v5.1: revert INT8 — too lossy for 0.6B, keep LoRA merge + float32

app.py

@@ -80,18 +80,7 @@ def _load_model():
             )
 
         MODEL.eval()
-
-        # Dynamic INT8 quantization — now safe since LoRA is merged
-        try:
-            MODEL = torch.quantization.quantize_dynamic(
-                MODEL, {torch.nn.Linear}, dtype=torch.qint8,
-            )
-            INFERENCE_MODE = "pytorch-int8"
-            print("INT8 quantization applied")
-        except Exception as e:
-            INFERENCE_MODE = "pytorch"
-            print(f"Quantization skipped: {e}")
-
+        INFERENCE_MODE = "pytorch-merged"
         print(f"Model ready: {INFERENCE_MODE}")
 
     except Exception as e:
@@ -110,7 +99,7 @@ def health():
         "version": "5.0.0",
         "architecture": "Qwen3-0.6B + LoRA r=64 (merged)",
         "inference": INFERENCE_MODE,
-        "features": ["streaming", "thinking", "int8"],
+        "features": ["streaming", "thinking", "merged"],
         "model_dir": MODEL_DIR,
         "error": LOAD_ERROR,
     }