feat: Add Bitsandbytes quantization for transformer backend enhancement #1775 and fix: Transformer backend error on CUDA #1774 (#1823)

* fixes #1775 and #1774

Add BitsAndBytes Quantization and fixes embedding on CUDA devices

* Manage 4bit and 8 bit quantization

Manage different BitsAndBytes options with the quantization: parameter in yaml

* fix compilation errors on non CUDA environment

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'))