dearwolf/LocalAI
1
0
Fork 0
mirror of https://github.com/mudler/LocalAI.git synced 2025-05-28 22:44:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 12

feat: move llama to a grpc

Browse source 
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
This commit is contained in:
Ettore Di Giacinto 2023-07-15 01:19:43 +02:00
parent b816009db0
commit 58f6aab637
13 changed files with 454 additions and 340 deletions
Download patch file Download diff file Expand all files Collapse all files

205
pkg/grpc/proto/llmserver.pb.go
View file

@ -87,7 +87,6 @@ type PredictOptions struct {
MirostatTAU float32 `protobuf:"fixed32,21,opt,name=MirostatTAU,proto3" json:"MirostatTAU,omitempty"`
PenalizeNL bool `protobuf:"varint,22,opt,name=PenalizeNL,proto3" json:"PenalizeNL,omitempty"`
LogitBias string `protobuf:"bytes,23,opt,name=LogitBias,proto3" json:"LogitBias,omitempty"`
PathPromptCache string `protobuf:"bytes,24,opt,name=PathPromptCache,proto3" json:"PathPromptCache,omitempty"`
MLock bool `protobuf:"varint,25,opt,name=MLock,proto3" json:"MLock,omitempty"`
MMap bool `protobuf:"varint,26,opt,name=MMap,proto3" json:"MMap,omitempty"`
PromptCacheAll bool `protobuf:"varint,27,opt,name=PromptCacheAll,proto3" json:"PromptCacheAll,omitempty"`
@ -98,6 +97,8 @@ type PredictOptions struct {
TopP float32 `protobuf:"fixed32,32,opt,name=TopP,proto3" json:"TopP,omitempty"`
PromptCachePath string `protobuf:"bytes,33,opt,name=PromptCachePath,proto3" json:"PromptCachePath,omitempty"`
Debug bool `protobuf:"varint,34,opt,name=Debug,proto3" json:"Debug,omitempty"`
EmbeddingTokens []int32 `protobuf:"varint,35,rep,packed,name=EmbeddingTokens,proto3" json:"EmbeddingTokens,omitempty"`
Embeddings string `protobuf:"bytes,36,opt,name=Embeddings,proto3" json:"Embeddings,omitempty"`
}
func (x *PredictOptions) Reset() {
@ -293,13 +294,6 @@ func (x *PredictOptions) GetLogitBias() string {
return ""
}
func (x *PredictOptions) GetPathPromptCache() string {
if x != nil {
return x.PathPromptCache
}
return ""
}
func (x *PredictOptions) GetMLock() bool {
if x != nil {
return x.MLock
@ -370,6 +364,20 @@ func (x *PredictOptions) GetDebug() bool {
return false
}
func (x *PredictOptions) GetEmbeddingTokens() []int32 {
if x != nil {
return x.EmbeddingTokens
}
return nil
}
func (x *PredictOptions) GetEmbeddings() string {
if x != nil {
return x.Embeddings
}
return ""
}
// The response message containing the result
type Reply struct {
state protoimpl.MessageState
@ -624,13 +632,60 @@ func (x *Result) GetSuccess() bool {
return false
}
type EmbeddingResult struct {
state protoimpl.MessageState
sizeCache protoimpl.SizeCache
unknownFields protoimpl.UnknownFields
Embeddings []float32 `protobuf:"fixed32,1,rep,packed,name=embeddings,proto3" json:"embeddings,omitempty"`
}
func (x *EmbeddingResult) Reset() {
*x = EmbeddingResult{}
if protoimpl.UnsafeEnabled {
mi := &file_pkg_grpc_proto_llmserver_proto_msgTypes[5]
ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
ms.StoreMessageInfo(mi)
}
}
func (x *EmbeddingResult) String() string {
return protoimpl.X.MessageStringOf(x)
}
func (*EmbeddingResult) ProtoMessage() {}
func (x *EmbeddingResult) ProtoReflect() protoreflect.Message {
mi := &file_pkg_grpc_proto_llmserver_proto_msgTypes[5]
if protoimpl.UnsafeEnabled && x != nil {
ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
if ms.LoadMessageInfo() == nil {
ms.StoreMessageInfo(mi)
}
return ms
}
return mi.MessageOf(x)
}
// Deprecated: Use EmbeddingResult.ProtoReflect.Descriptor instead.
