Warmup cublas (#566)

benchmark/gsm8k/README.md

 ## Download data
 ```
-wget https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl
+bash download_data.sh
 ```
 
 ## Run benchmark

benchmark/gsm8k/download_data.sh

+wget https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/train.jsonl
+wget https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl
\ No newline at end of file

benchmark/mmlu/README.md

 ## Download data
 ```
-wget https://people.eecs.berkeley.edu/~hendrycks/data.tar
-tar xf data.tar
+bash download_data.sh
 ```
 
 ## Run benchmark

benchmark/mmlu/download_data.sh

+wget https://people.eecs.berkeley.edu/~hendrycks/data.tar
+tar xf data.tar
\ No newline at end of file

python/sglang/srt/managers/controller/model_runner.py

@@ -270,6 +270,7 @@ class ModelRunner:
         # Load the model and create memory pool
         self.load_model()
         self.init_memory_pool(total_gpu_memory)
+        self.init_cublas()
         self.init_flash_infer()
 
     def load_model(self):
@@ -346,6 +347,15 @@ class ModelRunner:
             f"avail mem={get_available_gpu_memory(self.gpu_id):.2f} GB"
         )
 
+    def init_cublas(self):
+        """We need to run a small matmul to init cublas. Otherwise, it will raise some errors later."""
+        dtype = torch.float16
+        device = "cuda"
+        a = torch.ones((16, 16), dtype=dtype, device=device)
+        b = torch.ones((16, 16), dtype=dtype, device=device)
+        c = a @ b
+        return c
+
     def init_flash_infer(self):
         if global_server_args_dict.get("enable_flashinfer", False):
             from flashinfer import (

python/sglang/srt/managers/controller/tp_worker.py

@@ -410,7 +410,7 @@ class ModelTpServer:
                 self.tree_cache_metrics["hit"] / self.tree_cache_metrics["total"]
             )
             logger.info(
-                f"[gpu_id={self.gpu_id}] Prefil batch. "
+                f"[gpu_id={self.gpu_id}] Prefill batch. "
                 f"#new-seq: {len(can_run_list)}, "
                 f"#new-token: {new_batch_input_tokens}, "
                 f"#cached-token: {hit_tokens}, "