Update handler.py

handler.py

@@ -1,53 +1,38 @@
-from typing import Dict, Any
-import logging
+from typing import Dict, List, Any
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import transformers
+import torch
 
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from peft import PeftConfig, PeftModel
-import torch.cuda
+dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
 
-
-LOGGER = logging.getLogger(__name__)
-logging.basicConfig(level=logging.INFO)
-device = "cuda" if torch.cuda.is_available() else "cpu"
-
-
-class EndpointHandler():
+class EndpointHandler:
     def __init__(self, path=""):
-        config = PeftConfig.from_pretrained(path)
-        model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, load_in_4bit=True, device_map='auto')
-        self.tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
-        # Load the Lora model
-        self.model = PeftModel.from_pretrained(model, path)
 
-    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
-        """
-        Args:
-            data (Dict): The payload with the text prompt and generation parameters.
-        """
-        LOGGER.info(f"Received data: {data}")
-        # Get inputs
-        query = data.pop("inputs", None)
-        prompt_template = """
-        Below is a screenplay prompt followed by a screenplay response. Generate only screenplay response.
-        ### Screenplay Prompt:
-        {query}
-        ### Screenplay Response:
-        """
-        prompt = prompt_template.format(query=query)
-        parameters = data.pop("parameters", None)
-        if prompt is None:
-            raise ValueError("Missing prompt.")
-        # Preprocess
-        encodeds = self.tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
-
-        model_inputs = encodeds.to(device)
+        tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
+        model = AutoModelForCausalLM.from_pretrained(
+            path,
+            return_dict=True,
+            device_map="auto",
+            load_in_8bit=True,
+            torch_dtype=dtype,
+            trust_remote_code=True,
+        )
+
+        generation_config = model.generation_config
+        generation_config.max_new_tokens = 1000000
+        generation_config.temperature = 0
+        generation_config.num_return_sequences = 1
+        generation_config.pad_token_id = tokenizer.eos_token_id
+        generation_config.eos_token_id = tokenizer.eos_token_id
+        self.generation_config = generation_config
+
+        self.pipeline = transformers.pipeline(
+            "text-generation",
+            model=model,
+            tokenizer=tokenizer
+        )
 
-        # Forward
-        LOGGER.info(f"Start generation.")
-        eos_tok = self.tokenizer.eos_token_id
-        LOGGER.info(f"Generating Ids")
-        generated_ids = self.model.generate(**model_inputs, max_new_tokens=9999999, do_sample=True, pad_token_id=eos_tok)
-        LOGGER.info(f"Ids Generated.")
-        decoded = self.tokenizer.batch_decode(generated_ids)
-        LOGGER.info(f"Generated text length: {len(decoded[0])}")
-        return {"generated_text": decoded[0]}
+    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
+        prompt = data.pop("inputs", data)
+        result = self.pipeline(prompt, generation_config=self.generation_config)
+        return result