func (*EmbeddingResult) Descriptor() ([]byte, []int) {
return file_pkg_grpc_proto_llmserver_proto_rawDescGZIP(), []int{5}
}
func (x *EmbeddingResult) GetEmbeddings() []float32 {
if x != nil {
return x.Embeddings
}
return nil
}
var File_pkg_grpc_proto_llmserver_proto protoreflect.FileDescriptor
var file_pkg_grpc_proto_llmserver_proto_rawDesc = []byte{
0x0a, 0x1e, 0x70, 0x6b, 0x67, 0x2f, 0x67, 0x72, 0x70, 0x63, 0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f,
0x2f, 0x6c, 0x6c, 0x6d, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x70, 0x72, 0x6f, 0x74, 0x6f,
0x12, 0x03, 0x6c, 0x6c, 0x6d, 0x22, 0x0f, 0x0a, 0x0d, 0x48, 0x65, 0x61, 0x6c, 0x74, 0x68, 0x4d,
0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x80, 0x08, 0x0a, 0x0e, 0x50, 0x72, 0x65, 0x64, 0x69,
0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0xa0, 0x08, 0x0a, 0x0e, 0x50, 0x72, 0x65, 0x64, 0x69,
0x63, 0x74, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x12, 0x16, 0x0a, 0x06, 0x50, 0x72, 0x6f,
0x6d, 0x70, 0x74, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x52, 0x06, 0x50, 0x72, 0x6f, 0x6d, 0x70,
0x74, 0x12, 0x12, 0x0a, 0x04, 0x53, 0x65, 0x65, 0x64, 0x18, 0x02, 0x20, 0x01, 0x28, 0x05, 0x52,
@ -673,28 +728,30 @@ var file_pkg_grpc_proto_llmserver_proto_rawDesc = []byte{
0x1e, 0x0a, 0x0a, 0x50, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x4e, 0x4c, 0x18, 0x16, 0x20,
0x01, 0x28, 0x08, 0x52, 0x0a, 0x50, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x4e, 0x4c, 0x12,
0x1c, 0x0a, 0x09, 0x4c, 0x6f, 0x67, 0x69, 0x74, 0x42, 0x69, 0x61, 0x73, 0x18, 0x17, 0x20, 0x01,
0x28, 0x09, 0x52, 0x09, 0x4c, 0x6f, 0x67, 0x69, 0x74, 0x42, 0x69, 0x61, 0x73, 0x12, 0x28, 0x0a,
0x0f, 0x50, 0x61, 0x74, 0x68, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65,
0x18, 0x18, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0f, 0x50, 0x61, 0x74, 0x68, 0x50, 0x72, 0x6f, 0x6d,
0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x12, 0x14, 0x0a, 0x05, 0x4d, 0x4c, 0x6f, 0x63, 0x6b,
0x18, 0x19, 0x20, 0x01, 0x28, 0x08, 0x52, 0x05, 0x4d, 0x4c, 0x6f, 0x63, 0x6b, 0x12, 0x12, 0x0a,
0x04, 0x4d, 0x4d, 0x61, 0x70, 0x18, 0x1a, 0x20, 0x01, 0x28, 0x08, 0x52, 0x04, 0x4d, 0x4d, 0x61,
0x70, 0x12, 0x26, 0x0a, 0x0e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65,
0x41, 0x6c, 0x6c, 0x18, 0x1b, 0x20, 0x01, 0x28, 0x08, 0x52, 0x0e, 0x50, 0x72, 0x6f, 0x6d, 0x70,
0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x41, 0x6c, 0x6c, 0x12, 0x24, 0x0a, 0x0d, 0x50, 0x72, 0x6f,
0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x52, 0x4f, 0x18, 0x1c, 0x20, 0x01, 0x28, 0x08,
0x52, 0x0d, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x52, 0x4f, 0x12,
0x18, 0x0a, 0x07, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x18, 0x1d, 0x20, 0x01, 0x28, 0x09,
0x52, 0x07, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x12, 0x18, 0x0a, 0x07, 0x4d, 0x61, 0x69,
0x6e, 0x47, 0x50, 0x55, 0x18, 0x1e, 0x20, 0x01, 0x28, 0x09, 0x52, 0x07, 0x4d, 0x61, 0x69, 0x6e,
