highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-07-02 08:01:03 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix: add back bf16 support

Browse Source 
this was accidentally removed when moving fs/ggml from its previous
location
This commit is contained in:
Michael Yang
2025-02-25 09:18:44 -08:00
parent a499390648
commit b16367b4b2
2 changed files with 75 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
fs/ggml/ggml.go
View File

@ -207,11 +207,26 @@ func (t Tensor) block() (n int) {
func (t Tensor) blockSize() uint64 { func (t Tensor) blockSize() uint64 {
switch t.Kind { switch t.Kind {
case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16 case
0, // F32
1, // F16
24, // I8
25, // I16
26, // I32
27, // I64
28, // F64
30: // BF16
return 1 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, // Q4_0
3, // Q4_1
6, // Q5_0
7, // Q5_1
8, // Q8_0
9, // Q8_1
20: // IQ4_NL
return 32 return 32
default: // All others default:
return 256 return 256
} }
} }
@ -235,7 +250,7 @@ func (t Tensor) typeSize() uint64 {
case 8: // Q8_0 case 8: // Q8_0
return 2 + blockSize return 2 + blockSize
case 9: // Q8_1 case 9: // Q8_1
return 4 + 4 + blockSize return 2 + 2 + blockSize
case 10: // Q2_K case 10: // Q2_K
return blockSize/16 + blockSize/4 + 2 + 2 return blockSize/16 + blockSize/4 + 2 + 2
case 11: // Q3_K case 11: // Q3_K
@ -247,7 +262,7 @@ func (t Tensor) typeSize() uint64 {
case 14: // Q6_K case 14: // Q6_K
return blockSize/2 + blockSize/4 + blockSize/16 + 2 return blockSize/2 + blockSize/4 + blockSize/16 + 2
case 15: // Q8_K case 15: // Q8_K
return 2 + blockSize + 2*blockSize/16 return 4 + blockSize + 2*blockSize/16
case 16: // IQ2_XXS case 16: // IQ2_XXS
return 2 + 2*blockSize/8 return 2 + 2*blockSize/8
case 17: // IQ2_XS case 17: // IQ2_XS
@ -276,6 +291,8 @@ func (t Tensor) typeSize() uint64 {
return 8 return 8
case 29: // IQ1_M case 29: // IQ1_M
return blockSize/8 + blockSize/16 + blockSize/32 return blockSize/8 + blockSize/16 + blockSize/32
case 30: // BF16
return 2
default: default:
return 0 return 0
} }

53
fs/ggml/ggml_test.go
View File

@ -3,6 +3,7 @@ package ggml
import ( import (
"maps" "maps"
"slices" "slices"
"strconv"
"strings" "strings"
"testing" "testing"
@ -157,3 +158,55 @@ func TestTensorLayers(t *testing.T) {
}) })
} }
} }
// ref: https://github.com/ggml-org/llama.cpp/blob/a82c9e7c23ef6db48cebfa194dc9cebbc4ac3552/ggml/src/ggml.c#L572
func TestTensorTypes(t *testing.T) {
cases := []struct {
kind uint32
blockSize uint64
typeSize uint64
}{
{0, 1, 4},
{1, 1, 2},
{2, 32, 18},
{3, 32, 20},
{6, 32, 22},
{7, 32, 24},
{8, 32, 34},
{9, 32, 36},
{10, 256, 84},
{11, 256, 110},
{12, 256, 144},
{13, 256, 176},
{14, 256, 210},
{15, 256, 292},
{16, 256, 66},
{17, 256, 74},
{18, 256, 98},
{19, 256, 50},
{20, 32, 18},
{21, 256, 110},
{22, 256, 82},
{23, 256, 136},
{24, 1, 1},
{25, 1, 2},
{26, 1, 4},
{27, 1, 8},
{28, 1, 8},
{29, 256, 56},
{30, 1, 2},
}
for _, tt := range cases {
t.Run(strconv.Itoa(int(tt.kind)), func(t *testing.T) {
tensor := Tensor{Kind: tt.kind}
if tensor.blockSize() != tt.blockSize {
t.Errorf("unexpected block size: got=%d want=%d", tensor.blockSize(), tt.blockSize)
}
if tensor.typeSize() != tt.typeSize {
t.Errorf("unexpected type size: got=%d want=%d", tensor.typeSize(), tt.typeSize)
}
})
}
}