[RL] refactor flash rl weight reload in sglang (#14870)

Co-authored-by: eternally-z <zzywzj@gmail.com>

python/sglang/srt/model_loader/loader.py

@@ -788,6 +788,19 @@ class QuantizedRLModelLoader(DefaultModelLoader):
         """
         logger.info("[QuantizedRL] Initial load with FP8 quantization")
 
+        original_load_weights = model.load_weights
+
+        def load_weights_proxy(weights):
+            if QuantizedRLModelLoader.is_reload_scenario(model):
+                logger.info("[QuantizedRL] Using fast path reload in load_weights")
+                QuantizedRLModelLoader.rebinding_and_load_weights(
+                    model, original_load_weights, weights
+                )
+            else:
+                original_load_weights(weights)
+
+        model.load_weights = load_weights_proxy
+
         model.load_weights(weights)
         original_weights = dict(model.named_parameters())