updated reqquirements for flash

app.py

@@ -7,6 +7,8 @@ import os
 import random
 import re
 import subprocess
+torch.cuda.empty_cache()  # Clears unused GPU memory
+torch.cuda.memory_allocated()  # Checks available GPU memory
 
 # Define the model name
 OUTPUT_DIR = "output"
@@ -20,10 +22,12 @@ tokenizer = AutoTokenizer.from_pretrained(model_name)
 # Load the model (without Flash Attention)
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
-    torch_dtype=torch.float16,  # Use float16 for performance
-    device_map="auto"
+    torch_dtype=torch.float16,  # ✅ Keeps precision while reducing memory usage
+    device_map="auto",  # ✅ Auto-allocates GPU/CPU memory
+    attn_implementation="triton",  # ✅ Uses Triton for faster inference
+    low_cpu_mem_usage=True  # ✅ Reduces RAM usage
 )
-
+torch.backends.cuda.matmul.allow_tf32 = True
 # Use torch.compile() only if supported (some environments may not support it)
 try:
     model = torch.compile(model, mode="max-autotune")  # Optimize inference performance
@@ -78,7 +82,6 @@ emotion_keywords = {
     "curious": ["curious", "wonder", "question", "thinking"],
     "thoughtful": ["sad", "thoughtful", "hmm", "ponder", "upset"],
     "concerned": ["error", "wrong", "problem", "issue", "stuck"],
-    "mischievous": ["trouble", "mischief", "sneaky", "prank"],
     "naughty": ["sexy", "naughty", "slut", "bad girl","slutty", "tease", "whore", "cum", "tits", "ass", "shake", "tail", "pussy" ]
 }
     

requirements.txt

@@ -6,6 +6,7 @@ accelerate
 spaces
 bitsandbytes
 autoawq
+triton
 transformers>=4.37.0