[SOT] Extend SOT warmup support to new hardware (#3032)

* add new hardware

* add_sot_warmup4new_hardware

* fix conflict

* rm Optional

fastdeploy/worker/gcu_model_runner.py

@@ -26,6 +26,10 @@ from paddleformers.utils.log import logger
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request
 from fastdeploy.model_executor.forward_meta import ForwardMeta
+from fastdeploy.model_executor.graph_optimization.utils import (
+    profile_run_guard,
+    sot_warmup_guard,
+)
 from fastdeploy.model_executor.guided_decoding import get_guided_backend
 from fastdeploy.model_executor.guided_decoding.base_guided_decoding import (
     LogitsProcessorBase,
@@ -79,8 +83,10 @@ class GCUModelRunner(ModelRunnerBase):
             self.sampler = SpeculativeSampler(fd_config)
 
         # Cuda Graph
+        self.graph_opt_level = self.graph_opt_config.graph_opt_level
         self.use_cudagraph = self.graph_opt_config.use_cudagraph
         self.cudagraph_capture_sizes = list(reversed(self.graph_opt_config.cudagraph_capture_sizes))
+        self.sot_warmup_sizes = self.graph_opt_config.sot_warmup_sizes
 
         # Initialize share inputs
         self._init_share_inputs(self.parallel_config.max_num_seqs)
@@ -851,6 +857,17 @@ class GCUModelRunner(ModelRunnerBase):
         time_after_capture = time.perf_counter()
         logger.info(f"Cuda Graph capturing took {time_after_capture - time_before_capture} seconds")
 
+    @sot_warmup_guard(True)
+    def sot_warmup(self) -> None:
+        start_time = time.perf_counter()
+        for batch_size in self.sot_warmup_sizes:
+            self._dummy_run(
+                num_tokens=self.parallel_config.max_num_batched_tokens,
+                batch_size=batch_size,
+            )
+            logger.info(f"SOT warmup the model with the batch size:{batch_size}")
+        logger.info(f"SOT warmup took {time.perf_counter() - start_time} seconds")
+
     def _get_skip_idx(self, model_forward_batch: Optional[List[Request]] = None):
         """
         Get the index of the request that needs to be skipped during execution.
@@ -1041,6 +1058,7 @@ class GCUModelRunner(ModelRunnerBase):
         else:
             raise ValueError(f"{type(self.model)} has no attribute 'empty_input_forward")
 
+    @profile_run_guard(True)
     def profile_run(self) -> None:
         """Execute a forward pass with dummy inputs to profile the memory usage of the model"""