highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-03-26 17:51:48 +01:00
Code Issues Packages Projects Releases Wiki Activity

wip

Browse Source
This commit is contained in:
jmorganca 2025-03-16 21:58:21 -04:00
parent afb34b0e60
commit 8025781dce
5 changed files with 25 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
convert/convert_mistral.go
View File

@ -63,10 +63,8 @@ func (p *mistralModel) KV(t *Tokenizer) ggml.KV {
kv["mistral.feed_forward_length"] = cmp.Or(p.IntermediateSize, p.NInner)
}
if headCount := cmp.Or(p.NumAttentionHeads, p.NHead); headCount > 0 {
kv["mistral.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
kv["mistral.rope.dimension_count"] = p.HiddenSize / headCount
}
kv["mistral.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
kv["mistral.rope.dimension_count"] = p.HiddenSize / cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)
if p.RopeTheta > 0 {
kv["mistral.rope.freq_base"] = p.RopeTheta
@ -136,6 +134,13 @@ func (p *mistralModel) Tensors(ts []Tensor) []ggml.Tensor {
t.SetRepacker(p.repack)
}
if strings.HasPrefix(t.Name(), "patch_merger.") ||
strings.HasPrefix(t.Name(), "pre_mm_projector_output_norm.") ||
strings.HasPrefix(t.Name(), "vision_encoder.") ||
strings.HasPrefix(t.Name(), "vision_language_adapter.") {
continue
}
out = append(out, ggml.Tensor{
Name: t.Name(),
Kind: t.Kind(),
@ -149,7 +154,7 @@ func (p *mistralModel) Tensors(ts []Tensor) []ggml.Tensor {
func (p *mistralModel) Replacements() []string {
return []string{
"tok_embeddings.weight", "token_embd",
"tok_embeddings", "token_embd",
"norm", "output_norm",
"layers", "blk",
"attention_norm", "attn_norm",
@ -160,6 +165,8 @@ func (p *mistralModel) Replacements() []string {
"feed_forward.w1", "ffn_gate",
"feed_forward.w2", "ffn_down",
"feed_forward.w3", "ffn_up",
"ffn_norm", "ffn_norm",
"output", "output",
}
}

5
model/models/mistral/model.go
View File

@ -37,7 +37,10 @@ func New(c ml.Config) (model.Model, error) {
m := Model{
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+`),
// TODO: need to set this in the conversion for mistral:
// 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+
c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
// 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.Uints("tokenizer.ggml.token_type"),

1
model/models/models.go
View File

@ -4,5 +4,6 @@ import (
_ "github.com/ollama/ollama/model/models/gemma2"
_ "github.com/ollama/ollama/model/models/gemma3"
_ "github.com/ollama/ollama/model/models/llama"
_ "github.com/ollama/ollama/model/models/mistral"
_ "github.com/ollama/ollama/model/models/mllama"
)

4
model/process_text.go
View File

@ -263,6 +263,10 @@ func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
continue
}
if id := bpe.vocab.Encode(pair.value); id < 0 {
continue
}
merges[pair.a].runes = append(left.runes, right.runes...)
merges[pair.b].runes = nil

4
runner/ollamarunner/runner.go
View File

@ -179,6 +179,10 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, *
return nil, nil, err
}
for _, t := range tokens {
decoded, _ := s.model.(model.TextProcessor).Decode([]int32{t})
fmt.Println("token", t, "decoded", decoded)
}
for _, t := range tokens {
inputs = append(inputs, input.Input{Token: t})
}