Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.
gemma3: make embedding non-causal (#12297)
Michael Yangec9eb28f4c34
diff --git a/model/models/gemma3/embed.go b/model/models/gemma3/embed.goindex 525547767..9251111cf 100644--- a/model/models/gemma3/embed.go+++ b/model/models/gemma3/embed.go@@ -2,7 +2,6 @@ package gemma3import ("github.com/ollama/ollama/fs"- "github.com/ollama/ollama/kvcache""github.com/ollama/ollama/ml""github.com/ollama/ollama/ml/nn""github.com/ollama/ollama/ml/nn/pooling"@@ -53,10 +52,5 @@ func newEmbedModel(c fs.Config) (model.Model, error) {poolingType: pooling.Type(c.Uint("pooling_type", 0)),}- m.Cache = kvcache.NewWrapperCache(- kvcache.NewSWACache(int32(c.Uint("attention.sliding_window")), m.Shift),- kvcache.NewCausalCache(m.Shift),- )-return m, nil}diff --git a/model/models/gemma3/model_text.go b/model/models/gemma3/model_text.goindex 631baeccd..d5bdd410d 100644--- a/model/models/gemma3/model_text.go+++ b/model/models/gemma3/model_text.go@@ -182,16 +182,18 @@ func (m *TextModel) Forward(ctx ml.Context, batch input.Batch, cache kvcache.Cacfor i, layer := range m.Layers {// gemma alternates between the sliding window (local) and causal (global)// kv cache every 6 layers- cacheType := cacheTypeSWA- if (i+1)%gemmaGlobalCacheCount == 0 {- cacheType = cacheTypeCausal- }- cache.SetLayer(i)- wc := cache.(*kvcache.WrapperCache)- wc.SetLayerType(cacheType)-- if causal, ok := wc.UnderlyingCache().(*kvcache.Causal); ok {- causal.SetCausal(ctx, kvcache.CausalOptions{Except: except})+ if cache != nil {+ cacheType := cacheTypeSWA+ if (i+1)%gemmaGlobalCacheCount == 0 {+ cacheType = cacheTypeCausal+ }+ cache.SetLayer(i)+ wc := cache.(*kvcache.WrapperCache)+ wc.SetLayerType(cacheType)++ if causal, ok := wc.UnderlyingCache().(*kvcache.Causal); ok {+ causal.SetCausal(ctx, kvcache.CausalOptions{Except: except})+ }}var lastLayerOutputs ml.Tensor