mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

model: handle multiple eos tokens (#10577)

Browse Source 
* get eos_token_id from generation_config.json

* refactor

* include both ids and strings in trace

* comments

* remove special case for gemma3 special vocab (#10743)
This commit is contained in:
Michael Yang
2025-05-16 13:40:23 -07:00
committed by GitHub
parent 27da2cddc5
commit 333e360422
18 changed files with 282 additions and 182 deletions
Download Patch File Download Diff File Expand all files Collapse all files

126
model/process_text.go → model/bytepairencoding.go
View File

@@ -5,116 +5,13 @@ import (
"context"
"iter"
"log/slog"
"slices"
"strings"
"sync"
"github.com/dlclark/regexp2"
heap "github.com/emirpasic/gods/v2/trees/binaryheap"
"github.com/ollama/ollama/logutil"
)
type Special int32
const (
SpecialBOS Special = iota
SpecialEOS
)
const (
TOKEN_TYPE_NORMAL = iota + 1
TOKEN_TYPE_UNKNOWN
TOKEN_TYPE_CONTROL
TOKEN_TYPE_USER_DEFINED
TOKEN_TYPE_UNUSED
TOKEN_TYPE_BYTE
)
type TextProcessor interface {
Encode(s string, addSpecial bool) ([]int32, error)
Decode([]int32) (string, error)
Is(int32, Special) bool
Vocabulary() *Vocabulary
}
type Vocabulary struct {
Values []string
Types []int32
Scores []float32
Merges []string
BOS, EOS, EOT int32
AddBOS, AddEOS, AddEOT bool
specialOnce sync.Once
special []string
valuesOnce sync.Once
values map[string]int32
mergeOnce sync.Once
merge map[string]int32
}
func (v *Vocabulary) Is(id int32, special Special) bool {
switch special {
case SpecialBOS:
return id == v.BOS
case SpecialEOS:
return id == v.EOS || id == v.EOT
default:
return false
}
}
func (v *Vocabulary) Encode(s string) int32 {
v.valuesOnce.Do(func() {
v.values = make(map[string]int32, len(v.Values))
for i, value := range v.Values {
v.values[value] = int32(i)
}
})
if id, ok := v.values[s]; ok {
return id
}
return -1
}
func (v *Vocabulary) Decode(id int32) string {
return v.Values[id]
}
func (v *Vocabulary) SpecialVocabulary() []string {
v.specialOnce.Do(func() {
for i := range v.Values {
if slices.Contains([]int{105, 106}, i) {
v.special = append(v.special, v.Values[i])
} else if v.Types[i] == TOKEN_TYPE_CONTROL {
v.special = append(v.special, v.Values[i])
}
}
})
return v.special
}
func (v *Vocabulary) Merge(left, right string) int {
v.mergeOnce.Do(func() {
v.merge = make(map[string]int32, len(v.Merges))
for i, merge := range v.Merges {
v.merge[merge] = int32(i)
}
})
if id, ok := v.merge[left+" "+right]; ok {
return int(id)
}
return -1
}
type BytePairEncoding struct {
pre *regexp2.Regexp
vocab *Vocabulary
@@ -304,27 +201,12 @@ func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
}
}
slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "string", s, "ids", ids)
if addSpecial && len(ids) > 0 {
if bpe.vocab.AddBOS {
if ids[0] == bpe.vocab.BOS {
slog.Warn("adding bos token to prompt which already has it", "id", bpe.vocab.BOS)
}
slog.Debug("adding bos token to prompt", "id", bpe.vocab.BOS)
ids = append([]int32{bpe.vocab.BOS}, ids...)
}
if bpe.vocab.AddEOS {
if ids[len(ids)-1] == bpe.vocab.EOS {
slog.Warn("adding eos token to prompt which already has it", "id", bpe.vocab.EOS)
}
slog.Debug("adding eos token to prompt", "id", bpe.vocab.EOS)
ids = append(ids, bpe.vocab.EOS)
}
ids = bpe.vocab.addSpecials(ids)
}
slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "ids", ids)
return ids, nil
}
@@ -352,6 +234,6 @@ func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
}
}
slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "string", sb.String())
slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "ids", ids, "string", sb.String())
return sb.String(), nil
}

0
model/process_text_test.go → model/bytepairencoding_test.go
View File

11
model/models/gemma2/model.go
View File

@@ -43,10 +43,13 @@ func New(c fs.Config) (model.Model, error) {
Values: c.Strings("tokenizer.ggml.tokens"),
Scores: c.Floats("tokenizer.ggml.scores"),
Types: c.Ints("tokenizer.ggml.token_type"),
BOS: int32(c.Uint("tokenizer.ggml.bos_token_id")),
EOS: int32(c.Uint("tokenizer.ggml.eos_token_id")),
// TODO: set EOT to EOS otherwise 0 will stop generation
EOT: int32(c.Uint("tokenizer.ggml.eos_token_id")),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
BOS: []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
EOS: append(
[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
c.Ints("tokenizer.ggml.eos_token_ids")...,
),
},
),
Layers: make([]Layer, c.Uint("block_count")),

12
model/models/gemma3/model.go
View File

@@ -60,12 +60,16 @@ func New(c fs.Config) (model.Model, error) {
Values: c.Strings("tokenizer.ggml.tokens"),
Scores: c.Floats("tokenizer.ggml.scores"),
Types: c.Ints("tokenizer.ggml.token_type"),
BOS: int32(c.Uint("tokenizer.ggml.bos_token_id")),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
EOS: int32(1),
BOS: []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
EOT: int32(106),
AddEOT: c.Bool("tokenizer.ggml.add_eot_token", false),
EOS: append(
[]int32{
int32(c.Uint("tokenizer.ggml.eos_token_id")),
int32(c.Uint("tokenizer.ggml.eot_token_id", 106)),
},
c.Ints("tokenizer.ggml.eos_token_ids")...,
),
},
),
ImageProcessor: newImageProcessor(c),

10
model/models/llama/model.go
View File

@@ -43,13 +43,13 @@ func New(c fs.Config) (model.Model, error) {
Values: c.Strings("tokenizer.ggml.tokens"),
Types: c.Ints("tokenizer.ggml.token_type"),
Merges: c.Strings("tokenizer.ggml.merges"),
BOS: int32(c.Uint("tokenizer.ggml.bos_token_id")),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
EOS: int32(c.Uint("tokenizer.ggml.eos_token_id")),
BOS: []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
// TODO: set EOT to EOS otherwise 0 will stop generation
EOT: int32(c.Uint("tokenizer.ggml.eos_token_id")),
AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
EOS: append(
[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
c.Ints("tokenizer.ggml.eos_token_ids")...,
),
},
),
Layers: make([]Layer, c.Uint("block_count")),

10
model/models/llama4/model.go
View File

@@ -40,13 +40,13 @@ func New(c fs.Config) (model.Model, error) {
Values: c.Strings("tokenizer.ggml.tokens"),
Types: c.Ints("tokenizer.ggml.token_type"),
Merges: c.Strings("tokenizer.ggml.merges"),
BOS: int32(c.Uint("tokenizer.ggml.bos_token_id")),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
EOS: int32(c.Uint("tokenizer.ggml.eos_token_id")),
BOS: []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
// TODO: set EOT to EOS otherwise 0 will stop generation
EOT: int32(c.Uint("tokenizer.ggml.eos_token_id")),
AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
EOS: append(
[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
c.Ints("tokenizer.ggml.eos_token_ids")...,
),
},
),
ImageProcessor: newImageProcessor(c),

18
model/models/mistral3/model.go
View File

@@ -37,25 +37,25 @@ func New(c fs.Config) (model.Model, error) {
}
m := &Model{
TextModel: textModel,
VisionModel: newVisionModel(c),
ImageProcessor: newImageProcessor(c),
MultiModalProjector: newMultiModalProjector(c),
BytePairEncoding: model.NewBytePairEncoding(
c.String("tokenizer.ggml.pretokenizer", `[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]*[\p{Ll}\p{Lm}\p{Lo}\p{M}]+|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]+[\p{Ll}\p{Lm}\p{Lo}\p{M}]*|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n/]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
&model.Vocabulary{
Values: c.Strings("tokenizer.ggml.tokens"),
Types: c.Ints("tokenizer.ggml.token_type"),
Merges: c.Strings("tokenizer.ggml.merges"),
BOS: int32(c.Uint("tokenizer.ggml.bos_token_id", 1)),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
EOS: int32(c.Uint("tokenizer.ggml.eos_token_id", 2)),
BOS: []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
// TODO: set EOT to EOS otherwise 0 will stop generation
EOT: int32(c.Uint("tokenizer.ggml.eos_token_id")),
AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
EOS: append(
[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
c.Ints("tokenizer.ggml.eos_token_ids")...,
),
},
),
TextModel: textModel,
VisionModel: newVisionModel(c),
ImageProcessor: newImageProcessor(c),
MultiModalProjector: newMultiModalProjector(c),
}
m.Cache = kvcache.NewCausalCache(m.TextModel.Shift)

10
model/models/mllama/model.go
View File

@@ -38,13 +38,13 @@ func New(c fs.Config) (model.Model, error) {
Values: c.Strings("tokenizer.ggml.tokens"),
Types: c.Ints("tokenizer.ggml.token_type"),
Merges: c.Strings("tokenizer.ggml.merges"),
BOS: int32(c.Uint("tokenizer.ggml.bos_token_id")),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
EOS: int32(c.Uint("tokenizer.ggml.eos_token_id")),
BOS: []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
// TODO: set EOT to EOS otherwise 0 will stop generation
EOT: int32(c.Uint("tokenizer.ggml.eos_token_id")),
AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
EOS: append(
[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
c.Ints("tokenizer.ggml.eos_token_ids")...,
),
},
),
ImageProcessor: newImageProcessor(c),

