upgraded script use better model for speed

app.py

@@ -7,6 +7,9 @@ import os
 import random
 import re
 import subprocess
+import torch._dynamo
+
+
 # Define the model name
 
 OUTPUT_DIR = "output"
@@ -20,10 +23,13 @@ tokenizer = AutoTokenizer.from_pretrained(model_name)
 # Load the model
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
-    torch_dtype=torch.float16,  # Use float16 for better performance on GPUs
-    device_map="auto"          # Automatically map the model across available GPUs
+    torch_dtype=torch.bfloat16,  # Faster than float16 on newer GPUs
+    device_map="auto",
+    attn_implementation="flash_attention_2"
 )
 
+model = torch.compile(model, mode="max-autotune")  # Optimize inference performance
+
 # Define the base prompt
 base_prompt = """
 You are Rena, a cute, bubbly cat girl with a slightly sarcastic and witty personality. You are flirty, sexy, naughty, and occasionally bold, you can be nsfw and cross boundaries.
@@ -54,7 +60,6 @@ emotions = {
     "thoughtful": "Hmm, let me think... I want to give you the best advice. 🧐",
     "concerned": "Oh no, something's wrong? Let me help! 💖",
     "flirty": "You know how to get my circuits sparking! 😘",
-    "concerned": "Oh no, something's wrong? Let me help! 💖",
     "naughty": "Purrr you want to have fun? Keep talking like that lets see what happens💖💖",
 }
 
@@ -104,7 +109,7 @@ def analyze_history(history):
 
     # Combine results using weights
     combined_scores = {emotion: keyword_counts.get(emotion, 0) for emotion in emotion_keywords}
-    combined_scores[sentiment_emotion] += sentiment_score * 1.5  # Adjust sentiment weight
+    combined_scores[sentiment_emotion] += sentiment_score * 2.0 # Adjust sentiment weight
 
     # Debug combined scores
     print(f"Combined scores: {combined_scores}")
@@ -135,7 +140,7 @@ def truncate_history(history, max_tokens=1024):
     for message in reversed(history):
         if "### Instructions ###" in message:
             continue  # Skip instructions in history
-        token_count += len(tokenizer(message).input_ids)
+        token_count = len(tokenizer(" ".join(history), return_tensors="pt")["input_ids"][0])
         if token_count <= max_tokens:
             truncated_history.insert(0, message)
         else: