Update app/main.py

app/main.py

@@ -1,21 +1,29 @@
 from fastapi import FastAPI
 from pydantic import BaseModel
-from transformers import pipeline, AutoTokenizer
+from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
 from glin_profanity import Filter
 import re
 
-# 1. Models
+# 1. Toxicity classifier
 toxicity_pipe = pipeline("text-classification", model="textdetox/twitter-xlmr-toxicity-classifier")
+
+# 2. Phi-3 Model - The FIX is here
 model_id = "microsoft/Phi-3-mini-4k-instruct"
 tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-llm_pipe = pipeline("text-generation", model=model_id, tokenizer=tokenizer, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    trust_remote_code=True,
+    attn_implementation="eager",  # ✅ This resolves the window_size error
+    device_map="auto"
+)
+llm_pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
+
 profanity_filter = Filter({"languages": ["english", "spanish", "french"], "replace_with": "***"})
 
 app = FastAPI(title="Multilingual Moderation API")
 
-# 2. Request Model - THIS IS THE CRITICAL PART
 class TextRequest(BaseModel):
-    text: str  # <--- It MUST be 'text', not 'data'
+    text: str
 
 class ModerationResponse(BaseModel):
     original_text: str
@@ -32,7 +40,7 @@ threat_pattern = re.compile("|".join(THREAT_KEYWORDS), re.IGNORECASE)
 
 @app.post("/moderate", response_model=ModerationResponse)
 async def moderate_text(request: TextRequest):
-    text = request.text # Using 'text' correctly
+    text = request.text
     
     profanity_result = profanity_filter.check_profanity(text)
     tox_result = toxicity_pipe(text)[0]