ggla.go 2.5 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140
  1. package llm
  2. import (
  3. "encoding/binary"
  4. "errors"
  5. "io"
  6. "slices"
  7. )
  8. type containerGGLA struct {
  9. version uint32
  10. }
  11. func (c *containerGGLA) Name() string {
  12. return "ggla"
  13. }
  14. func (c *containerGGLA) Decode(rs io.ReadSeeker) (model, error) {
  15. if err := binary.Read(rs, binary.LittleEndian, &c.version); err != nil {
  16. return nil, err
  17. }
  18. switch c.version {
  19. case 1:
  20. default:
  21. return nil, errors.New("invalid version")
  22. }
  23. model := newGGLA(c)
  24. err := model.decode(rs)
  25. return model, err
  26. }
  27. type ggla struct {
  28. *containerGGLA
  29. offset int64
  30. kv KV
  31. tensors []*Tensor
  32. }
  33. func newGGLA(container *containerGGLA) *ggla {
  34. return &ggla{
  35. containerGGLA: container,
  36. kv: make(KV),
  37. }
  38. }
  39. func (llm *ggla) KV() KV {
  40. return llm.kv
  41. }
  42. func (llm *ggla) Tensors() Tensors {
  43. return llm.tensors
  44. }
  45. func (llm *ggla) Offset() int64 {
  46. return llm.offset
  47. }
  48. func (llm *ggla) decode(rs io.ReadSeeker) error {
  49. var r uint32
  50. if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
  51. return err
  52. }
  53. llm.kv["r"] = r
  54. var alpha uint32
  55. if err := binary.Read(rs, binary.LittleEndian, &alpha); err != nil {
  56. return err
  57. }
  58. llm.kv["alpha"] = alpha
  59. offset, err := rs.Seek(0, io.SeekCurrent)
  60. if err != nil {
  61. return err
  62. }
  63. llm.offset = offset
  64. for {
  65. var dims uint32
  66. if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
  67. return err
  68. }
  69. var namesize uint32
  70. if err := binary.Read(rs, binary.LittleEndian, &namesize); err != nil {
  71. return err
  72. }
  73. var t Tensor
  74. if err := binary.Read(rs, binary.LittleEndian, &t.Kind); err != nil {
  75. return err
  76. }
  77. t.Shape = make([]uint64, dims)
  78. for i := 0; uint32(i) < dims; i++ {
  79. var shape32 uint32
  80. if err := binary.Read(rs, binary.LittleEndian, &shape32); err != nil {
  81. return err
  82. }
  83. t.Shape[i] = uint64(shape32)
  84. }
  85. // ggla tensor shape is reversed
  86. // ref: https://github.com/ggerganov/llama.cpp/blob/29ae62d2ae163e2b68aa0ad3bf2ab4636de0c957/convert-lora-to-ggml.py#L44
  87. slices.Reverse(t.Shape)
  88. name := make([]byte, namesize)
  89. if err := binary.Read(rs, binary.LittleEndian, &name); err != nil {
  90. return err
  91. }
  92. t.Name = string(name)
  93. offset, err := rs.Seek(0, io.SeekCurrent)
  94. if err != nil {
  95. return err
  96. }
  97. if _, err := rs.Seek((offset+31)&-32, io.SeekStart); err != nil {
  98. return err
  99. }
  100. offset, err = rs.Seek(0, io.SeekCurrent)
  101. if err != nil {
  102. return err
  103. }
  104. t.Offset = uint64(offset)
  105. if _, err := rs.Seek(int64(t.size()), io.SeekCurrent); err != nil {
  106. return err
  107. }
  108. llm.tensors = append(llm.tensors, &t)
  109. }
  110. }