Fix logging (#796)

python/sglang/lang/backend/openai.py

@@ -18,7 +18,7 @@ except ImportError as e:
     openai = tiktoken = e
 
 
-logger = logging.getLogger("openai")
+logger = logging.getLogger(__name__)
 
 
 def create_logit_bias_int(tokenizer):

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

@@ -15,6 +15,7 @@ limitations under the License.
 
 """Meta data for requests and batches"""
 
+import logging
 import warnings
 from dataclasses import dataclass
 from enum import IntEnum, auto
@@ -40,6 +41,9 @@ global_server_args_dict = {
 }
 
 
+logger = logging.getLogger(__name__)
+
+
 class ForwardMode(IntEnum):
     # Prefill a new sequence. This is deprecated now. "EXTEND" covers this case.
     PREFILL = auto()
@@ -379,7 +383,7 @@ class Batch:
             out_cache_loc = self.token_to_kv_pool.alloc(extend_num_tokens)
 
             if out_cache_loc is None:
-                print("Prefill out of memory. This should never happen.")
+                logger.error("Prefill out of memory. This should never happen.")
                 self.tree_cache.pretty_print()
                 exit()
 
@@ -613,7 +617,7 @@ class Batch:
         self.out_cache_loc = self.token_to_kv_pool.alloc(bs)
 
         if self.out_cache_loc is None:
-            print("Decode out of memory. This should never happen.")
+            logger.error("Decode out of memory. This should never happen.")
             self.tree_cache.pretty_print()
             exit()
 

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

@@ -39,7 +39,7 @@ from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.utils import kill_parent_process
 from sglang.utils import get_exception_traceback
 
-logger = logging.getLogger("srt.controller")
+logger = logging.getLogger(__name__)
 
 
 class LoadBalanceMethod(Enum):

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

@@ -31,7 +31,7 @@ from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.utils import kill_parent_process
 from sglang.utils import get_exception_traceback
 
-logger = logging.getLogger("srt.controller")
+logger = logging.getLogger(__name__)
 
 
 class ControllerSingle:

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

@@ -57,7 +57,7 @@ from sglang.srt.utils import (
     monkey_patch_vllm_qvk_linear_loader,
 )
 
-logger = logging.getLogger("srt.model_runner")
+logger = logging.getLogger(__name__)
 
 
 class ModelRunner:
@@ -90,7 +90,7 @@ class ModelRunner:
 
         # Init torch distributed
         torch.cuda.set_device(self.gpu_id)
-        logger.info(f"[gpu_id={self.gpu_id}] Init nccl begin.")
+        logger.info(f"[gpu={self.gpu_id}] Init nccl begin.")
 
         if not server_args.enable_p2p_check:
             monkey_patch_vllm_p2p_access_check(self.gpu_id)
@@ -130,7 +130,7 @@ class ModelRunner:
 
     def load_model(self):
         logger.info(
-            f"[gpu_id={self.gpu_id}] Load weight begin. "
+            f"[gpu={self.gpu_id}] Load weight begin. "
             f"avail mem={get_available_gpu_memory(self.gpu_id):.2f} GB"
         )
 
@@ -178,7 +178,7 @@ class ModelRunner:
             cache_config=None,
         )
         logger.info(
-            f"[gpu_id={self.gpu_id}] Load weight end. "
+            f"[gpu={self.gpu_id}] Load weight end. "
             f"type={type(self.model).__name__}, "
             f"dtype={self.dtype}, "
             f"avail mem={get_available_gpu_memory(self.gpu_id):.2f} GB"
@@ -229,7 +229,7 @@ class ModelRunner:
             layer_num=self.model_config.num_hidden_layers,
         )
         logger.info(
-            f"[gpu_id={self.gpu_id}] Memory pool end. "
+            f"[gpu={self.gpu_id}] Memory pool end. "
             f"avail mem={get_available_gpu_memory(self.gpu_id):.2f} GB"
         )
 
@@ -280,7 +280,7 @@ class ModelRunner:
             return
 
         logger.info(
-            f"[gpu_id={self.gpu_id}] Capture cuda graph begin. This can take up to several minutes."
+            f"[gpu={self.gpu_id}] Capture cuda graph begin. This can take up to several minutes."
         )
         batch_size_list = [1, 2, 4] + [i * 8 for i in range(1, 17)]
         self.cuda_graph_runner = CudaGraphRunner(

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

@@ -55,7 +55,7 @@ from sglang.srt.utils import (
 )
 from sglang.utils import get_exception_traceback
 
-logger = logging.getLogger("srt.tp_worker")
+logger = logging.getLogger(__name__)
 
 
 class ModelTpServer:
@@ -132,7 +132,7 @@ class ModelTpServer:
 
         # Print info
         logger.info(
-            f"[gpu_id={self.gpu_id}] "
+            f"[gpu={self.gpu_id}] "
             f"max_total_num_tokens={self.max_total_num_tokens}, "
             f"max_prefill_tokens={self.max_prefill_tokens}, "
             f"max_running_requests={self.max_running_requests}, "
@@ -256,7 +256,7 @@ class ModelTpServer:
         self.num_generated_tokens = 0
         self.last_stats_tic = time.time()
         logger.info(
-            f"[gpu_id={self.gpu_id}] Decode batch. "
+            f"[gpu={self.gpu_id}] Decode batch. "
             f"#running-req: {len(self.running_batch.reqs)}, "
             f"#token: {num_used}, "
             f"token usage: {num_used / self.max_total_num_tokens:.2f}, "
@@ -434,7 +434,7 @@ class ModelTpServer:
                 self.tree_cache_metrics["hit"] / self.tree_cache_metrics["total"]
             )
             logger.info(
-                f"[gpu_id={self.gpu_id}] Prefill batch. "
+                f"[gpu={self.gpu_id}] Prefill batch. "
                 f"#new-seq: {len(can_run_list)}, "
                 f"#new-token: {new_batch_input_tokens}, "
                 f"#cached-token: {hit_tokens}, "

python/sglang/srt/model_loader/utils.py

@@ -38,7 +38,7 @@ from vllm.model_executor.layers.quantization.base_config import QuantizationConf
 
 from sglang.srt.layers.quantization import get_quantization_config
 
-logger = logging.getLogger("srt.model_loader")
+logger = logging.getLogger(__name__)
 temp_dir = tempfile.gettempdir()