0x47, 0x50, 0x55, 0x12, 0x20, 0x0a, 0x0b, 0x54, 0x65, 0x6e, 0x73, 0x6f, 0x72, 0x53, 0x70, 0x6c,
0x69, 0x74, 0x18, 0x1f, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0b, 0x54, 0x65, 0x6e, 0x73, 0x6f, 0x72,
0x53, 0x70, 0x6c, 0x69, 0x74, 0x12, 0x12, 0x0a, 0x04, 0x54, 0x6f, 0x70, 0x50, 0x18, 0x20, 0x20,
0x01, 0x28, 0x02, 0x52, 0x04, 0x54, 0x6f, 0x70, 0x50, 0x12, 0x28, 0x0a, 0x0f, 0x50, 0x72, 0x6f,
0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x50, 0x61, 0x74, 0x68, 0x18, 0x21, 0x20, 0x01,
0x28, 0x09, 0x52, 0x0f, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x50,
0x61, 0x74, 0x68, 0x12, 0x14, 0x0a, 0x05, 0x44, 0x65, 0x62, 0x75, 0x67, 0x18, 0x22, 0x20, 0x01,
0x28, 0x08, 0x52, 0x05, 0x44, 0x65, 0x62, 0x75, 0x67, 0x22, 0x21, 0x0a, 0x05, 0x52, 0x65, 0x70,
0x28, 0x09, 0x52, 0x09, 0x4c, 0x6f, 0x67, 0x69, 0x74, 0x42, 0x69, 0x61, 0x73, 0x12, 0x14, 0x0a,
0x05, 0x4d, 0x4c, 0x6f, 0x63, 0x6b, 0x18, 0x19, 0x20, 0x01, 0x28, 0x08, 0x52, 0x05, 0x4d, 0x4c,
0x6f, 0x63, 0x6b, 0x12, 0x12, 0x0a, 0x04, 0x4d, 0x4d, 0x61, 0x70, 0x18, 0x1a, 0x20, 0x01, 0x28,
0x08, 0x52, 0x04, 0x4d, 0x4d, 0x61, 0x70, 0x12, 0x26, 0x0a, 0x0e, 0x50, 0x72, 0x6f, 0x6d, 0x70,
0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x41, 0x6c, 0x6c, 0x18, 0x1b, 0x20, 0x01, 0x28, 0x08, 0x52,
0x0e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x41, 0x6c, 0x6c, 0x12,
0x24, 0x0a, 0x0d, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x52, 0x4f,
0x18, 0x1c, 0x20, 0x01, 0x28, 0x08, 0x52, 0x0d, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x61,
0x63, 0x68, 0x65, 0x52, 0x4f, 0x12, 0x18, 0x0a, 0x07, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72,
0x18, 0x1d, 0x20, 0x01, 0x28, 0x09, 0x52, 0x07, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x12,
0x18, 0x0a, 0x07, 0x4d, 0x61, 0x69, 0x6e, 0x47, 0x50, 0x55, 0x18, 0x1e, 0x20, 0x01, 0x28, 0x09,
0x52, 0x07, 0x4d, 0x61, 0x69, 0x6e, 0x47, 0x50, 0x55, 0x12, 0x20, 0x0a, 0x0b, 0x54, 0x65, 0x6e,
0x73, 0x6f, 0x72, 0x53, 0x70, 0x6c, 0x69, 0x74, 0x18, 0x1f, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0b,
0x54, 0x65, 0x6e, 0x73, 0x6f, 0x72, 0x53, 0x70, 0x6c, 0x69, 0x74, 0x12, 0x12, 0x0a, 0x04, 0x54,
0x6f, 0x70, 0x50, 0x18, 0x20, 0x20, 0x01, 0x28, 0x02, 0x52, 0x04, 0x54, 0x6f, 0x70, 0x50, 0x12,
0x28, 0x0a, 0x0f, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x61, 0x63, 0x68, 0x65, 0x50, 0x61,
0x74, 0x68, 0x18, 0x21, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0f, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74,
0x43, 0x61, 0x63, 0x68, 0x65, 0x50, 0x61, 0x74, 0x68, 0x12, 0x14, 0x0a, 0x05, 0x44, 0x65, 0x62,
0x75, 0x67, 0x18, 0x22, 0x20, 0x01, 0x28, 0x08, 0x52, 0x05, 0x44, 0x65, 0x62, 0x75, 0x67, 0x12,
0x28, 0x0a, 0x0f, 0x45, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x54, 0x6f, 0x6b, 0x65,
0x6e, 0x73, 0x18, 0x23, 0x20, 0x03, 0x28, 0x05, 0x52, 0x0f, 0x45, 0x6d, 0x62, 0x65, 0x64, 0x64,
0x69, 0x6e, 0x67, 0x54, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x12, 0x1e, 0x0a, 0x0a, 0x45, 0x6d, 0x62,
