Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

models: enable deepseek2 (deepseek v3.1 w/ MLA) on the new engine (#13151)

Patrick Devine604e43b28d12
diff --git a/fs/ggml/ggml.go b/fs/ggml/ggml.go
index 205279c67..dcd1234dc 100644
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -250,6 +250,7 @@ func (kv KV) OllamaEngineRequired() bool {
"qwen3", "qwen3moe",
"qwen3vl", "qwen3vlmoe",
"deepseekocr",
+ "deepseek2",
}, kv.Architecture())
}
diff --git a/model/models/deepseek2/model.go b/model/models/deepseek2/model.go
index c1251ecfa..99c57959d 100644
--- a/model/models/deepseek2/model.go
+++ b/model/models/deepseek2/model.go
@@ -236,6 +236,11 @@ type Model struct {
}
func New(c fs.Config) (model.Model, error) {
+ if c.Uint("attention.key_length_mla") == 0 {
+ // non-MLA models aren't yet supported
+ return nil, model.ErrUnsupportedModel
+ }
+
layers := make([]Layer, c.Uint("block_count"))
firstDenseLayerIndex := int(c.Uint("leading_dense_block_count"))