no message

main.py

@@ -4,18 +4,23 @@ from pydantic import BaseModel
 from huggingface_hub import InferenceClient
 import uvicorn
 from typing import Generator
-import json  # Asegúrate de que esta línea esté al principio del archivo
+import json
 import nltk
 import os
+from transformers import pipeline
 
-
+# Set up the environment for NLTK
 nltk.data.path.append(os.getenv('NLTK_DATA'))
 
+# Initialize the FastAPI app
 app = FastAPI()
 
 # Initialize the InferenceClient with your model
 client = InferenceClient("mistralai/Mistral-7B-Instruct-v0.2")
 
+# Initialize the summarization pipeline
+summarizer = pipeline("summarization", model="sshleifer/distilbart-cnn-12-6")
+
 class Item(BaseModel):
     prompt: str
     history: list
@@ -25,23 +30,24 @@ class Item(BaseModel):
     top_p: float = 0.15
     repetition_penalty: float = 1.0
 
+def summarize_history(history):
+    # Concatenate all history entries into a single string
+    full_history = " ".join(entry['content'] for entry in history if entry['role'] == 'user')
+    # Summarize the history
+    summarized_history = summarizer(full_history, max_length=1024, truncation=True)
+    return summarized_history[0]['summary_text']
+
 def format_prompt(current_prompt, history):
     formatted_history = "<s>"
-    for entry in history:
-        if entry["role"] == "user":
-            formatted_history += f"[USER] {entry['content']} [/USER]"
-        elif entry["role"] == "assistant":
-            formatted_history += f"[ASSISTANT] {entry['content']} [/ASSISTANT]"
+    formatted_history += f"[HISTORY] {history} [/HISTORY]"
     formatted_history += f"[USER] {current_prompt} [/USER]</s>"
     return formatted_history
 
-
 def generate_stream(item: Item) -> Generator[bytes, None, None]:
-    formatted_prompt = format_prompt(f"{item.system_prompt}, {item.prompt}", item.history)
-    # Estimate token count for the formatted_prompt
-    input_token_count = len(nltk.word_tokenize(formatted_prompt))  # NLTK tokenization
+    summarized_history = summarize_history(item.history)
+    formatted_prompt = format_prompt(item.prompt, summarized_history)
+    input_token_count = len(nltk.word_tokenize(formatted_prompt))
 
-    # Ensure total token count doesn't exceed the maximum limit
     max_tokens_allowed = 32768
     max_new_tokens_adjusted = max(1, min(item.max_new_tokens, max_tokens_allowed - input_token_count))
 
@@ -54,18 +60,15 @@ def generate_stream(item: Item) -> Generator[bytes, None, None]:
         "seed": 42,
     }
 
-    # Stream the response from the InferenceClient
     for response in client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True):
-        # This assumes 'details=True' gives you a structure where you can access the text like this
         chunk = {
             "text": response.token.text,
-            "complete": response.generated_text is not None  # Adjust based on how you detect completion
+            "complete": response.generated_text is not None
         }
         yield json.dumps(chunk).encode("utf-8") + b"\n"
 
 @app.post("/generate/")
 async def generate_text(item: Item):
-    # Stream response back to the client
     return StreamingResponse(generate_stream(item), media_type="application/x-ndjson")
 
 if __name__ == "__main__":

requirements.txt

@@ -3,4 +3,5 @@ uvicorn
 huggingface_hub
 pydantic
 torch==2.0.0
-nltk
+nltk
+transformers