0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x73, 0x18, 0x24, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0a, 0x45,
0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x21, 0x0a, 0x05, 0x52, 0x65, 0x70,
0x6c, 0x79, 0x12, 0x18, 0x0a, 0x07, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x18, 0x01, 0x20,
0x01, 0x28, 0x09, 0x52, 0x07, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x82, 0x03, 0x0a,
0x0c, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x12, 0x14, 0x0a,
@ -724,26 +781,33 @@ var file_pkg_grpc_proto_llmserver_proto_rawDesc = []byte{
0x74, 0x22, 0x3c, 0x0a, 0x06, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x12, 0x18, 0x0a, 0x07, 0x6d,
0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x52, 0x07, 0x6d, 0x65,
0x73, 0x73, 0x61, 0x67, 0x65, 0x12, 0x18, 0x0a, 0x07, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73,
0x18, 0x02, 0x20, 0x01, 0x28, 0x08, 0x52, 0x07, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x32,
0xc4, 0x01, 0x0a, 0x03, 0x4c, 0x4c, 0x4d, 0x12, 0x2a, 0x0a, 0x06, 0x48, 0x65, 0x61, 0x6c, 0x74,
0x68, 0x12, 0x12, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x48, 0x65, 0x61, 0x6c, 0x74, 0x68, 0x4d, 0x65,
0x73, 0x73, 0x61, 0x67, 0x65, 0x1a, 0x0a, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x52, 0x65, 0x70, 0x6c,
0x79, 0x22, 0x00, 0x12, 0x2c, 0x0a, 0x07, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x12, 0x13,
0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x4f, 0x70, 0x74, 0x69,
0x6f, 0x6e, 0x73, 0x1a, 0x0a, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x52, 0x65, 0x70, 0x6c, 0x79, 0x22,
0x00, 0x12, 0x2d, 0x0a, 0x09, 0x4c, 0x6f, 0x61, 0x64, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x12, 0x11,
0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e,
0x73, 0x1a, 0x0b, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x22, 0x00,
0x12, 0x34, 0x0a, 0x0d, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x53, 0x74, 0x72, 0x65, 0x61,
0x6d, 0x12, 0x13, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x4f,
0x18, 0x02, 0x20, 0x01, 0x28, 0x08, 0x52, 0x07, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x22,
0x31, 0x0a, 0x0f, 0x45, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x52, 0x65, 0x73, 0x75,
0x6c, 0x74, 0x12, 0x1e, 0x0a, 0x0a, 0x65, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x73,
0x18, 0x01, 0x20, 0x03, 0x28, 0x02, 0x52, 0x0a, 0x65, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e,
0x67, 0x73, 0x32, 0xfe, 0x01, 0x0a, 0x03, 0x4c, 0x4c, 0x4d, 0x12, 0x2a, 0x0a, 0x06, 0x48, 0x65,
0x61, 0x6c, 0x74, 0x68, 0x12, 0x12, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x48, 0x65, 0x61, 0x6c, 0x74,
0x68, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x1a, 0x0a, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x52,
0x65, 0x70, 0x6c, 0x79, 0x22, 0x00, 0x12, 0x2c, 0x0a, 0x07, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63,
0x74, 0x12, 0x13, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x4f,
0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x0a, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x52, 0x65, 0x70,
