pali112
/
ollama
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fixed converter

This commit is contained in:
nicole pardal 2025-12-10 16:11:52 -08:00
parent b6f769ae60
commit 29a2d6d931
1 changed files with 54 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

56
convert/convert_olmo.go
View File

@ -6,6 +6,16 @@ import (
"github.com/ollama/ollama/fs/ggml" "github.com/ollama/ollama/fs/ggml"
) )
type ropeScaling struct {
Factor float32 `json:"factor"`
OriginalMaxPositionEmbeds uint32 `json:"original_max_position_embeddings"`
AttentionFactor float32 `json:"attention_factor"`
BetaFast float32 `json:"beta_fast"`
BetaSlow float32 `json:"beta_slow"`
RopeType string `json:"rope_type"`
ExtrapolationFactor float32 `json:"extrapolation_factor"`
}
type olmoModel struct { type olmoModel struct {
ModelParameters ModelParameters
@ -17,6 +27,7 @@ type olmoModel struct {
MaxPositionEmbeddings uint32 `json:"max_position_embeddings"` MaxPositionEmbeddings uint32 `json:"max_position_embeddings"`
RMSNormEPS float32 `json:"rms_norm_eps"` RMSNormEPS float32 `json:"rms_norm_eps"`
RopeTheta float32 `json:"rope_theta"` RopeTheta float32 `json:"rope_theta"`
RopeScaling *ropeScaling `json:"rope_scaling"`
ClampKQV float32 `json:"f_clamp_kqv"` ClampKQV float32 `json:"f_clamp_kqv"`
SlidingWindow uint32 `json:"sliding_window"` SlidingWindow uint32 `json:"sliding_window"`
LayerTypes []string `json:"layer_types"` LayerTypes []string `json:"layer_types"`
@ -26,34 +37,53 @@ var _ ModelConverter = (*olmoModel)(nil)
func (p *olmoModel) KV(t *Tokenizer) ggml.KV { func (p *olmoModel) KV(t *Tokenizer) ggml.KV {
kv := p.ModelParameters.KV(t) kv := p.ModelParameters.KV(t)
kv["general.architecture"] = "olmo" kv["general.architecture"] = "olmo2"
kv["olmo.block_count"] = p.NumHiddenLayers kv["olmo2.block_count"] = p.NumHiddenLayers
kv["olmo.context_length"] = p.MaxPositionEmbeddings kv["olmo2.context_length"] = p.MaxPositionEmbeddings
kv["olmo.embedding_length"] = p.HiddenSize kv["olmo2.embedding_length"] = p.HiddenSize
kv["olmo.feed_forward_length"] = p.IntermediateSize kv["olmo2.feed_forward_length"] = p.IntermediateSize
kv["olmo.attention.head_count"] = p.NumAttentionHeads kv["olmo2.attention.head_count"] = p.NumAttentionHeads
kv["olmo.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads) kv["olmo2.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
if p.RopeTheta > 0 { if p.RopeTheta > 0 {
kv["olmo.rope.freq_base"] = p.RopeTheta kv["olmo2.rope.freq_base"] = p.RopeTheta
} else { } else {
kv["olmo.rope.freq_base"] = float32(10000.0) kv["olmo2.rope.freq_base"] = float32(10000.0)
}
if p.RopeScaling != nil {
if p.RopeScaling.Factor > 0 {
kv["olmo2.rope.scaling.factor"] = p.RopeScaling.Factor
}
if p.RopeScaling.OriginalMaxPositionEmbeds > 0 {
kv["olmo2.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeds
}
if p.RopeScaling.AttentionFactor > 0 {
kv["olmo2.rope.scaling.attn_factor"] = p.RopeScaling.AttentionFactor
}
if p.RopeScaling.RopeType != "" {
kv["olmo2.rope.scaling.type"] = p.RopeScaling.RopeType
}
} }
if p.RMSNormEPS > 0 { if p.RMSNormEPS > 0 {
kv["olmo.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS kv["olmo2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
} }
if p.ClampKQV > 0 { if p.ClampKQV > 0 {
kv["olmo.attention.clamp_kqv"] = p.ClampKQV kv["olmo2.attention.clamp_kqv"] = p.ClampKQV
} }
if p.SlidingWindow > 0 { if p.SlidingWindow > 0 {
kv["olmo.attention.sliding_window"] = p.SlidingWindow kv["olmo2.attention.sliding_window"] = p.SlidingWindow
} }
if len(p.LayerTypes) > 0 { if len(p.LayerTypes) > 0 {
kv["olmo.attention.layer_types"] = p.LayerTypes slidingPattern := make([]bool, len(p.LayerTypes))
for i, layerType := range p.LayerTypes {
slidingPattern[i] = (layerType == "sliding_attention")
}
kv["olmo2.attention.sliding_window_pattern"] = slidingPattern
} }
return kv return kv