manual merge

Signed-off-by: Dave Lee <[email protected]>

Author: dave-gray101

.github/workflows/notify-models.yaml

@@ -18,7 +18,7 @@ jobs:
       with:
         model: 'hermes-2-theta-llama-3-8b' # Any from models.localai.io, or from huggingface.com with: "huggingface://<repository>/file"
         # Check the PR diff using the current branch and the base branch of the PR
-    - uses: GrantBirki/git-diff-action@v2.7.0
+    - uses: GrantBirki/git-diff-action@v2.8.0
       id: git-diff-action
       with:
             json_diff_file_output: diff.json
@@ -99,7 +99,7 @@ jobs:
         docker run -e -ti -d --name local-ai -p 8080:8080 localai/localai:master-ffmpeg-core run --debug $MODEL_NAME
         until [ "`docker inspect -f {{.State.Health.Status}} local-ai`" == "healthy" ]; do echo "Waiting for container to be ready";  docker logs --tail 10 local-ai; sleep 2; done
       # Check the PR diff using the current branch and the base branch of the PR
-    - uses: GrantBirki/git-diff-action@v2.7.0
+    - uses: GrantBirki/git-diff-action@v2.8.0
       id: git-diff-action
       with:
             json_diff_file_output: diff.json

Makefile

@@ -8,7 +8,7 @@ DETECT_LIBS?=true
 # llama.cpp versions
 GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
 GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
-CPPLLAMA_VERSION?=53debe6f3c9cca87e9520a83ee8c14d88977afa4
+CPPLLAMA_VERSION?=3ec9fd4b77b6aca03a3c2bf678eae3f9517d6904
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp

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 }
+            });
+        }
+    }
 }
 
 
@@ -2522,6 +2542,18 @@ class BackendServiceImpl final : public backend::Backend::Service {
         return grpc::Status::OK;
     }
 
+    grpc::Status TokenizeString(ServerContext* context, const backend::PredictOptions* request, backend::TokenizationResponse* response){
+         json data = parse_options(false, request, llama);
+
+         std::vector<llama_token> tokens = llama.tokenize(data["prompt"],false);
+
+         for (int i=0 ; i< tokens.size(); i++){
+            response->add_tokens(tokens[i]);
+         }
+
+        return grpc::Status::OK;
+    }
+
     grpc::Status GetMetrics(ServerContext* context, const backend::MetricsRequest* request, backend::MetricsResponse* response) {
         llama_client_slot* active_slot = llama.get_active_slot();
 

backend/python/transformers/requirements-cpu.txt

@@ -5,4 +5,4 @@ accelerate
 transformers
 bitsandbytes
 outetts
-sentence-transformers==3.4.0
+sentence-transformers==3.4.1

backend/python/transformers/requirements-cublas11.txt

@@ -6,4 +6,4 @@ accelerate
 transformers
 bitsandbytes
 outetts
-sentence-transformers==3.4.0
+sentence-transformers==3.4.1

backend/python/transformers/requirements-cublas12.txt

@@ -5,4 +5,4 @@ numba==0.60.0
 transformers
 bitsandbytes
 outetts
-sentence-transformers==3.4.0
+sentence-transformers==3.4.1

backend/python/transformers/requirements-hipblas.txt

@@ -7,4 +7,4 @@ numba==0.60.0
 bitsandbytes
 outetts
 bitsandbytes
-sentence-transformers==3.4.0
+sentence-transformers==3.4.1

backend/python/transformers/requirements-intel.txt

@@ -8,4 +8,4 @@ numba==0.60.0
 intel-extension-for-transformers
 bitsandbytes
 outetts
-sentence-transformers==3.4.0
+sentence-transformers==3.4.1

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,