0x6c, 0x79, 0x22, 0x00, 0x30, 0x01, 0x42, 0x57, 0x0a, 0x1b, 0x69, 0x6f, 0x2e, 0x73, 0x6b, 0x79,
0x6e, 0x65, 0x74, 0x2e, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x61, 0x69, 0x2e, 0x6c, 0x6c, 0x6d, 0x73,
0x65, 0x72, 0x76, 0x65, 0x72, 0x42, 0x09, 0x4c, 0x4c, 0x4d, 0x53, 0x65, 0x72, 0x76, 0x65, 0x72,
0x50, 0x01, 0x5a, 0x2b, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67,
0x6f, 0x2d, 0x73, 0x6b, 0x79, 0x6e, 0x65, 0x74, 0x2f, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x41, 0x49,
0x2f, 0x70, 0x6b, 0x67, 0x2f, 0x67, 0x72, 0x70, 0x63, 0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x62,
0x06, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x33,
0x6c, 0x79, 0x22, 0x00, 0x12, 0x2d, 0x0a, 0x09, 0x4c, 0x6f, 0x61, 0x64, 0x4d, 0x6f, 0x64, 0x65,
0x6c, 0x12, 0x11, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x4f, 0x70, 0x74,
0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x0b, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x52, 0x65, 0x73, 0x75, 0x6c,
0x74, 0x22, 0x00, 0x12, 0x34, 0x0a, 0x0d, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x53, 0x74,
0x72, 0x65, 0x61, 0x6d, 0x12, 0x13, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69,
0x63, 0x74, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x0a, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e,
0x52, 0x65, 0x70, 0x6c, 0x79, 0x22, 0x00, 0x30, 0x01, 0x12, 0x38, 0x0a, 0x09, 0x45, 0x6d, 0x62,
0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x12, 0x13, 0x2e, 0x6c, 0x6c, 0x6d, 0x2e, 0x50, 0x72, 0x65,
0x64, 0x69, 0x63, 0x74, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x14, 0x2e, 0x6c, 0x6c,
0x6d, 0x2e, 0x45, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x52, 0x65, 0x73, 0x75, 0x6c,
0x74, 0x22, 0x00, 0x42, 0x57, 0x0a, 0x1b, 0x69, 0x6f, 0x2e, 0x73, 0x6b, 0x79, 0x6e, 0x65, 0x74,
0x2e, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x61, 0x69, 0x2e, 0x6c, 0x6c, 0x6d, 0x73, 0x65, 0x72, 0x76,
0x65, 0x72, 0x42, 0x09, 0x4c, 0x4c, 0x4d, 0x53, 0x65, 0x72, 0x76, 0x65, 0x72, 0x50, 0x01, 0x5a,
0x2b, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, 0x6f, 0x2d, 0x73,
0x6b, 0x79, 0x6e, 0x65, 0x74, 0x2f, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x41, 0x49, 0x2f, 0x70, 0x6b,
0x67, 0x2f, 0x67, 0x72, 0x70, 0x63, 0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x62, 0x06, 0x70, 0x72,
0x6f, 0x74, 0x6f, 0x33,
}
var (
@ -758,25 +822,28 @@ func file_pkg_grpc_proto_llmserver_proto_rawDescGZIP() []byte {
return file_pkg_grpc_proto_llmserver_proto_rawDescData
}
var file_pkg_grpc_proto_llmserver_proto_msgTypes = make([]protoimpl.MessageInfo, 5)
var file_pkg_grpc_proto_llmserver_proto_msgTypes = make([]protoimpl.MessageInfo, 6)
var file_pkg_grpc_proto_llmserver_proto_goTypes = []interface{}{
(*HealthMessage)(nil), // 0: llm.HealthMessage
(*PredictOptions)(nil), // 1: llm.PredictOptions
(*Reply)(nil), // 2: llm.Reply
(*ModelOptions)(nil), // 3: llm.ModelOptions
(*Result)(nil), // 4: llm.Result
