fix typing.

Author: wanghao7

lightllm/common/basemodel/triton_kernel/add_in_place.py

@@ -21,17 +21,18 @@ def _add_in_place(
     x = x + y * alpha
     tl.store(input_ptr + offsets, x, mask=mask)
 
+
 @torch.no_grad()
 def add_in_place(input: torch.Tensor, other: torch.Tensor, *, alpha=1):
     assert input.is_contiguous(), "input tensor must be contiguous"
     assert other.is_contiguous(), "other tensor must be contiguous"
     n_elements = input.numel()
-    grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']), )
+    grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),)
     _add_in_place[grid](
         input,
         other,
         n_elements,
         alpha,
         BLOCK_SIZE=1024,
     )
-    return input
+    return input

lightllm/server/router/manager.py

@@ -108,8 +108,6 @@ def __init__(self, args, router_port, detokenization_port, metric_port):
         self.schedule_task = None
         self.overlap_event = threading.Event()
         return
-    
-
 
     async def wait_to_model_ready(self):
         # 初始化模型
@@ -322,11 +320,11 @@ async def _step(self):
                 self.running_batch = new_batch
                 await self._prefill_batch(self.running_batch)
                 self._filter_runing_batch()
-                
+
                 # 激进调度控制
                 if not self.args.disable_aggressive_schedule:
                     self.has_wait_tokens = self.max_wait_tokens
-                
+
             elif self.is_multinode_and_multidp:
                 # 在多节点多 dp 的模式下，如果当前 running_batch 为None, 也需要不断的调用 decode 操作，
                 # 因为其他节点上的dp可能存在运行的请求，所以本节点也需要调用decode，推理后端的backend会
@@ -341,11 +339,11 @@ async def _step(self):
             new_mini_batch = await self.get_schedule_result(self.running_batch)
             self.has_wait_tokens = 0
             if new_mini_batch is not None:
-                
+
                 # 激进调度控制
                 if not self.args.disable_aggressive_schedule:
                     self.has_wait_tokens = self.max_wait_tokens
-                
+
                 self.stats_tool.count_prompt_tokens(new_mini_batch)
                 await self._prefill_batch(new_mini_batch)
                 if not new_mini_batch.is_clear():

lightllm/utils/log_utils.py

@@ -14,6 +14,7 @@
 _LOG_LEVEL = getattr(logging, _LOG_LEVEL.upper(), 0)
 _LOG_DIR = os.environ.get("LIGHTLLM_LOG_DIR", None)
 
+
 class NewLineFormatter(logging.Formatter):
     """Adds logging prefix to newlines to align multi-line messages."""
 
@@ -45,14 +46,14 @@ def _setup_logger():
         _default_handler.flush = sys.stdout.flush  # type: ignore
         _default_handler.setLevel(_LOG_LEVEL)
         _root_logger.addHandler(_default_handler)
-    
+
     if _default_file_handler is None and _LOG_DIR is not None:
         if not os.path.exists(_LOG_DIR):
             try:
                 os.makedirs(_LOG_DIR)
             except OSError as e:
                 _root_logger.warn(f"Error creating directory {_LOG_DIR} : {e}")
-        _default_file_handler = logging.FileHandler(_LOG_DIR + '/default.log')
+        _default_file_handler = logging.FileHandler(_LOG_DIR + "/default.log")
         _default_file_handler.setLevel(_LOG_LEVEL)
         _default_file_handler.setFormatter(fmt)
         _root_logger.addHandler(_default_file_handler)
@@ -62,6 +63,7 @@ def _setup_logger():
     # being propagated to the parent logger.
     _root_logger.propagate = False
 
+
 # The logger is initialized when the module is imported.
 # This is thread-safe as the module is only imported once,
 # guaranteed by the Python GIL.
@@ -96,7 +98,8 @@ def init_logger(name: str):
 
 _log_time_mark_dict = {}
 
-def log_time_ready(mark_name, time_count:int):
+
+def log_time_ready(mark_name, time_count: int):
     """
     time_count 间隔时间超过多少s调用该函数会返回True，否则返回False
     用于控制一些日志输出的频率
@@ -112,5 +115,3 @@ def log_time_ready(mark_name, time_count:int):
         return True
     else:
         return False
-    
-

unit_tests/common/basemodel/triton_kernel/test_add_in_place.py

@@ -24,7 +24,8 @@ def test_add_in_place(dim1, dim2, alpha):
     output = input + other * alpha
     add_in_place(input, other, alpha=alpha)
     rlt = torch.allclose(input, output, atol=1e-5, rtol=0)
-    assert(rlt)
+    assert rlt
+
 
 if __name__ == "__main__":
     pytest.main()