[log] Make forward iter count optional (#12116)

python/sglang/srt/environ.py

@@ -111,18 +111,21 @@ class Envs:
     # Model & File Download
     SGLANG_USE_MODELSCOPE = EnvBool(False)
 
+    # Logging Options
+    SGLANG_LOG_GC = EnvBool(False)
+    SGLANG_LOG_FORWARD_ITERS = EnvBool(False)
+    SGLANG_DISABLE_REQUEST_LOGGING = EnvBool(False)
+
     # Test & Debug
     SGLANG_IS_IN_CI = EnvBool(False)
     SGLANG_IS_IN_CI_AMD = EnvBool(False)
     SGLANG_SET_CPU_AFFINITY = EnvBool(False)
     SGLANG_PROFILE_WITH_STACK = EnvBool(True)
     SGLANG_RECORD_STEP_TIME = EnvBool(False)
-    SGLANG_GC_LOG = EnvBool(False)
     SGLANG_FORCE_SHUTDOWN = EnvBool(False)
     SGLANG_DEBUG_MEMORY_POOL = EnvBool(False)
     SGLANG_TEST_REQUEST_TIME_STATS = EnvBool(False)
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK = EnvBool(False)
-    SGLANG_DISABLE_REQUEST_LOGGING = EnvBool(False)
     SGLANG_SIMULATE_ACC_LEN = EnvFloat(-1)
     SGLANG_SIMULATE_ACC_METHOD = EnvStr("multinomial")
     SGLANG_TORCH_PROFILER_DIR = EnvStr("/tmp")
@@ -251,7 +254,17 @@ class Envs:
 envs = Envs()
 
 
+def _print_deprecated_env(new_name: str, old_name: str):
+    if old_name in os.environ:
+        warnings.warn(
+            f"Environment variable {old_name} will be deprecated, please use {new_name} instead"
+        )
+        os.environ[new_name] = os.environ[old_name]
+
+
 def _convert_SGL_to_SGLANG():
+    _print_deprecated_env("SGLANG_LOG_GC", "SGLANG_GC_LOG")
+
     for key, value in os.environ.items():
         if key.startswith("SGL_"):
             new_key = key.replace("SGL_", "SGLANG_", 1)

python/sglang/srt/managers/scheduler.py

@@ -494,7 +494,7 @@ class Scheduler(
         )
         self.init_disaggregation()
 
-        if get_bool_env_var("SGLANG_GC_LOG"):
+        if envs.SGLANG_LOG_GC.get():
             configure_gc_logger()
 
         # Init prefill kv split size when deterministic inference is enabled with various attention backends

python/sglang/srt/managers/scheduler_metrics_mixin.py

@@ -7,6 +7,7 @@ from typing import TYPE_CHECKING, List, Optional
 
 from sglang.srt.disaggregation.kv_events import EventPublisherFactory, KVEventBatch
 from sglang.srt.disaggregation.utils import DisaggregationMode
+from sglang.srt.environ import envs
 from sglang.srt.managers.schedule_policy import PrefillAdder
 from sglang.srt.managers.scheduler import Req, ScheduleBatch
 from sglang.srt.metrics.collector import SchedulerMetricsCollector, SchedulerStats
@@ -18,6 +19,7 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 RECORD_STEP_TIME = get_bool_env_var("SGLANG_RECORD_STEP_TIME")
+LOG_FORWARD_ITERS = envs.SGLANG_LOG_FORWARD_ITERS.get()
 
 
 class KvMetrics:
@@ -125,8 +127,10 @@ class SchedulerMetricsMixin:
             num_used, token_usage, _, _ = self._get_token_info()
             token_usage_msg = f"token usage: {token_usage:.2f}, "
 
+        iter_msg = f" [{self.forward_ct + 1}]" if LOG_FORWARD_ITERS else ""
+
         f = (
-            f"Prefill batch. "
+            f"Prefill batch{iter_msg}, "
             f"#new-seq: {len(can_run_list)}, "
             f"#new-token: {adder.log_input_tokens}, "
             f"#cached-token: {adder.log_hit_tokens}, "
@@ -249,7 +253,8 @@ class SchedulerMetricsMixin:
                 gap_latency / self.server_args.decode_log_interval
             )
 
-        msg = f"Decode batch. #running-req: {num_running_reqs}, {token_usage_msg}"
+        iter_msg = f" [{self.forward_ct}]" if LOG_FORWARD_ITERS else ""
+        msg = f"Decode batch{iter_msg}, #running-req: {num_running_reqs}, {token_usage_msg}"
 
         if self.spec_algorithm.is_none():
             spec_accept_length = 0