Unify the memory pool api and tp worker API (#1724)

python/sglang/srt/model_executor/forward_batch_info.py

@@ -25,6 +25,8 @@ ScheduleBatch -> ModelWorkerBatch -> ForwardBatch
 - ScheduleBatch is managed by `scheduler.py::Scheduler`.
   It contains high-level scheduling data. Most of the data is on the CPU.
 - ModelWorkerBatch is managed by `tp_worker.py::TpModelWorker`.
+  It is a subset of `ScheduleBatch` that only contains data related to the model forward on GPU.
+  It will be transformed from CPU scheduler to GPU model runner.
 - ForwardBatch is managed by `model_runner.py::ModelRunner`.
   It contains low-level tensor data. Most of the data consists of GPU tensors.
 """

python/sglang/srt/model_executor/model_runner.py

@@ -131,6 +131,13 @@ class ModelRunner:
             ]:
                 server_args.disable_cuda_graph = True
 
+        if self.server_args.enable_overlap_schedule:
+            logger.warning(
+                "Overlap scheduler is enabled. This is an experimental feature. "
+                "Sampling penalizer (e.g., frequency and repetition penalty), constrained decoding (e.g., regex, JSON), "
+                "and embedding APIs are not supported and will lead to wrong results."
+            )
+
         # Global vars
         if server_args.show_time_cost:
             enable_show_time_cost()