listening on %s", ln.Addr())
  293. s := &http.Server{
  294. Handler: r,
  295. }
  296. return s.Serve(ln)
  297. }
  298. func streamResponse(c *gin.Context, ch chan any) {
  299. c.Stream(func(w io.Writer) bool {
  300. val, ok := <-ch
  301. if !ok {
  302. return false
  303. }
  304. bts, err := json.Marshal(val)
  305. if err != nil {
  306. log.Printf("streamResponse: json.Marshal failed with %s", err)
  307. return false
  308. }
  309. bts = append(bts, '\n')
  310. if _, err := w.Write(bts); err != nil {
  311. log.Printf("streamResponse: w.Write failed with %s", err)
  312. return false
  313. }
  314. return true
  315. })
  316. }