types.go 4.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179
  1. package api
  2. import (
  3. "fmt"
  4. "os"
  5. "runtime"
  6. "time"
  7. )
  8. type StatusError struct {
  9. StatusCode int
  10. Status string
  11. Message string
  12. }
  13. func (e StatusError) Error() string {
  14. if e.Message != "" {
  15. return fmt.Sprintf("%s: %s", e.Status, e.Message)
  16. }
  17. return e.Status
  18. }
  19. type GenerateRequest struct {
  20. Model string `json:"model"`
  21. Prompt string `json:"prompt"`
  22. Context []int `json:"context,omitempty"`
  23. Options `json:"options"`
  24. }
  25. type CreateRequest struct {
  26. Name string `json:"name"`
  27. Path string `json:"path"`
  28. }
  29. type CreateProgress struct {
  30. Status string `json:"status"`
  31. }
  32. type DeleteRequest struct {
  33. Name string `json:"name"`
  34. }
  35. type PullRequest struct {
  36. Name string `json:"name"`
  37. Insecure bool `json:"insecure,omitempty"`
  38. Username string `json:"username"`
  39. Password string `json:"password"`
  40. }
  41. type ProgressResponse struct {
  42. Status string `json:"status"`
  43. Digest string `json:"digest,omitempty"`
  44. Total int `json:"total,omitempty"`
  45. Completed int `json:"completed,omitempty"`
  46. }
  47. type PushRequest struct {
  48. Name string `json:"name"`
  49. Insecure bool `json:"insecure,omitempty"`
  50. Username string `json:"username"`
  51. Password string `json:"password"`
  52. }
  53. type ListResponse struct {
  54. Models []ListResponseModel `json:"models"`
  55. }
  56. type ListResponseModel struct {
  57. Name string `json:"name"`
  58. ModifiedAt time.Time `json:"modified_at"`
  59. Size int `json:"size"`
  60. }
  61. type GenerateResponse struct {
  62. Model string `json:"model"`
  63. CreatedAt time.Time `json:"created_at"`
  64. Response string `json:"response,omitempty"`
  65. Done bool `json:"done"`
  66. Context []int `json:"context,omitempty"`
  67. TotalDuration time.Duration `json:"total_duration,omitempty"`
  68. PromptEvalCount int `json:"prompt_eval_count,omitempty"`
  69. PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
  70. EvalCount int `json:"eval_count,omitempty"`
  71. EvalDuration time.Duration `json:"eval_duration,omitempty"`
  72. }
  73. func (r *GenerateResponse) Summary() {
  74. if r.TotalDuration > 0 {
  75. fmt.Fprintf(os.Stderr, "total duration: %v\n", r.TotalDuration)
  76. }
  77. if r.PromptEvalCount > 0 {
  78. fmt.Fprintf(os.Stderr, "prompt eval count: %d token(s)\n", r.PromptEvalCount)
  79. }
  80. if r.PromptEvalDuration > 0 {
  81. fmt.Fprintf(os.Stderr, "prompt eval duration: %s\n", r.PromptEvalDuration)
  82. fmt.Fprintf(os.Stderr, "prompt eval rate: %.2f tokens/s\n", float64(r.PromptEvalCount)/r.PromptEvalDuration.Seconds())
  83. }
  84. if r.EvalCount > 0 {
  85. fmt.Fprintf(os.Stderr, "eval count: %d token(s)\n", r.EvalCount)
  86. }
  87. if r.EvalDuration > 0 {
  88. fmt.Fprintf(os.Stderr, "eval duration: %s\n", r.EvalDuration)
  89. fmt.Fprintf(os.Stderr, "eval rate: %.2f tokens/s\n", float64(r.EvalCount)/r.EvalDuration.Seconds())
  90. }
  91. }
  92. type Options struct {
  93. Seed int `json:"seed,omitempty"`
  94. // Backend options
  95. UseNUMA bool `json:"numa,omitempty"`
  96. // Model options
  97. NumCtx int `json:"num_ctx,omitempty"`
  98. NumBatch int `json:"num_batch,omitempty"`
  99. NumGPU int `json:"num_gpu,omitempty"`
  100. MainGPU int `json:"main_gpu,omitempty"`
  101. LowVRAM bool `json:"low_vram,omitempty"`
  102. F16KV bool `json:"f16_kv,omitempty"`
  103. LogitsAll bool `json:"logits_all,omitempty"`
  104. VocabOnly bool `json:"vocab_only,omitempty"`
  105. UseMMap bool `json:"use_mmap,omitempty"`
  106. UseMLock bool `json:"use_mlock,omitempty"`
  107. EmbeddingOnly bool `json:"embedding_only,omitempty"`
  108. // Predict options
  109. RepeatLastN int `json:"repeat_last_n,omitempty"`
  110. RepeatPenalty float32 `json:"repeat_penalty,omitempty"`
  111. FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
  112. PresencePenalty float32 `json:"presence_penalty,omitempty"`
  113. Temperature float32 `json:"temperature,omitempty"`
  114. TopK int `json:"top_k,omitempty"`
  115. TopP float32 `json:"top_p,omitempty"`
  116. TFSZ float32 `json:"tfs_z,omitempty"`
  117. TypicalP float32 `json:"typical_p,omitempty"`
  118. Mirostat int `json:"mirostat,omitempty"`
  119. MirostatTau float32 `json:"mirostat_tau,omitempty"`
  120. MirostatEta float32 `json:"mirostat_eta,omitempty"`
  121. NumThread int `json:"num_thread,omitempty"`
  122. }
  123. func DefaultOptions() Options {
  124. return Options{
  125. Seed: -1,
  126. UseNUMA: false,
  127. NumCtx: 2048,
  128. NumBatch: 512,
  129. NumGPU: 1,
  130. LowVRAM: false,
  131. F16KV: true,
  132. UseMMap: true,
  133. UseMLock: false,
  134. RepeatLastN: 512,
  135. RepeatPenalty: 1.1,
  136. FrequencyPenalty: 0.0,
  137. PresencePenalty: 0.0,
  138. Temperature: 0.8,
  139. TopK: 40,
  140. TopP: 0.9,
  141. TFSZ: 1.0,
  142. TypicalP: 1.0,
  143. Mirostat: 0,
  144. MirostatTau: 5.0,
  145. MirostatEta: 0.1,
  146. NumThread: runtime.NumCPU(),
  147. }
  148. }