routes.go 7.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354
  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. "dario.cat/mergo"
  17. "github.com/gin-contrib/cors"
  18. "github.com/gin-gonic/gin"
  19. "github.com/jmorganca/ollama/api"
  20. "github.com/jmorganca/ollama/llama"
  21. )
  22. var loaded struct {
  23. mu sync.Mutex
  24. llm *llama.LLM
  25. expireAt time.Time
  26. expireTimer *time.Timer
  27. digest string
  28. options api.Options
  29. }
  30. func GenerateHandler(c *gin.Context) {
  31. loaded.mu.Lock()
  32. defer loaded.mu.Unlock()
  33. checkpointStart := time.Now()
  34. var req api.GenerateRequest
  35. if err := c.ShouldBindJSON(&req); err != nil {
  36. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  37. return
  38. }
  39. model, err := GetModel(req.Model)
  40. if err != nil {
  41. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  42. return
  43. }
  44. if model.Digest != loaded.digest || !reflect.DeepEqual(loaded.options, req.Options) {
  45. if loaded.llm != nil {
  46. loaded.llm.Close()
  47. loaded.llm = nil
  48. }
  49. opts := model.Options
  50. if err := mergo.Merge(&opts, req.Options, mergo.WithOverride); err != nil {
  51. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  52. return
  53. }
  54. llm, err := llama.New(model.ModelPath, opts)
  55. if err != nil {
  56. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  57. return
  58. }
  59. loaded.llm = llm
  60. loaded.digest = model.Digest
  61. }
  62. sessionDuration := 5 * time.Minute
  63. loaded.expireAt = time.Now().Add(sessionDuration)
  64. if loaded.expireTimer == nil {
  65. loaded.expireTimer = time.AfterFunc(sessionDuration, func() {
  66. loaded.mu.Lock()
  67. defer loaded.mu.Unlock()
  68. if time.Now().Before(loaded.expireAt) {
  69. return
  70. }
  71. if loaded.llm == nil {
  72. return
  73. }
  74. loaded.llm.Close()
  75. loaded.llm = nil
  76. })
  77. }
  78. loaded.expireTimer.Reset(sessionDuration)
  79. checkpointLoaded := time.Now()
  80. prompt, err := model.Prompt(req)
  81. if err != nil {
  82. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  83. return
  84. }
  85. ch := make(chan any)
  86. go func() {
  87. defer close(ch)
  88. fn := func(r api.GenerateResponse) {
  89. loaded.expireAt = time.Now().Add(sessionDuration)
  90. loaded.expireTimer.Reset(sessionDuration)
  91. r.Model = req.Model
  92. r.CreatedAt = time.Now().UTC()
  93. if r.Done {
  94. r.TotalDuration = time.Since(checkpointStart)
  95. r.LoadDuration = checkpointLoaded.Sub(checkpointStart)
  96. }
  97. ch <- r
  98. }
  99. if err := loaded.llm.Predict(req.Context, prompt, fn); err != nil {
  100. ch <- gin.H{"error": err.Error()}
  101. }
  102. }()
  103. streamResponse(c, ch)
  104. }
  105. func PullModelHandler(c *gin.Context) {
  106. var req api.PullRequest
  107. if err := c.ShouldBindJSON(&req); err != nil {
  108. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  109. return
  110. }
  111. ch := make(chan any)
  112. go func() {
  113. defer close(ch)
  114. fn := func(r api.ProgressResponse) {
  115. ch <- r
  116. }
  117. regOpts := &RegistryOptions{
  118. Insecure: req.Insecure,
  119. Username: req.Username,
  120. Password: req.Password,
  121. }
  122. if err := PullModel(req.Name, regOpts, fn); err != nil {
  123. ch <- gin.H{"error": err.Error()}
  124. }
  125. }()
  126. streamResponse(c, ch)
  127. }
  128. func PushModelHandler(c *gin.Context) {
  129. var req api.PushRequest
  130. if err := c.ShouldBindJSON(&req); err != nil {
  131. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  132. return
  133. }
  134. ch := make(chan any)
  135. go func() {
  136. defer close(ch)
  137. fn := func(r api.ProgressResponse) {
  138. ch <- r
  139. }
  140. regOpts := &RegistryOptions{
  141. Insecure: req.Insecure,
  142. Username: req.Username,
  143. Password: req.Password,
  144. }
  145. if err := PushModel(req.Name, regOpts, fn); err != nil {
  146. ch <- gin.H{"error": err.Error()}
  147. }
  148. }()
  149. streamResponse(c, ch)
  150. }
  151. func CreateModelHandler(c *gin.Context) {
  152. var req api.CreateRequest
  153. if err := c.ShouldBindJSON(&req); err != nil {
  154. c.JSON(http.StatusBadRequest, gin.H{"message": err.Error()})
  155. return
  156. }
