this is work. First prompt

app/config.py

@@ -12,7 +12,7 @@ class Config(BaseSettings):
     MAX_CHUNK_SIZE: int = 2048
     CHUNK_OVERLAP: int = 50
     LM_MODEL_NAME: str = "/models/Qwen3-Embedding-0.6B"
-    LOCAL_LLM_NAME: str = "/models/Qwen3-8B"
+    LOCAL_LLM_NAME: str = "/models/Qwen3-4B"
     QWEN_MODEL_NAME: str = "Qwen3-8B"
 
 

app/main.py

@@ -3,6 +3,7 @@ import gradio as gr
 import torch
 import asyncio
 import threading
+from transformers import AutoTokenizer as HFTokenizer
 from loguru import logger
 from langchain_chroma import Chroma
 from langchain_huggingface import HuggingFaceEmbeddings
@@ -31,13 +32,13 @@ class ChatWithAI:
                 tokenizer_mode="auto",
                 trust_remote_code=True,
                 dtype=torch.float16,  # float16
-                gpu_memory_utilization=0.85,
-                max_model_len=16384,
+                gpu_memory_utilization=0.5,
+                max_model_len=32768,
                 enable_prefix_caching=True,
                 tensor_parallel_size=1,  # Увеличь, если у тебя несколько GPU
             )
             self.engine = AsyncLLMEngine.from_engine_args(engine_args)
-            self.tokenizer = self.engine.engine.tokenizer.tokenizer  # доступ к токенизатору
+            self.tokenizer = HFTokenizer.from_pretrained(settings.LM_MODEL_NAME, trust_remote_code=True)
         else:
             raise ValueError(f"Неподдерживаемый провайдер: {provider}")