feat: Add UseFastTokenizer

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

api/config/config.go

@@ -56,9 +56,10 @@ type Config struct {
 	NGQA       int32   `yaml:"ngqa"`
 
 	// AutoGPTQ
-	ModelBaseName string `yaml:"model_base_name"`
-	Device        string `yaml:"device"`
-	Triton        bool   `yaml:"triton"`
+	ModelBaseName    string `yaml:"model_base_name"`
+	Device           string `yaml:"device"`
+	Triton           bool   `yaml:"triton"`
+	UseFastTokenizer bool   `yaml:"use_fast_tokenizer"`
 }
 
 type Functions struct {

api/config/prediction.go

@@ -39,4 +39,6 @@ type PredictionOptions struct {
 	RopeFreqBase        float32 `json:"rope_freq_base" yaml:"rope_freq_base"`
 	RopeFreqScale       float32 `json:"rope_freq_scale" yaml:"rope_freq_scale"`
 	NegativePromptScale float32 `json:"negative_prompt_scale" yaml:"negative_prompt_scale"`
+	// AutoGPTQ
+	UseFastTokenizer bool `json:"use_fast_tokenizer" yaml:"use_fast_tokenizer"`
 }