  157. ch := make(chan any)
  158. go func() {
  159. defer close(ch)
  160. fn := func(resp api.ProgressResponse) {
  161. ch <- resp
  162. }
  163. if err := CreateModel(req.Name, req.Path, fn); err != nil {
  164. ch <- gin.H{"error": err.Error()}
  165. }
  166. }()
  167. streamResponse(c, ch)
  168. }
  169. func DeleteModelHandler(c *gin.Context) {
  170. var req api.DeleteRequest
  171. if err := c.ShouldBindJSON(&req); err != nil {
  172. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  173. return
  174. }
  175. if err := DeleteModel(req.Name); err != nil {
  176. if os.IsNotExist(err) {
  177. c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Name)})
  178. } else {
  179. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  180. }
  181. return
  182. }
  183. }
  184. func ListModelsHandler(c *gin.Context) {
  185. var models []api.ListResponseModel
  186. fp, err := GetManifestPath()
  187. if err != nil {
  188. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  189. return
  190. }
  191. err = filepath.Walk(fp, func(path string, info os.FileInfo, err error) error {
  192. if err != nil {
  193. if errors.Is(err, os.ErrNotExist) {
  194. log.Printf("manifest file does not exist: %s", fp)
  195. return nil
  196. }
  197. return err
  198. }
  199. if !info.IsDir() {
  200. fi, err := os.Stat(path)
  201. if err != nil {
  202. log.Printf("skipping file: %s", fp)
  203. return nil
  204. }
  205. path := path[len(fp)+1:]
  206. slashIndex := strings.LastIndex(path, "/")
  207. if slashIndex == -1 {
  208. return nil
  209. }
  210. tag := path[:slashIndex] + ":" + path[slashIndex+1:]
  211. mp := ParseModelPath(tag)
  212. manifest, err := GetManifest(mp)
  213. if err != nil {
  214. log.Printf("skipping file: %s", fp)
  215. return nil
  216. }
  217. model := api.ListResponseModel{
  218. Name: mp.GetShortTagname(),
  219. Size: manifest.GetTotalSize(),
  220. ModifiedAt: fi.ModTime(),
  221. }
  222. models = append(models, model)
  223. }
  224. return nil
  225. })
  226. if err != nil {
  227. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  228. return
  229. }
  230. c.JSON(http.StatusOK, api.ListResponse{Models: models})
  231. }
  232. func CopyModelHandler(c *gin.Context) {
  233. var req api.CopyRequest
  234. if err := c.ShouldBindJSON(&req); err != nil {
  235. c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
  236. return
  237. }
  238. if err := CopyModel(req.Source, req.Destination); err != nil {
  239. if os.IsNotExist(err) {
  240. c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Source)})
  241. } else {
  242. c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
  243. }
  244. return
  245. }
  246. }
  247. func Serve(ln net.Listener) error {
  248. config := cors.DefaultConfig()
  249. config.AllowWildcard = true
  250. // only allow http/https from localhost
  251. config.AllowOrigins = []string{
  252. "http://localhost",
  253. "http://localhost:*",
  254. "https://localhost",
  255. "https://localhost:*",
  256. "http://127.0.0.1",
  257. "http://127.0.0.1:*",
  258. "https://127.0.0.1",
  259. "https://127.0.0.1:*",
  260. }
  261. r := gin.Default()
  262. r.Use(cors.New(config))
  263. r.GET("/", func(c *gin.Context) {
  264. c.String(http.StatusOK, "Ollama is running")
  265. })
  266. r.POST("/api/pull", PullModelHandler)
  267. r.POST("/api/generate", GenerateHandler)
  268. r.POST("/api/create", CreateModelHandler)
  269. r.POST("/api/push", PushModelHandler)
  270. r.POST("/api/copy", CopyModelHandler)
  271. r.GET("/api/tags", ListModelsHandler)
  272. r.DELETE("/api/delete", DeleteModelHandler)
  273. log.Printf("Listening on %s", ln.Addr())
  274. s := &http.Server{
  275. Handler: r,
  276. }
  277. return s.Serve(ln)
  278. }
  279. func streamResponse(c *gin.Context, ch chan any) {
  280. c.Stream(func(w io.Writer) bool {
  281. val, ok := <-ch
  282. if !ok {
  283. return false
  284. }
  285. bts, err := json.Marshal(val)
  286. if err != nil {
  287. log.Printf("streamResponse: json.Marshal failed with %s", err)
  288. return false
  289. }
  290. bts = append(bts, '\n')
  291. if _, err := w.Write(bts); err != nil {
  292. log.Printf("streamResponse: w.Write failed with %s", err)
  293. return false
  294. }
  295. return true
  296. })
  297. }