feat(llama.cpp): Add support to grammar triggers (#4733)

Signed-off-by: Ettore Di Giacinto <[email protected]>

Author: mudler

backend/backend.proto

@@ -163,6 +163,11 @@ message Reply {
   double timing_token_generation = 5;
 }
 
+message GrammarTrigger {
+  string word = 1;
+  bool at_start = 2; 
+}
+
 message ModelOptions {
   string Model = 1;
   int32 ContextSize = 2;
@@ -247,6 +252,8 @@ message ModelOptions {
 
   string CacheTypeKey = 63;
   string CacheTypeValue = 64;
+
+  repeated GrammarTrigger GrammarTriggers = 65;
 }
 
 message Result {

backend/cpp/llama/grpc-server.cpp

@@ -468,6 +468,9 @@ struct llama_server_context
     bool add_bos_token      = true;
     bool has_eos_token      = true;
 
+    bool grammar_lazy = false;
+    std::vector<common_grammar_trigger> grammar_trigger_words;
+
     int32_t n_ctx;  // total context for all clients / slots
 
     // system prompt
@@ -706,6 +709,8 @@ struct llama_server_context
         slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
         slot->sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
         slot->sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
+        slot->sparams.grammar_trigger_words = grammar_trigger_words;
+        slot->sparams.grammar_lazy = grammar_lazy;
 
         if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
             // Might be better to reject the request with a 400 ?
@@ -2374,6 +2379,21 @@ static void params_parse(const backend::ModelOptions* request,
     if ( request->ropefreqscale() != 0.0f ) {
         params.rope_freq_scale = request->ropefreqscale();
     }
+
+    if (request->grammartriggers_size() > 0) {
+        LOG_INFO("configuring grammar triggers", {});
+        llama.grammar_lazy = true;
+        for (int i = 0; i < request->grammartriggers_size(); i++) {
+            common_grammar_trigger trigger;
+            trigger.word = request->grammartriggers(i).word();
+            trigger.at_start = request->grammartriggers(i).at_start();
+            llama.grammar_trigger_words.push_back(trigger);
+            LOG_INFO("grammar trigger", {
+                { "word", trigger.word },
+                { "at_start", trigger.at_start }
+            });
+        }
+    }
 }
 
 

core/backend/options.go

@@ -118,9 +118,19 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		nGPULayers = *c.NGPULayers
 	}
 
+	triggers := make([]*pb.GrammarTrigger, 0)
+	for _, t := range c.FunctionsConfig.GrammarConfig.GrammarTriggers {
+		triggers = append(triggers, &pb.GrammarTrigger{
+			Word:    t.Word,
+			AtStart: t.AtStart,
+		})
+
+	}
+
 	return &pb.ModelOptions{
 		CUDA:                 c.CUDA || c.Diffusers.CUDA,
 		SchedulerType:        c.Diffusers.SchedulerType,
+		GrammarTriggers:      triggers,
 		PipelineType:         c.Diffusers.PipelineType,
 		CFGScale:             c.CFGScale,
 		LoraAdapter:          c.LoraAdapter,

pkg/functions/parse.go

@@ -47,6 +47,14 @@ type GrammarConfig struct {
 	// SchemaType can be configured to use a specific schema type to force the grammar
 	// available : json, llama3.1
 	SchemaType string `yaml:"schema_type"`
+
+	GrammarTriggers []GrammarTrigger `yaml:"triggers"`
+}
+
+type GrammarTrigger struct {
+	// Trigger is the string that triggers the grammar
+	Word    string `yaml:"word"`
+	AtStart bool   `yaml:"at_start"`
 }
 
 // FunctionsConfig is the configuration for the tool/function call.
@@ -361,6 +369,6 @@ func ParseFunctionCallArgs(functionArguments string, functionConfig FunctionsCon
 	}
 
 	jsonBytes, _ := json.Marshal(args)
-	
+
 	return string(jsonBytes)
 }