Spaces:

Gaston895
/

Q3

Sleeping

App Files Files Community

Gaston895 commited on Mar 3

Commit

1cd3db1

verified ·

1 Parent(s): ca9c3b8

Upload app.py

Browse files

Files changed (1) hide show

app.py +45 -115

app.py CHANGED Viewed

@@ -9,9 +9,10 @@ from huggingface_hub import hf_hub_download
 app = Flask(__name__)
 CORS(app)
-# Qwen3-14B Claude 4.5 Opus High Reasoning Distill Model
-MODEL_REPO = "Gaston895/Qwen3-14B-Claude-4.5-Opus-High-Reasoning-Distill-GGUF"
-MODEL_FILE = "Qwen3-14B-Claude-4.5-Opus-Distill.q4_k_m.gguf"
 model = None
 loading_error = None
@@ -22,7 +23,7 @@ def load_model():
         # Use HF_TOKEN if your space is private
         token = os.environ.get("HF_TOKEN")
-        print(f"📥 Downloading model from: {MODEL_REPO}/{MODEL_FILE}...")
         # Download from HuggingFace model repository
         model_path = hf_hub_download(
@@ -34,23 +35,18 @@ def load_model():
         print(f"✅ Model downloaded to: {model_path}")
         print("🏗️ Initializing model engine (llama-cpp)...")
-        # Use extremely conservative settings for 14B model stability
-        # The model is crashing due to memory/batch issues
         model = Llama(
             model_path=model_path,
-            n_ctx=1024,  # Very small context for stability
-            n_threads=1,  # Single thread to avoid race conditions
-            n_batch=32,   # Very small batch size
-            n_gpu_layers=0,  # Force CPU-only for stability
-            use_mmap=True,
-            use_mlock=False,
-            verbose=False,
-            seed=42  # Fixed seed for reproducibility
         )
-        print("✅ Model loaded successfully!")
     except Exception as e:
         loading_error = str(e)
-        print(f"❌ Error loading model: {e}")
 # Start loading in background
 threading.Thread(target=load_model, daemon=True).start()
@@ -62,7 +58,7 @@ def index():
     <!DOCTYPE html>
     <html>
     <head>
-        <title>OpenGSSTEC AI API</title>
         <style>
             body { font-family: Arial, sans-serif; max-width: 800px; margin: 50px auto; padding: 20px; }
             h1 { color: #333; }
@@ -71,10 +67,14 @@ def index():
             .status { padding: 10px; border-radius: 5px; margin: 20px 0; }
             .online { background: #d4edda; color: #155724; }
             .loading { background: #fff3cd; color: #856404; }
         </style>
     </head>
     <body>
-        <h1>🤖 OpenGSSTEC AI API - Qwen3-14B</h1>
         <div class="status """ + ("online" if model else "loading") + """">
             Status: """ + ("✅ Online and Ready" if model else "⏳ Loading Model...") + """
         </div>
@@ -84,31 +84,21 @@ def index():
         <div class="endpoint">
             <h3>GET /health</h3>
             <p>Check API health and model status</p>
-            <code>curl https://gsstec-open.hf.space/health</code>
         </div>
         <div class="endpoint">
             <h3>POST /chat</h3>
             <p>Send chat messages to the AI model</p>
-            <pre><code>curl -X POST https://gsstec-open.hf.space/chat \\
-  -H "Content-Type: application/json" \\
-  -d '{
-    "messages": [
-      {"role": "user", "content": "Hello!"}
-    ]
-  }'</code></pre>
         </div>
         <h2>Model Information</h2>
         <ul>
-            <li><strong>Model:</strong> Qwen3-14B Claude 4.5 Opus High Reasoning Distill (Q4_K_M)</li>
             <li><strong>Repository:</strong> """ + MODEL_REPO + """</li>
-            <li><strong>Parameters:</strong> 14B</li>
-            <li><strong>Context Length:</strong> 8192 tokens (limited to 2048 for stability)</li>
-            <li><strong>Specialization:</strong> High reasoning and instruction following</li>
         </ul>
-        <p><a href="/health">Check Health Status →</a></p>
     </body>
     </html>
     """
@@ -120,102 +110,46 @@ def health():
         "status": "online" if model else "loading",
         "repo": MODEL_REPO,
         "file": MODEL_FILE,
-        "memory_limit": "1GB",
         "error": loading_error
     })
 @app.route('/chat', methods=['POST'])
 def chat():
-    print("=" * 60)
-    print("📨 Received /chat request")
     if not model:
-        print("❌ Model not loaded yet")
         return jsonify({"error": "Model still loading"}), 503
     try:
-        # Log raw request
-        print(f"📦 Content-Type: {request.content_type}")
-        print(f"📦 Raw data (first 500 bytes): {request.data[:500]}")
