types.go 4.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159
  1. package api
  2. import (
  3. "fmt"
  4. "os"
  5. "runtime"
  6. "time"
  7. )
  8. type GenerateRequest struct {
  9. Model string `json:"model"`
  10. Prompt string `json:"prompt"`
  11. Context []int `json:"context,omitempty"`
  12. Options `json:"options"`
  13. }
  14. type CreateRequest struct {
  15. Name string `json:"name"`
  16. Path string `json:"path"`
  17. }
  18. type CreateProgress struct {
  19. Status string `json:"status"`
  20. }
  21. type PullRequest struct {
  22. Name string `json:"name"`
  23. Username string `json:"username"`
  24. Password string `json:"password"`
  25. }
  26. type PullProgress struct {
  27. Status string `json:"status"`
  28. Digest string `json:"digest,omitempty"`
  29. Total int `json:"total,omitempty"`
  30. Completed int `json:"completed,omitempty"`
  31. Percent float64 `json:"percent,omitempty"`
  32. }
  33. type PushRequest struct {
  34. Name string `json:"name"`
  35. Username string `json:"username"`
  36. Password string `json:"password"`
  37. }
  38. type PushProgress struct {
  39. Status string `json:"status"`
  40. Digest string `json:"digest,omitempty"`
  41. Total int `json:"total,omitempty"`
  42. Completed int `json:"completed,omitempty"`
  43. Percent float64 `json:"percent,omitempty"`
  44. }
  45. type GenerateResponse struct {
  46. Model string `json:"model"`
  47. CreatedAt time.Time `json:"created_at"`
  48. Response string `json:"response,omitempty"`
  49. Done bool `json:"done"`
  50. Context []int `json:"context,omitempty"`
  51. TotalDuration time.Duration `json:"total_duration,omitempty"`
  52. PromptEvalCount int `json:"prompt_eval_count,omitempty"`
  53. PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
  54. EvalCount int `json:"eval_count,omitempty"`
  55. EvalDuration time.Duration `json:"eval_duration,omitempty"`
  56. }
  57. func (r *GenerateResponse) Summary() {
  58. if r.TotalDuration > 0 {
  59. fmt.Fprintf(os.Stderr, "total duration: %v\n", r.TotalDuration)
  60. }
  61. if r.PromptEvalCount > 0 {
  62. fmt.Fprintf(os.Stderr, "prompt eval count: %d token(s)\n", r.PromptEvalCount)
  63. }
  64. if r.PromptEvalDuration > 0 {
  65. fmt.Fprintf(os.Stderr, "prompt eval duration: %s\n", r.PromptEvalDuration)
  66. fmt.Fprintf(os.Stderr, "prompt eval rate: %.2f tokens/s\n", float64(r.PromptEvalCount)/r.PromptEvalDuration.Seconds())
  67. }
  68. if r.EvalCount > 0 {
  69. fmt.Fprintf(os.Stderr, "eval count: %d token(s)\n", r.EvalCount)
  70. }
  71. if r.EvalDuration > 0 {
  72. fmt.Fprintf(os.Stderr, "eval duration: %s\n", r.EvalDuration)
  73. fmt.Fprintf(os.Stderr, "eval rate: %.2f tokens/s\n", float64(r.EvalCount)/r.EvalDuration.Seconds())
  74. }
  75. }
  76. type Options struct {
  77. Seed int `json:"seed,omitempty"`
  78. // Backend options
  79. UseNUMA bool `json:"numa,omitempty"`
  80. // Model options
  81. NumCtx int `json:"num_ctx,omitempty"`
  82. NumBatch int `json:"num_batch,omitempty"`
  83. NumGPU int `json:"num_gpu,omitempty"`
  84. MainGPU int `json:"main_gpu,omitempty"`
  85. LowVRAM bool `json:"low_vram,omitempty"`
  86. F16KV bool `json:"f16_kv,omitempty"`
  87. LogitsAll bool `json:"logits_all,omitempty"`
  88. VocabOnly bool `json:"vocab_only,omitempty"`
  89. UseMMap bool `json:"use_mmap,omitempty"`
  90. UseMLock bool `json:"use_mlock,omitempty"`
  91. EmbeddingOnly bool `json:"embedding_only,omitempty"`
  92. // Predict options
  93. RepeatLastN int `json:"repeat_last_n,omitempty"`
  94. RepeatPenalty float32 `json:"repeat_penalty,omitempty"`
  95. FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
  96. PresencePenalty float32 `json:"presence_penalty,omitempty"`
  97. Temperature float32 `json:"temperature,omitempty"`
  98. TopK int `json:"top_k,omitempty"`
  99. TopP float32 `json:"top_p,omitempty"`
  100. TFSZ float32 `json:"tfs_z,omitempty"`
  101. TypicalP float32 `json:"typical_p,omitempty"`
  102. Mirostat int `json:"mirostat,omitempty"`
  103. MirostatTau float32 `json:"mirostat_tau,omitempty"`
  104. MirostatEta float32 `json:"mirostat_eta,omitempty"`
  105. NumThread int `json:"num_thread,omitempty"`
  106. }
  107. func DefaultOptions() Options {
  108. return Options{
  109. Seed: -1,
  110. UseNUMA: false,
  111. NumCtx: 2048,
  112. NumBatch: 512,
  113. NumGPU: 1,
  114. LowVRAM: false,
  115. F16KV: true,
  116. UseMMap: true,
  117. UseMLock: false,
  118. RepeatLastN: 512,
  119. RepeatPenalty: 1.1,
  120. FrequencyPenalty: 0.0,
  121. PresencePenalty: 0.0,
  122. Temperature: 0.8,
  123. TopK: 40,
  124. TopP: 0.9,
  125. TFSZ: 1.0,
  126. TypicalP: 1.0,
  127. Mirostat: 0,
  128. MirostatTau: 5.0,
  129. MirostatEta: 0.1,
  130. NumThread: runtime.NumCPU(),
  131. }
  132. }