(*HealthMessage)(nil), // 0: llm.HealthMessage
(*PredictOptions)(nil), // 1: llm.PredictOptions
(*Reply)(nil), // 2: llm.Reply
(*ModelOptions)(nil), // 3: llm.ModelOptions
(*Result)(nil), // 4: llm.Result
(*EmbeddingResult)(nil), // 5: llm.EmbeddingResult
}
var file_pkg_grpc_proto_llmserver_proto_depIdxs = []int32{
0, // 0: llm.LLM.Health:input_type -> llm.HealthMessage
1, // 1: llm.LLM.Predict:input_type -> llm.PredictOptions
3, // 2: llm.LLM.LoadModel:input_type -> llm.ModelOptions
1, // 3: llm.LLM.PredictStream:input_type -> llm.PredictOptions
2, // 4: llm.LLM.Health:output_type -> llm.Reply
2, // 5: llm.LLM.Predict:output_type -> llm.Reply
4, // 6: llm.LLM.LoadModel:output_type -> llm.Result
2, // 7: llm.LLM.PredictStream:output_type -> llm.Reply
4, // [4:8] is the sub-list for method output_type
0, // [0:4] is the sub-list for method input_type
1, // 4: llm.LLM.Embedding:input_type -> llm.PredictOptions
2, // 5: llm.LLM.Health:output_type -> llm.Reply
2, // 6: llm.LLM.Predict:output_type -> llm.Reply
4, // 7: llm.LLM.LoadModel:output_type -> llm.Result
2, // 8: llm.LLM.PredictStream:output_type -> llm.Reply
5, // 9: llm.LLM.Embedding:output_type -> llm.EmbeddingResult
5, // [5:10] is the sub-list for method output_type
0, // [0:5] is the sub-list for method input_type
0, // [0:0] is the sub-list for extension type_name
0, // [0:0] is the sub-list for extension extendee
0, // [0:0] is the sub-list for field type_name
@ -848,6 +915,18 @@ func file_pkg_grpc_proto_llmserver_proto_init() {
return nil
}
}
file_pkg_grpc_proto_llmserver_proto_msgTypes[5].Exporter = func(v interface{}, i int) interface{} {
switch v := v.(*EmbeddingResult); i {
case 0:
return &v.state
case 1:
return &v.sizeCache
case 2:
return &v.unknownFields
default:
return nil
}
}
}
type x struct{}
out := protoimpl.TypeBuilder{
@ -855,7 +934,7 @@ func file_pkg_grpc_proto_llmserver_proto_init() {
GoPackagePath: reflect.TypeOf(x{}).PkgPath(),
RawDescriptor: file_pkg_grpc_proto_llmserver_proto_rawDesc,
NumEnums: 0,
NumMessages: 5,
NumMessages: 6,
NumExtensions: 0,
NumServices: 1,
},

8
pkg/grpc/proto/llmserver.proto
View file

@ -12,6 +12,7 @@ service LLM {
rpc Predict(PredictOptions) returns (Reply) {}
rpc LoadModel(ModelOptions) returns (Result) {}
rpc PredictStream(PredictOptions) returns (stream Reply) {}
rpc Embedding(PredictOptions) returns (EmbeddingResult) {}
}
message HealthMessage {}
@ -41,7 +42,6 @@ message PredictOptions {
float MirostatTAU = 21;
bool PenalizeNL = 22;
string LogitBias = 23;
string PathPromptCache = 24;
bool MLock = 25;
bool MMap = 26;
bool PromptCacheAll = 27;
@ -52,6 +52,8 @@ message PredictOptions {
float TopP = 32;
string PromptCachePath = 33;
bool Debug = 34;
repeated int32 EmbeddingTokens = 35;
string Embeddings = 36;
}
// The response message containing the result
@ -79,4 +81,8 @@ message ModelOptions {
message Result {
string message = 1;
bool success = 2;
}
message EmbeddingResult {
repeated float embeddings = 1;
}

36
pkg/grpc/proto/llmserver_grpc.pb.go