11
model/models/qwen25vl/model.go
View File

@@ -34,12 +34,13 @@ func New(c fs.Config) (model.Model, error) {
Values: c.Strings("tokenizer.ggml.tokens"),
Types: c.Ints("tokenizer.ggml.token_type"),
Merges: c.Strings("tokenizer.ggml.merges"),
BOS: int32(c.Uint("tokenizer.ggml.bos_token_id")),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", false),
EOS: int32(c.Uint("tokenizer.ggml.eos_token_id")),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
BOS: []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
EOT: int32(c.Uint("tokenizer.ggml.eos_token_id")),
AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
EOS: append(
[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
c.Ints("tokenizer.ggml.eos_token_ids")...,
),
},
),
TextModel: NewTextModel(c),

23
model/process_text_spm.go → model/sentencepiece.go
View File

@@ -182,27 +182,12 @@ func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error)
}
}
slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "string", s, "ids", ids)
if addSpecial && len(ids) > 0 {
if spm.vocab.AddBOS {
if ids[0] == spm.vocab.BOS {
slog.Warn("adding bos token to prompt which already has it", "id", spm.vocab.BOS)
}
slog.Debug("adding bos token to prompt", "id", spm.vocab.BOS)
ids = append([]int32{spm.vocab.BOS}, ids...)
}
if spm.vocab.AddEOS {
if ids[len(ids)-1] == spm.vocab.EOS {
slog.Warn("adding eos token to prompt which already has it", "id", spm.vocab.EOS)
}
slog.Debug("adding eos token to prompt", "id", spm.vocab.EOS)
ids = append(ids, spm.vocab.EOS)
}
ids = spm.vocab.addSpecials(ids)
}
slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "ids", ids)
return ids, nil
}
@@ -261,6 +246,6 @@ func (spm SentencePieceModel) Decode(ids []int32) (string, error) {
}
}
slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "string", sb.String())
slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "ids", ids, "string", sb.String())
return sb.String(), nil
}

0
model/process_text_spm_test.go → model/sentencepiece_test.go
View File

17
model/textprocessor.go Normal file
View File

@@ -0,0 +1,17 @@
package model
const (
TOKEN_TYPE_NORMAL = iota + 1
TOKEN_TYPE_UNKNOWN
TOKEN_TYPE_CONTROL
TOKEN_TYPE_USER_DEFINED
TOKEN_TYPE_UNUSED
TOKEN_TYPE_BYTE
)
type TextProcessor interface {
Encode(s string, addSpecial bool) ([]int32, error)
Decode([]int32) (string, error)
Is(int32, Special) bool
Vocabulary() *Vocabulary
}

112
model/vocabulary.go Normal file
View File

@@ -0,0 +1,112 @@
package model
import (
"log/slog"
"slices"
"sync"
)
type Special int32
const (
SpecialBOS Special = iota
SpecialEOS
)
type Vocabulary struct {
Values []string
Types []int32
Scores []float32
Merges []string
BOS, EOS []int32
AddBOS, AddEOS bool
specialOnce sync.Once
special []string
valuesOnce sync.Once
values map[string]int32
mergeOnce sync.Once
merge map[string]int32
}
func (v *Vocabulary) Is(id int32, special Special) bool {
switch special {
case SpecialBOS:
return slices.Contains(v.BOS, id)
case SpecialEOS:
return slices.Contains(v.EOS, id)
default:
return false
}
}
func (v *Vocabulary) addSpecials(ids []int32) []int32 {
if v.AddBOS && len(v.BOS) > 0 {
if slices.Contains(v.BOS, ids[0]) {
slog.Warn("adding bos token to prompt which already has it", "id", v.BOS)
}
slog.Debug("adding bos token to prompt", "id", v.BOS)
ids = append([]int32{v.BOS[0]}, ids...)
}
if v.AddEOS && len(v.EOS) > 0 {
if slices.Contains(v.BOS, ids[len(ids)-1]) {
slog.Warn("adding eos token to prompt which already has it", "id", v.EOS)
}
slog.Debug("adding eos token to prompt", "id", v.EOS)
ids = append(ids, v.EOS[0])
}
return ids
}
func (v *Vocabulary) Encode(s string) int32 {
v.valuesOnce.Do(func() {
v.values = make(map[string]int32, len(v.Values))
for i, value := range v.Values {
v.values[value] = int32(i)
}
})
if id, ok := v.values[s]; ok {
return id
}
return -1
}
func (v *Vocabulary) Decode(id int32) string {
return v.Values[id]
}
func (v *Vocabulary) SpecialVocabulary() []string {
v.specialOnce.Do(func() {
for i := range v.Values {
if v.Types[i] == TOKEN_TYPE_CONTROL {
v.special = append(v.special, v.Values[i])
}
}
})
return v.special
}
func (v *Vocabulary) Merge(left, right string) int {
v.mergeOnce.Do(func() {
v.merge = make(map[string]int32, len(v.Merges))
for i, merge := range v.Merges {
v.merge[merge] = int32(i)
}
})
if id, ok := v.merge[left+" "+right]; ok {
return int(id)
}
return -1
}