ggml.go 4.3 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241
  1. package llm
  2. import (
  3. "encoding/binary"
  4. "errors"
  5. "io"
  6. )
  7. type GGML struct {
  8. container
  9. model
  10. Size int64
  11. }
  12. const (
  13. fileTypeF32 uint32 = iota
  14. fileTypeF16
  15. fileTypeQ4_0
  16. fileTypeQ4_1
  17. fileTypeQ4_1_F16
  18. fileTypeQ8_0 uint32 = iota + 2
  19. fileTypeQ5_0
  20. fileTypeQ5_1
  21. fileTypeQ2_K
  22. fileTypeQ3_K_S
  23. fileTypeQ3_K_M
  24. fileTypeQ3_K_L
  25. fileTypeQ4_K_S
  26. fileTypeQ4_K_M
  27. fileTypeQ5_K_S
  28. fileTypeQ5_K_M
  29. fileTypeQ6_K
  30. )
  31. func fileType(fileType uint32) string {
  32. switch fileType {
  33. case fileTypeF32:
  34. return "F32"
  35. case fileTypeF16:
  36. return "F16"
  37. case fileTypeQ4_0:
  38. return "Q4_0"
  39. case fileTypeQ4_1:
  40. return "Q4_1"
  41. case fileTypeQ4_1_F16:
  42. return "Q4_1_F16"
  43. case fileTypeQ8_0:
  44. return "Q8_0"
  45. case fileTypeQ5_0:
  46. return "Q5_0"
  47. case fileTypeQ5_1:
  48. return "Q5_1"
  49. case fileTypeQ2_K:
  50. return "Q2_K"
  51. case fileTypeQ3_K_S:
  52. return "Q3_K_S"
  53. case fileTypeQ3_K_M:
  54. return "Q3_K_M"
  55. case fileTypeQ3_K_L:
  56. return "Q3_K_L"
  57. case fileTypeQ4_K_S:
  58. return "Q4_K_S"
  59. case fileTypeQ4_K_M:
  60. return "Q4_K_M"
  61. case fileTypeQ5_K_S:
  62. return "Q5_K_S"
  63. case fileTypeQ5_K_M:
  64. return "Q5_K_M"
  65. case fileTypeQ6_K:
  66. return "Q6_K"
  67. default:
  68. return "unknown"
  69. }
  70. }
  71. type model interface {
  72. ModelFamily() string
  73. ModelType() string
  74. FileType() string
  75. NumLayers() int64
  76. }
  77. type container interface {
  78. Name() string
  79. Decode(*readSeekOffset) (model, error)
  80. }
  81. type containerGGML struct{}
  82. func (c *containerGGML) Name() string {
  83. return "ggml"
  84. }
  85. func (c *containerGGML) Decode(ro *readSeekOffset) (model, error) {
  86. return nil, nil
  87. }
  88. type containerGGMF struct {
  89. version uint32
  90. }
  91. func (c *containerGGMF) Name() string {
  92. return "ggmf"
  93. }
  94. func (c *containerGGMF) Decode(ro *readSeekOffset) (model, error) {
  95. var version uint32
  96. binary.Read(ro, binary.LittleEndian, &version)
  97. switch version {
  98. case 1:
  99. default:
  100. return nil, errors.New("invalid version")
  101. }
  102. c.version = version
  103. return nil, nil
  104. }
  105. type containerGGJT struct {
  106. version uint32
  107. }
  108. func (c *containerGGJT) Name() string {
  109. return "ggjt"
  110. }
  111. func (c *containerGGJT) Decode(ro *readSeekOffset) (model, error) {
  112. var version uint32
  113. binary.Read(ro, binary.LittleEndian, &version)
  114. switch version {
  115. case 1, 2, 3:
  116. default:
  117. return nil, errors.New("invalid version")
  118. }
  119. c.version = version
  120. // different model types may have different layouts for hyperparameters
  121. var llama llamaModel
  122. binary.Read(ro, binary.LittleEndian, &llama.hyperparameters)
  123. return &llama, nil
  124. }
  125. type containerLORA struct {
  126. version uint32
  127. }
  128. func (c *containerLORA) Name() string {
  129. return "ggla"
  130. }
  131. func (c *containerLORA) Decode(ro *readSeekOffset) (model, error) {
  132. var version uint32
  133. binary.Read(ro, binary.LittleEndian, &version)
  134. switch version {
  135. case 1:
  136. default:
  137. return nil, errors.New("invalid version")
  138. }
  139. c.version = version
  140. return nil, nil
  141. }
  142. const (
  143. // Magic constant for `ggml` files (unversioned).
  144. FILE_MAGIC_GGML = 0x67676d6c
  145. // Magic constant for `ggml` files (versioned, ggmf).
  146. FILE_MAGIC_GGMF = 0x67676d66
  147. // Magic constant for `ggml` files (versioned, ggjt).
  148. FILE_MAGIC_GGJT = 0x67676a74
  149. // Magic constant for `ggla` files (LoRA adapter).
  150. FILE_MAGIC_GGLA = 0x67676C61
  151. // Magic constant for `gguf` files (versioned, gguf)
  152. FILE_MAGIC_GGUF_LE = 0x46554747
  153. FILE_MAGIC_GGUF_BE = 0x47475546
  154. )
  155. func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
  156. ro := readSeekOffset{ReadSeeker: r}
  157. var magic uint32
  158. if err := binary.Read(&ro, binary.LittleEndian, &magic); err != nil {
  159. return nil, err
  160. }
  161. var c container
  162. switch magic {
  163. case FILE_MAGIC_GGML:
  164. c = &containerGGML{}
  165. case FILE_MAGIC_GGMF:
  166. c = &containerGGMF{}
  167. case FILE_MAGIC_GGJT:
  168. c = &containerGGJT{}
  169. case FILE_MAGIC_GGLA:
  170. c = &containerLORA{}
  171. case FILE_MAGIC_GGUF_LE:
  172. c = &containerGGUF{bo: binary.LittleEndian}
  173. case FILE_MAGIC_GGUF_BE:
  174. c = &containerGGUF{bo: binary.BigEndian}
  175. default:
  176. return nil, errors.New("invalid file magic")
  177. }
  178. model, err := c.Decode(&ro)
  179. if err != nil {
  180. return nil, err
  181. }
  182. // final model type
  183. return &GGML{
  184. container: c,
  185. model: model,
  186. Size: ro.offset,
  187. }, nil
  188. }
  189. type readSeekOffset struct {
  190. io.ReadSeeker
  191. offset int64
  192. }
  193. func (rso *readSeekOffset) Seek(offset int64, whence int) (int64, error) {
  194. offset, err := rso.ReadSeeker.Seek(offset, whence)
  195. if err != nil {
  196. return 0, err
  197. }
  198. rso.offset = offset
  199. return offset, nil
  200. }
  201. func (rso *readSeekOffset) Read(p []byte) (int, error) {
  202. n, err := rso.ReadSeeker.Read(p)
  203. rso.offset += int64(n)
  204. return n, err
  205. }