mirror of
https://github.com/ollama/ollama.git
synced 2025-12-05 18:46:22 -06:00
models: enable deepseek2 (deepseek v3.1 w/ MLA) on the new engine (#13151)
This commit is contained in:
@@ -250,6 +250,7 @@ func (kv KV) OllamaEngineRequired() bool {
|
||||
"qwen3", "qwen3moe",
|
||||
"qwen3vl", "qwen3vlmoe",
|
||||
"deepseekocr",
|
||||
"deepseek2",
|
||||
}, kv.Architecture())
|
||||
}
|
||||
|
||||
|
||||
@@ -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"))
|
||||
|
||||
Reference in New Issue
Block a user