feat(llama.cpp): add flash_attention and no_kv_offloading (#2310)

feat(llama.cpp): add flash_attn and no_kv_offload

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

Author: mudler

backend/backend.proto

@@ -212,6 +212,9 @@ message ModelOptions {
   float YarnBetaSlow = 47;
 
   string Type = 49;
+
+  bool FlashAttention = 56;
+  bool NoKVOffload = 57;
 }
 
 message Result {

backend/cpp/llama/grpc-server.cpp

@@ -2254,6 +2254,9 @@ static void params_parse(const backend::ModelOptions* request,
     }
     params.use_mlock = request->mlock();
     params.use_mmap = request->mmap();
+    params.flash_attn = request->flashattention();
+    params.no_kv_offload = request->nokvoffload();
+
     params.embedding = request->embeddings();
 
     if (request->ropescaling() == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }

core/backend/options.go

@@ -77,6 +77,8 @@ func gRPCModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		MaxModelLen:          int32(c.MaxModelLen),
 		TensorParallelSize:   int32(c.TensorParallelSize),
 		MMProj:               c.MMProj,
+		FlashAttention:       c.FlashAttention,
+		NoKVOffload:          c.NoKVOffloading,
 		YarnExtFactor:        c.YarnExtFactor,
 		YarnAttnFactor:       c.YarnAttnFactor,
 		YarnBetaFast:         c.YarnBetaFast,

core/config/backend_config.go

@@ -132,6 +132,9 @@ type LLMConfig struct {
 	TensorParallelSize   int     `yaml:"tensor_parallel_size"`   // vLLM
 	MMProj               string  `yaml:"mmproj"`
 
+	FlashAttention bool `yaml:"flash_attention"`
+	NoKVOffloading bool `yaml:"no_kv_offloading"`
+
 	RopeScaling string `yaml:"rope_scaling"`
 	ModelType   string `yaml:"type"`