         data = request.json
-        print(f"📊 Parsed JSON keys: {list(data.keys()) if data else 'None'}")
         messages = data.get('messages', [])
-        print(f"💬 Message count: {len(messages)}")
-        # Calculate approximate token count (rough estimate: 1 token ≈ 4 chars)
-        total_chars = sum(len(str(msg.get('content', ''))) for msg in messages)
-        estimated_tokens = total_chars // 4
-        print(f"📊 Estimated input tokens: {estimated_tokens}")
-        # If too large, truncate conversation history but keep system prompt
-        MAX_INPUT_TOKENS = 500  # Much smaller for 14B model stability
-        if estimated_tokens > MAX_INPUT_TOKENS:
-            print(f"⚠️ Input too large, truncating conversation history...")
-            # Keep system message and recent messages only
-            system_msg = messages[0] if messages and messages[0].get('role') == 'system' else None
-            user_messages = [m for m in messages if m.get('role') != 'system']
-            # Keep only last message for maximum stability
-            keep_count = 1  # Only keep the most recent message
-            truncated_messages = []
-            if system_msg:
-                # Truncate system message if it's too long
-                system_content = system_msg.get('content', '')
-                if len(system_content) > 1000:  # ~250 tokens
-                    system_content = system_content[:1000] + "\n\n[System prompt truncated for context limit]"
-                    print(f"⚠️ System prompt truncated from {len(system_msg.get('content', ''))} to {len(system_content)} chars")
-                truncated_messages.append({'role': 'system', 'content': system_content})
-            truncated_messages.extend(user_messages[-keep_count:])
-            messages = truncated_messages
-            print(f"✂️ Truncated to {len(messages)} messages")
-        # Format for Qwen2.5 Instruct (ChatML format)
         prompt = ""
         for msg in messages:
             role = msg.get('role', 'user')
             content = msg.get('content', '')
-            prompt += f"<|im_start|>{role}\n{content}<|im_end|>\n"
-        prompt += "<|im_start|>assistant\n"
-        # Additional safety: Hard limit on prompt length to avoid batch decode errors
-        MAX_PROMPT_CHARS = 2000  # Much smaller for 14B model stability
-        if len(prompt) > MAX_PROMPT_CHARS:
-            print(f"⚠️ Prompt too long ({len(prompt)} chars), truncating to {MAX_PROMPT_CHARS}...")
-            # Keep the beginning (system prompt) and end (recent context)
-            keep_start = MAX_PROMPT_CHARS // 2
-            keep_end = MAX_PROMPT_CHARS // 2
-            prompt = prompt[:keep_start] + "\n\n[...conversation truncated...]\n\n" + prompt[-keep_end:]
-            print(f"✂️ Prompt truncated to {len(prompt)} chars")
-        print(f"📝 Prompt length: {len(prompt)} chars")
-        print(f"📝 Prompt preview: {prompt[:200]}...")
-        print("🤖 Calling model...")
-        try:
-            output = model(
-                prompt,
-                max_tokens=200,  # Much smaller for stability
-                temperature=0.3,  # Lower temperature for more stable output
-                top_p=0.8,
-                stop=["<|im_end|>", "<|endoftext|>"],
-                echo=False
-            )
-        except Exception as model_error:
-            print(f"❌ Model inference failed: {model_error}")
-            # Return a fallback response
-            return jsonify({
-                "choices": [{
-                    "message": {"role": "assistant", "content": "I apologize, but I'm experiencing technical difficulties. The model is currently unstable. Please try a shorter message or try again later."},
-                    "finish_reason": "error"
-                }]
-            }), 503
-        print("✅ Model completed")
         response_text = output["choices"][0]["text"].strip()
-        print(f"💬 Response length: {len(response_text)} chars")
         return jsonify({
             "choices": [{
@@ -224,11 +158,7 @@ def chat():
             }]
         })
     except Exception as e:
-        print(f"❌ ERROR: {str(e)}")
-        print(f"❌ Type: {type(e).__name__}")
-        import traceback
-        traceback.print_exc()
-        return jsonify({"error": str(e), "type": type(e).__name__}), 500
 if __name__ == '__main__':
     port = int(os.environ.get('PORT', 7860))