View file

@ -26,6 +26,7 @@ type LLMClient interface {
Predict(ctx context.Context, in *PredictOptions, opts ...grpc.CallOption) (*Reply, error)
LoadModel(ctx context.Context, in *ModelOptions, opts ...grpc.CallOption) (*Result, error)
PredictStream(ctx context.Context, in *PredictOptions, opts ...grpc.CallOption) (LLM_PredictStreamClient, error)
Embedding(ctx context.Context, in *PredictOptions, opts ...grpc.CallOption) (*EmbeddingResult, error)
}
type lLMClient struct {
@ -95,6 +96,15 @@ func (x *lLMPredictStreamClient) Recv() (*Reply, error) {
return m, nil
}
func (c *lLMClient) Embedding(ctx context.Context, in *PredictOptions, opts ...grpc.CallOption) (*EmbeddingResult, error) {
out := new(EmbeddingResult)
err := c.cc.Invoke(ctx, "/llm.LLM/Embedding", in, out, opts...)
if err != nil {
return nil, err
}
return out, nil
}
// LLMServer is the server API for LLM service.
// All implementations must embed UnimplementedLLMServer
// for forward compatibility
@ -103,6 +113,7 @@ type LLMServer interface {
Predict(context.Context, *PredictOptions) (*Reply, error)
LoadModel(context.Context, *ModelOptions) (*Result, error)
PredictStream(*PredictOptions, LLM_PredictStreamServer) error
Embedding(context.Context, *PredictOptions) (*EmbeddingResult, error)
mustEmbedUnimplementedLLMServer()
}
@ -122,6 +133,9 @@ func (UnimplementedLLMServer) LoadModel(context.Context, *ModelOptions) (*Result
func (UnimplementedLLMServer) PredictStream(*PredictOptions, LLM_PredictStreamServer) error {
return status.Errorf(codes.Unimplemented, "method PredictStream not implemented")
}
func (UnimplementedLLMServer) Embedding(context.Context, *PredictOptions) (*EmbeddingResult, error) {
return nil, status.Errorf(codes.Unimplemented, "method Embedding not implemented")
}
func (UnimplementedLLMServer) mustEmbedUnimplementedLLMServer() {}
// UnsafeLLMServer may be embedded to opt out of forward compatibility for this service.
@ -210,6 +224,24 @@ func (x *lLMPredictStreamServer) Send(m *Reply) error {
return x.ServerStream.SendMsg(m)
}
func _LLM_Embedding_Handler(srv interface{}, ctx context.Context, dec func(interface{}) error, interceptor grpc.UnaryServerInterceptor) (interface{}, error) {
in := new(PredictOptions)
if err := dec(in); err != nil {
return nil, err
}
if interceptor == nil {
return srv.(LLMServer).Embedding(ctx, in)
}
info := &grpc.UnaryServerInfo{
Server: srv,
FullMethod: "/llm.LLM/Embedding",
}
handler := func(ctx context.Context, req interface{}) (interface{}, error) {
return srv.(LLMServer).Embedding(ctx, req.(*PredictOptions))
}
return interceptor(ctx, in, info, handler)
}
// LLM_ServiceDesc is the grpc.ServiceDesc for LLM service.
// It's only intended for direct use with grpc.RegisterService,
// and not to be introspected or modified (even as a copy)
@ -229,6 +261,10 @@ var LLM_ServiceDesc = grpc.ServiceDesc{
MethodName: "LoadModel",
Handler: _LLM_LoadModel_Handler,
},
{
MethodName: "Embedding",
Handler: _LLM_Embedding_Handler,
},
},
Streams: []grpc.StreamDesc{
{