[minor] Make DEFAULT_FORCE_STREAM_INTERVAL configurable via SGLANG_FORCE_STREAM_INTERVAL (#23215)

python/sglang/srt/environ.py

@@ -266,6 +266,12 @@ class Envs:
     SGLANG_REQ_RUNNING_TIMEOUT = EnvFloat(-1)  # in seconds
     SGLANG_DISAGGREGATION_BOOTSTRAP_ENTRY_CLEANUP_INTERVAL = EnvInt(120)
     SGLANG_SWA_EVICTION_INTERVAL_MULTIPLIER = EnvFloat(1.0)
+    # For non-streaming requests, the scheduler still flushes intermediate
+    # output batches to the tokenizer manager every N decoded tokens so that
+    # `first_token_time`/TTFT can be recorded. Lower this (e.g. to 1) to get
+    # an accurate TTFT for benchmarking; the upstream default of 50 trades
+    # off some TTFT-metric accuracy for less IPC overhead.
+    SGLANG_FORCE_STREAM_INTERVAL = EnvInt(50)
 
     # Test: pd-disaggregation
     SGLANG_TEST_PD_DISAGG_BACKEND = EnvStr("mooncake")

python/sglang/srt/managers/scheduler_output_processor_mixin.py

@@ -33,7 +33,9 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
-DEFAULT_FORCE_STREAM_INTERVAL = 50
+# How often (in decoded tokens) the scheduler force-flushes an intermediate
+# output batch for non-streaming requests.
+DEFAULT_FORCE_STREAM_INTERVAL = envs.SGLANG_FORCE_STREAM_INTERVAL.get()
 
 
 class SchedulerOutputProcessorMixin: