types.go 4.5 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173
  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 PullRequest struct {
  33. Name string `json:"name"`
  34. Username string `json:"username"`
  35. Password string `json:"password"`
  36. }
  37. type ProgressResponse struct {
  38. Status string `json:"status"`
  39. Digest string `json:"digest,omitempty"`
  40. Total int `json:"total,omitempty"`
  41. Completed int `json:"completed,omitempty"`
  42. }
  43. type PushRequest struct {
  44. Name string `json:"name"`
  45. Username string `json:"username"`
  46. Password string `json:"password"`
  47. }
  48. type ListResponse struct {
  49. Models []ListResponseModel `json:"models"`
  50. }
  51. type ListResponseModel struct {
  52. Name string `json:"name"`
  53. ModifiedAt time.Time `json:"modified_at"`
  54. Size int `json:"size"`
  55. }
  56. type GenerateResponse struct {
  57. Model string `json:"model"`
  58. CreatedAt time.Time `json:"created_at"`
  59. Response string `json:"response,omitempty"`
  60. Done bool `json:"done"`
  61. Context []int `json:"context,omitempty"`
  62. TotalDuration time.Duration `json:"total_duration,omitempty"`
  63. PromptEvalCount int `json:"prompt_eval_count,omitempty"`
  64. PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
  65. EvalCount int `json:"eval_count,omitempty"`
  66. EvalDuration time.Duration `json:"eval_duration,omitempty"`
  67. }
  68. func (r *GenerateResponse) Summary() {
  69. if r.TotalDuration > 0 {
  70. fmt.Fprintf(os.Stderr, "total duration: %v\n", r.TotalDuration)
  71. }
  72. if r.PromptEvalCount > 0 {
  73. fmt.Fprintf(os.Stderr, "prompt eval count: %d token(s)\n", r.PromptEvalCount)
  74. }
  75. if r.PromptEvalDuration > 0 {
  76. fmt.Fprintf(os.Stderr, "prompt eval duration: %s\n", r.PromptEvalDuration)
  77. fmt.Fprintf(os.Stderr, "prompt eval rate: %.2f tokens/s\n", float64(r.PromptEvalCount)/r.PromptEvalDuration.Seconds())
  78. }
  79. if r.EvalCount > 0 {
  80. fmt.Fprintf(os.Stderr, "eval count: %d token(s)\n", r.EvalCount)
  81. }
  82. if r.EvalDuration > 0 {
  83. fmt.Fprintf(os.Stderr, "eval duration: %s\n", r.EvalDuration)
  84. fmt.Fprintf(os.Stderr, "eval rate: %.2f tokens/s\n", float64(r.EvalCount)/r.EvalDuration.Seconds())
  85. }
  86. }
  87. type Options struct {
  88. Seed int `json:"seed,omitempty"`
  89. // Backend options
  90. UseNUMA bool `json:"numa,omitempty"`
  91. // Model options
  92. NumCtx int `json:"num_ctx,omitempty"`
  93. NumBatch int `json:"num_batch,omitempty"`
  94. NumGPU int `json:"num_gpu,omitempty"`
  95. MainGPU int `json:"main_gpu,omitempty"`
  96. LowVRAM bool `json:"low_vram,omitempty"`
  97. F16KV bool `json:"f16_kv,omitempty"`
  98. LogitsAll bool `json:"logits_all,omitempty"`
  99. VocabOnly bool `json:"vocab_only,omitempty"`
  100. UseMMap bool `json:"use_mmap,omitempty"`
  101. UseMLock bool `json:"use_mlock,omitempty"`
  102. EmbeddingOnly bool `json:"embedding_only,omitempty"`
  103. // Predict options
  104. RepeatLastN int `json:"repeat_last_n,omitempty"`
  105. RepeatPenalty float32 `json:"repeat_penalty,omitempty"`
  106. FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
  107. PresencePenalty float32 `json:"presence_penalty,omitempty"`
  108. Temperature float32 `json:"temperature,omitempty"`
  109. TopK int `json:"top_k,omitempty"`
  110. TopP float32 `json:"top_p,omitempty"`
  111. TFSZ float32 `json:"tfs_z,omitempty"`
  112. TypicalP float32 `json:"typical_p,omitempty"`
  113. Mirostat int `json:"mirostat,omitempty"`
  114. MirostatTau float32 `json:"mirostat_tau,omitempty"`
  115. MirostatEta float32 `json:"mirostat_eta,omitempty"`
  116. NumThread int `json:"num_thread,omitempty"`
  117. }
  118. func DefaultOptions() Options {
  119. return Options{
  120. Seed: -1,
  121. UseNUMA: false,
  122. NumCtx: 2048,
  123. NumBatch: 512,
  124. NumGPU: 1,
  125. LowVRAM: false,
  126. F16KV: true,
  127. UseMMap: true,
  128. UseMLock: false,
  129. RepeatLastN: 512,
  130. RepeatPenalty: 1.1,
  131. FrequencyPenalty: 0.0,
  132. PresencePenalty: 0.0,
  133. Temperature: 0.8,
  134. TopK: 40,
  135. TopP: 0.9,
  136. TFSZ: 1.0,
  137. TypicalP: 1.0,
  138. Mirostat: 0,
  139. MirostatTau: 5.0,
  140. MirostatEta: 0.1,
  141. NumThread: runtime.NumCPU(),
  142. }
  143. }