Update app.py

app.py

@@ -1,7 +1,6 @@
 from fastapi import FastAPI, Request
-from fastapi.responses import StreamingResponse, FileResponse
+from fastapi.responses import StreamingResponse
 from fastapi.middleware.cors import CORSMiddleware
-from fastapi.staticfiles import StaticFiles
 import onnxruntime as ort
 import numpy as np
 import tiktoken
@@ -10,22 +9,33 @@ import os
 
 app = FastAPI()
 
-# CORS erlauben
-app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
+# WICHTIG: Erlaubt deinem externen Frontend den Zugriff
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"], # Hier kannst du später deine Domain eintragen
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
 
 # Modell & Tokenizer laden
 tokenizer = tiktoken.get_encoding("gpt2")
 MODEL_PATH = "SmaLLMPro_350M_int8.onnx"
-# Nutzt optimierte CPU-Einstellungen
-session = ort.InferenceSession(MODEL_PATH, providers=['CPUExecutionProvider'])
+
+# Optimierte Session-Optionen für CPU
+sess_options = ort.SessionOptions()
+sess_options.intra_op_num_threads = 2 # HF Spaces haben meist 2 Kerne
+session = ort.InferenceSession(MODEL_PATH, sess_options, providers=['CPUExecutionProvider'])
 
 def top_k_sample(logits, k=50, temp=0.7):
-    logits = logits / temp
-    # Numpy ist VIEL schneller als JS-Schleifen
+    logits = logits / max(temp, 1e-6)
+    # Nur die Top-K Werte betrachten (spart massiv Zeit beim Sortieren)
     top_k_indices = np.argpartition(logits, -k)[-k:]
     top_k_logits = logits[top_k_indices]
+    
+    # Stabiler Softmax
     exp_logits = np.exp(top_k_logits - np.max(top_k_logits))
     probs = exp_logits / np.sum(exp_logits)
+    
     return int(np.random.choice(top_k_indices, p=probs))
 
 @app.post("/chat")
@@ -34,20 +44,34 @@ async def chat(request: Request):
     prompt = f"Instruction:\n{data['prompt']}\n\nResponse:\n"
     tokens = tokenizer.encode(prompt)
     
+    max_len = int(data.get('maxLen', 100))
+    temp = float(data.get('temp', 0.7))
+    top_k = int(data.get('topK', 40))
+
     async def generate():
         nonlocal tokens
-        for _ in range(data.get('maxLen', 100)):
+        for _ in range(max_len):
+            # Kontext auf 1024 beschränken
             ctx = tokens[-1024:]
+            # Padding (Rechtsbündig)
             padded = np.zeros((1, 1024), dtype=np.int64)
             padded[0, -len(ctx):] = ctx
             
+            # Inferenz
             outputs = session.run(None, {'input': padded})
+            # Wir nehmen nur die Logits des letzten Tokens
             logits = outputs[0][0, -1, :50304]
             
-            next_token = top_k_sample(logits, k=data.get('topK', 25), temp=data.get('temp', 0.5))
-            if next_token == 50256: break
+            next_token = top_k_sample(logits, k=top_k, temp=temp)
+            
+            if next_token == 50256: # EOS
+                break
             
             tokens.append(next_token)
             yield f"data: {json.dumps({'token': tokenizer.decode([next_token])})}\n\n"
 
-    return StreamingResponse(generate(), media_type="text/event-stream")
+    return StreamingResponse(generate(), media_type="text/event-stream")
+
+@app.get("/")
+def health():
+    return {"status": "SmaLLMPro API is online"}