fix(llama.cpp): fix eos without cache

Signed-off-by: GitHub <noreply@github.com>
Co-authored-by: mudler <mudler@users.noreply.github.com>

Makefile

@@ -4,11 +4,11 @@ GOVET=$(GOCMD) vet
 BINARY_NAME=local-ai
 
 # llama.cpp versions
-GOLLAMA_VERSION?=aeba71ee842819da681ea537e78846dc75949ac0
+GOLLAMA_VERSION?=6a8041ef6b46d4712afc3ae791d1c2d73da0ad1c
 
 GOLLAMA_STABLE_VERSION?=50cee7712066d9e38306eccadcfbb44ea87df4b7
 
-CPPLLAMA_VERSION?=19885d205e768579ab090d1e99281cae58c21b54
+CPPLLAMA_VERSION?=4755afd1cbd40d93c017e5b98c39796f52345314
 
 # gpt4all version
 GPT4ALL_REPO?=https://github.com/nomic-ai/gpt4all
@@ -19,13 +19,13 @@ RWKV_REPO?=https://github.com/donomii/go-rwkv.cpp
 RWKV_VERSION?=661e7ae26d442f5cfebd2a0881b44e8c55949ec6
 
 # whisper.cpp version
-WHISPER_CPP_VERSION?=37a709f6558c6d9783199e2b8cbb136e1c41d346
+WHISPER_CPP_VERSION?=a56f435fd475afd7edf02bfbf9f8c77f527198c2
 
 # bert.cpp version
 BERT_VERSION?=6abe312cded14042f6b7c3cd8edf082713334a4d
 
 # go-piper version
-PIPER_VERSION?=d6b6275ba037dabdba4a8b65dfdf6b2a73a67f07
+PIPER_VERSION?=9d0100873a7dbb0824dfea40e8cec70a1b110759
 
 # stablediffusion version
 STABLEDIFFUSION_VERSION?=362df9da29f882dbf09ade61972d16a1f53c3485
@@ -91,10 +91,13 @@ ifeq ($(BUILD_TYPE),openblas)
 	export WHISPER_OPENBLAS=1
 endif
 
+
 ifeq ($(BUILD_TYPE),cublas)
-	CGO_LDFLAGS+=-lcublas -lcudart -L$(CUDA_LIBPATH)
+	CGO_LDFLAGS+=-lcublas -lcudart -lculibos -lcublasLt -L$(CUDA_LIBPATH)
 	export LLAMA_CUBLAS=1
+# required by whisper.cpp
 	export WHISPER_CUBLAS=1
+	CGO_LDFLAGS+=-L$(CUDA_PATH)/stubs -lcuda
 endif
 
 ifeq ($(BUILD_TYPE),hipblas)
@@ -462,9 +465,6 @@ backend-assets/grpc/llama: backend-assets/grpc sources/go-llama/libbinding.a
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-llama LIBRARY_PATH=$(CURDIR)/sources/go-llama \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama ./backend/go/llm/llama/
 # TODO: every binary should have its own folder instead, so can have different  implementations
-ifeq ($(BUILD_TYPE),metal)
-	cp backend/cpp/llama/llama.cpp/ggml-metal.metal backend-assets/grpc/
-endif
 
 ## BACKEND CPP LLAMA START
 # Sets the variables in case it has to build the gRPC locally.
@@ -494,7 +494,7 @@ backend-assets/grpc/llama-cpp: backend-assets/grpc backend/cpp/llama/grpc-server
 	cp -rfv backend/cpp/llama/grpc-server backend-assets/grpc/llama-cpp
 # TODO: every binary should have its own folder instead, so can have different metal implementations
 ifeq ($(BUILD_TYPE),metal)
-	cp backend/cpp/llama/llama.cpp/build/bin/ggml-metal.metal backend-assets/grpc/
+	cp backend/cpp/llama/llama.cpp/build/bin/default.metallib backend-assets/grpc/
 endif
 
 backend-assets/grpc/llama-ggml: backend-assets/grpc sources/go-llama-ggml/libbinding.a

backend/cpp/llama/Makefile

@@ -18,6 +18,7 @@ else ifeq ($(BUILD_TYPE),clblas)
 # If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ 
 else ifeq ($(BUILD_TYPE),hipblas)
 	CMAKE_ARGS+=-DLLAMA_HIPBLAS=ON
+# If it's OSX, DO NOT embed the metal library - -DLLAMA_METAL_EMBED_LIBRARY=ON requires further investigation
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f16)

backend/cpp/llama/grpc-server.cpp

@@ -1084,7 +1084,7 @@ struct llama_server_context
             slot.has_next_token = false;
         }
 
-        if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(model))
+        if (result.tok == llama_token_eos(model))
         {
             slot.stopped_eos = true;
             slot.has_next_token = false;

backend/python/common-env/transformers/transformers-nvidia.yml

@@ -30,6 +30,7 @@ dependencies:
       - async-timeout==4.0.3
       - attrs==23.1.0
       - bark==0.1.5
+      - bitsandbytes==0.43.0
       - boto3==1.28.61
       - botocore==1.31.61
       - certifi==2023.7.22

backend/python/transformers/transformers_server.py

@@ -23,7 +23,7 @@ if XPU:
     from intel_extension_for_transformers.transformers.modeling import AutoModelForCausalLM
     from transformers import AutoTokenizer, AutoModel, set_seed
 else:
-    from transformers import AutoTokenizer, AutoModel, AutoModelForCausalLM, set_seed
+    from transformers import AutoTokenizer, AutoModel, AutoModelForCausalLM, set_seed, BitsAndBytesConfig
 
