types.go 4.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177
  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. Username string `json:"username"`
  38. Password string `json:"password"`
  39. }
  40. type ProgressResponse struct {
  41. Status string `json:"status"`
  42. Digest string `json:"digest,omitempty"`
  43. Total int `json:"total,omitempty"`
  44. Completed int `json:"completed,omitempty"`
  45. }
  46. type PushRequest struct {
  47. Name string `json:"name"`
  48. Username string `json:"username"`
  49. Password string `json:"password"`
  50. }
  51. type ListResponse struct {
  52. Models []ListResponseModel `json:"models"`
  53. }
  54. type ListResponseModel struct {
  55. Name string `json:"name"`
  56. ModifiedAt time.Time `json:"modified_at"`
  57. Size int `json:"size"`
  58. }
  59. type GenerateResponse struct {
  60. Model string `json:"model"`
  61. CreatedAt time.Time `json:"created_at"`
  62. Response string `json:"response,omitempty"`
  63. Done bool `json:"done"`
  64. Context []int `json:"context,omitempty"`
  65. TotalDuration time.Duration `json:"total_duration,omitempty"`
  66. PromptEvalCount int `json:"prompt_eval_count,omitempty"`
  67. PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
  68. EvalCount int `json:"eval_count,omitempty"`
  69. EvalDuration time.Duration `json:"eval_duration,omitempty"`
  70. }
  71. func (r *GenerateResponse) Summary() {
  72. if r.TotalDuration > 0 {
  73. fmt.Fprintf(os.Stderr, "total duration: %v\n", r.TotalDuration)
  74. }
  75. if r.PromptEvalCount > 0 {
  76. fmt.Fprintf(os.Stderr, "prompt eval count: %d token(s)\n", r.PromptEvalCount)
  77. }
  78. if r.PromptEvalDuration > 0 {
  79. fmt.Fprintf(os.Stderr, "prompt eval duration: %s\n", r.PromptEvalDuration)
  80. fmt.Fprintf(os.Stderr, "prompt eval rate: %.2f tokens/s\n", float64(r.PromptEvalCount)/r.PromptEvalDuration.Seconds())
  81. }
  82. if r.EvalCount > 0 {
  83. fmt.Fprintf(os.Stderr, "eval count: %d token(s)\n", r.EvalCount)
  84. }
  85. if r.EvalDuration > 0 {
  86. fmt.Fprintf(os.Stderr, "eval duration: %s\n", r.EvalDuration)
  87. fmt.Fprintf(os.Stderr, "eval rate: %.2f tokens/s\n", float64(r.EvalCount)/r.EvalDuration.Seconds())
  88. }
  89. }
  90. type Options struct {
  91. Seed int `json:"seed,omitempty"`
  92. // Backend options
  93. UseNUMA bool `json:"numa,omitempty"`
  94. // Model options
  95. NumCtx int `json:"num_ctx,omitempty"`
  96. NumBatch int `json:"num_batch,omitempty"`
  97. NumGPU int `json:"num_gpu,omitempty"`
  98. MainGPU int `json:"main_gpu,omitempty"`
  99. LowVRAM bool `json:"low_vram,omitempty"`
  100. F16KV bool `json:"f16_kv,omitempty"`
  101. LogitsAll bool `json:"logits_all,omitempty"`
  102. VocabOnly bool `json:"vocab_only,omitempty"`
  103. UseMMap bool `json:"use_mmap,omitempty"`
  104. UseMLock bool `json:"use_mlock,omitempty"`
  105. EmbeddingOnly bool `json:"embedding_only,omitempty"`
  106. // Predict options
  107. RepeatLastN int `json:"repeat_last_n,omitempty"`
  108. RepeatPenalty float32 `json:"repeat_penalty,omitempty"`
  109. FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
  110. PresencePenalty float32 `json:"presence_penalty,omitempty"`
  111. Temperature float32 `json:"temperature,omitempty"`
  112. TopK int `json:"top_k,omitempty"`
  113. TopP float32 `json:"top_p,omitempty"`
  114. TFSZ float32 `json:"tfs_z,omitempty"`
  115. TypicalP float32 `json:"typical_p,omitempty"`
  116. Mirostat int `json:"mirostat,omitempty"`
  117. MirostatTau float32 `json:"mirostat_tau,omitempty"`
  118. MirostatEta float32 `json:"mirostat_eta,omitempty"`
  119. NumThread int `json:"num_thread,omitempty"`
  120. }
  121. func DefaultOptions() Options {
  122. return Options{
  123. Seed: -1,
  124. UseNUMA: false,
  125. NumCtx: 2048,
  126. NumBatch: 512,
  127. NumGPU: 1,
  128. LowVRAM: false,
  129. F16KV: true,
  130. UseMMap: true,
  131. UseMLock: false,
  132. RepeatLastN: 512,
  133. RepeatPenalty: 1.1,
  134. FrequencyPenalty: 0.0,
  135. PresencePenalty: 0.0,
  136. Temperature: 0.8,
  137. TopK: 40,
  138. TopP: 0.9,
  139. TFSZ: 1.0,
  140. TypicalP: 1.0,
  141. Mirostat: 0,
  142. MirostatTau: 5.0,
  143. MirostatEta: 0.1,
  144. NumThread: runtime.NumCPU(),
  145. }
  146. }