Merge c24c79c7d08ff0ec29608c1baaf2769772f1c7e2 into 67691e410db7a50b07a64858820b14de9aa91314
This commit is contained in:
commit
71896db44f
@ -37,6 +37,11 @@ const (
|
||||
fileTypeIQ2_M
|
||||
fileTypeIQ1_M
|
||||
fileTypeBF16
|
||||
fileTypeQ4_0_4_4
|
||||
fileTypeQ4_0_4_8
|
||||
fileTypeQ4_0_8_8
|
||||
fileTypeTQ1_0
|
||||
fileTypeTQ2_0
|
||||
|
||||
fileTypeUnknown
|
||||
)
|
||||
@ -103,6 +108,16 @@ func ParseFileType(s string) (fileType, error) {
|
||||
return fileTypeIQ1_M, nil
|
||||
case "BF16":
|
||||
return fileTypeBF16, nil
|
||||
case "Q4_0_4_4":
|
||||
return fileTypeQ4_0_4_4, nil
|
||||
case "Q4_0_4_8":
|
||||
return fileTypeQ4_0_4_8, nil
|
||||
case "Q4_0_8_8":
|
||||
return fileTypeQ4_0_8_8, nil
|
||||
case "TQ1_0":
|
||||
return fileTypeTQ1_0, nil
|
||||
case "TQ2_0":
|
||||
return fileTypeTQ2_0, nil
|
||||
default:
|
||||
return fileTypeUnknown, fmt.Errorf("unknown fileType: %s", s)
|
||||
}
|
||||
@ -170,6 +185,16 @@ func (t fileType) String() string {
|
||||
return "IQ1_M"
|
||||
case fileTypeBF16:
|
||||
return "BF16"
|
||||
case fileTypeQ4_0_4_4:
|
||||
return "Q4_0_4_4"
|
||||
case fileTypeQ4_0_4_8:
|
||||
return "Q4_0_4_8"
|
||||
case fileTypeQ4_0_8_8:
|
||||
return "Q4_0_8_8"
|
||||
case fileTypeTQ1_0:
|
||||
return "TQ1_0"
|
||||
case fileTypeTQ2_0:
|
||||
return "TQ2_0"
|
||||
default:
|
||||
return "unknown"
|
||||
}
|
||||
|
12
llm/ggml.go
12
llm/ggml.go
@ -188,7 +188,7 @@ func (t Tensor) blockSize() uint64 {
|
||||
switch t.Kind {
|
||||
case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16
|
||||
return 1
|
||||
case 2, 3, 4, 5, 6, 7, 8, 9, 20: // Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, Q8_1, IQ4_NL
|
||||
case 2, 3, 4, 5, 6, 7, 8, 9, 20, 31, 32, 33: // Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, Q8_1, IQ4_NL, Q4_0_4_4, Q4_0_4_8, Q4_0_8_8
|
||||
return 32
|
||||
default: // All others
|
||||
return 256
|
||||
@ -257,6 +257,16 @@ func (t Tensor) typeSize() uint64 {
|
||||
return blockSize/8 + blockSize/16 + blockSize/32
|
||||
case 30: // BF16
|
||||
return 2
|
||||
case 31: // Q4_0_4_4
|
||||
return 2 + blockSize/2
|
||||
case 32: // Q4_0_4_8
|
||||
return 2 + blockSize/2
|
||||
case 33: // Q4_0_8_8
|
||||
return 2 + blockSize/2
|
||||
case 34: // TQ1_0
|
||||
return 2 + blockSize/64 + (blockSize-4*blockSize/64)/5
|
||||
case 35: // TQ2_0
|
||||
return 2 + blockSize/4
|
||||
default:
|
||||
return 0
|
||||
}
|
||||
|
Loading…
x
Reference in New Issue
Block a user