 
 _ONE_DAY_IN_SECONDS = 60 * 60 * 24
@@ -75,18 +75,50 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             A Result object that contains the result of the LoadModel operation.
         """
         model_name = request.Model
+
+        compute = "auto"
+        if request.F16Memory == True:
+            compute=torch.bfloat16
+
+        self.CUDA = request.CUDA
+
+        device_map="cpu"
+
+        quantization = None
+
+        if self.CUDA:
+            if request.Device:
+                device_map=request.Device
+            else:
+                device_map="cuda:0"
+            if request.Quantization == "bnb_4bit":
+                quantization = BitsAndBytesConfig(
+                    load_in_4bit = True,
+                    bnb_4bit_compute_dtype = compute,
+                    bnb_4bit_quant_type = "nf4",
+                    bnb_4bit_use_double_quant = True,
+                    load_in_8bit = False,
+                )
+            elif request.Quantization == "bnb_8bit":
+                quantization = BitsAndBytesConfig(
+                    load_in_4bit=False,
+                    bnb_4bit_compute_dtype = None,
+                    load_in_8bit=True,                                   
+                )
+                                                   
+    
         try:
             if request.Type == "AutoModelForCausalLM":
                 if XPU:
+                    if quantization == "xpu_4bit":
+                        xpu_4bit = True
                     self.model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=request.TrustRemoteCode,
-                                              device_map="xpu", load_in_4bit=True)
+                                              device_map="xpu", load_in_4bit=xpu_4bit)
                 else:
-                    self.model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=request.TrustRemoteCode)
+                    self.model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=request.TrustRemoteCode, use_safetensors=True, quantization_config=quantization, device_map=device_map, torch_dtype=compute)
             else:
-                self.model = AutoModel.from_pretrained(model_name, trust_remote_code=request.TrustRemoteCode)
+                self.model = AutoModel.from_pretrained(model_name, trust_remote_code=request.TrustRemoteCode,  use_safetensors=True,  quantization_config=quantization, device_map=device_map, torch_dtype=compute)
-
+            self.tokenizer = AutoTokenizer.from_pretrained(model_name, use_safetensors=True)
-            self.tokenizer = AutoTokenizer.from_pretrained(model_name)
-            self.CUDA = False
             self.XPU = False
 
             if XPU:
@@ -97,13 +129,6 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 except Exception as err:
                     print("Not using XPU:", err, file=sys.stderr)
 
-            if request.CUDA or torch.cuda.is_available():
-                try:
-                    print("Loading model", model_name, "to CUDA.", file=sys.stderr)
-                    self.model = self.model.to("cuda")
-                    self.CUDA = True
-                except Exception as err:
-                    print("Not using CUDA:", err, file=sys.stderr)
         except Exception as err:
             return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
         # Implement your logic here for the LoadModel service
@@ -130,13 +155,17 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         encoded_input = self.tokenizer(request.Embeddings, padding=True, truncation=True, max_length=max_length, return_tensors="pt")    
 
         # Create word embeddings
-        model_output = self.model(**encoded_input)
+        if self.CUDA:
+            encoded_input = encoded_input.to("cuda")
+
+        with torch.no_grad():    
+            model_output = self.model(**encoded_input)
 
         # Pool to get sentence embeddings; i.e. generate one 1024 vector for the entire sentence
-        sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask']).detach().numpy()
+        sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
         print("Calculated embeddings for: " + request.Embeddings, file=sys.stderr)
         print("Embeddings:", sentence_embeddings, file=sys.stderr)
-        return backend_pb2.EmbeddingResult(embeddings=sentence_embeddings)
+        return backend_pb2.EmbeddingResult(embeddings=sentence_embeddings[0])
 
     def Predict(self, request, context):
         """
@@ -163,12 +192,8 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         if XPU:
             inputs = inputs.to("xpu")
 
-        outputs = self.model.generate(inputs,max_new_tokens=max_tokens, temperature=request.Temperature, top_p=request.TopP)
+        outputs = self.model.generate(inputs,max_new_tokens=max_tokens, temperature=request.Temperature, top_p=request.TopP, do_sample=True, pad_token_id=self.tokenizer.eos_token_id)
-
+        generated_text = self.tokenizer.batch_decode(outputs[:, inputs.shape[1]:], skip_special_tokens=True)[0]
-        generated_text = self.tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
-        # Remove prompt from response if present
-        if request.Prompt in generated_text:
-            generated_text = generated_text.replace(request.Prompt, "")
 
         return backend_pb2.Reply(message=bytes(generated_text, encoding='utf-8'))
 

core/backend/embeddings.go

@@ -10,10 +10,6 @@ import (
 )
 
 func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, backendConfig config.BackendConfig, appConfig *config.ApplicationConfig) (func() ([]float32, error), error) {
-	if !backendConfig.Embeddings {
-		return nil, fmt.Errorf("endpoint disabled for this model by API configuration")
-	}
-
 	modelFile := backendConfig.Model
 
 	grpcOpts := gRPCModelOpts(backendConfig)