 app = Flask(__name__)
 CORS(app)
+# Fallback to a smaller, more stable model if the 14B model fails
+# This is a proven stable configuration
+MODEL_REPO = "microsoft/Phi-3-mini-4k-instruct-gguf"
+MODEL_FILE = "Phi-3-mini-4k-instruct-q4.gguf"
 model = None
 loading_error = None
         # Use HF_TOKEN if your space is private
         token = os.environ.get("HF_TOKEN")
+        print(f"📥 Downloading fallback model from: {MODEL_REPO}/{MODEL_FILE}...")
         # Download from HuggingFace model repository
         model_path = hf_hub_download(
         print(f"✅ Model downloaded to: {model_path}")
         print("🏗️ Initializing model engine (llama-cpp)...")
+        # Use very conservative settings for maximum stability
         model = Llama(
             model_path=model_path,
+            n_ctx=2048,  # Phi-3 mini works well with 2K context
+            n_threads=2,
+            n_batch=128,
+            verbose=False
         )
+        print("✅ Fallback model loaded successfully!")
     except Exception as e:
         loading_error = str(e)
+        print(f"❌ Error loading fallback model: {e}")
 # Start loading in background
 threading.Thread(target=load_model, daemon=True).start()
     <!DOCTYPE html>
     <html>
     <head>
+        <title>OpenGSSTEC AI API - Fallback Mode</title>
         <style>
             body { font-family: Arial, sans-serif; max-width: 800px; margin: 50px auto; padding: 20px; }
             h1 { color: #333; }
             .status { padding: 10px; border-radius: 5px; margin: 20px 0; }
             .online { background: #d4edda; color: #155724; }
             .loading { background: #fff3cd; color: #856404; }
+            .warning { background: #f8d7da; color: #721c24; }
         </style>
     </head>
     <body>
+        <h1>🤖 OpenGSSTEC AI API - Fallback Mode</h1>
+        <div class="status warning">
+            ⚠️ Running in fallback mode with Phi-3-mini due to 14B model instability
+        </div>
         <div class="status """ + ("online" if model else "loading") + """">
             Status: """ + ("✅ Online and Ready" if model else "⏳ Loading Model...") + """
         </div>
         <div class="endpoint">
             <h3>GET /health</h3>
             <p>Check API health and model status</p>
         </div>
         <div class="endpoint">
             <h3>POST /chat</h3>
             <p>Send chat messages to the AI model</p>
         </div>
         <h2>Model Information</h2>
         <ul>
+            <li><strong>Model:</strong> Phi-3-mini-4k-instruct (Q4)</li>
             <li><strong>Repository:</strong> """ + MODEL_REPO + """</li>
+            <li><strong>Parameters:</strong> 3.8B</li>
+            <li><strong>Context Length:</strong> 2048 tokens</li>
+            <li><strong>Status:</strong> Fallback mode - stable and reliable</li>
         </ul>
     </body>
     </html>
     """
         "status": "online" if model else "loading",
         "repo": MODEL_REPO,
         "file": MODEL_FILE,
+        "mode": "fallback",
         "error": loading_error
     })
 @app.route('/chat', methods=['POST'])
 def chat():
     if not model:
         return jsonify({"error": "Model still loading"}), 503
     try:
         data = request.json
         messages = data.get('messages', [])
+        # Simple prompt formatting for Phi-3
         prompt = ""
         for msg in messages:
             role = msg.get('role', 'user')
             content = msg.get('content', '')
+            if role == 'system':
+                prompt += f"System: {content}\n"
+            elif role == 'user':
+                prompt += f"User: {content}\n"
+            elif role == 'assistant':
+                prompt += f"Assistant: {content}\n"
+        prompt += "Assistant: "
+        # Keep prompt reasonable
+        if len(prompt) > 1500:
+            prompt = prompt[-1500:]
+        output = model(
+            prompt,
+            max_tokens=300,
+            temperature=0.7,
+            top_p=0.9,
+            stop=["User:", "System:"],
+            echo=False
+        )
         response_text = output["choices"][0]["text"].strip()
         return jsonify({
             "choices": [{
             }]
         })
     except Exception as e:
+        return jsonify({"error": str(e)}), 500
 if __name__ == '__main__':
     port = int(os.environ.get('PORT', 7860))