[Feat]Add scheduler recv skipper weights to environment configuration (#13855)

docs/references/environment_variables.md

@@ -31,6 +31,11 @@ SGLang supports various environment variables that can be used to configure its
 | `SGLANG_FUSED_MLA_ENABLE_ROPE_FUSION` | Enable RoPE fusion in Fused Multi-Layer Attention | `1` |
 | `SGLANG_DISABLE_CONSECUTIVE_PREFILL_OVERLAP` | Disable overlap schedule for consecutive prefill batches | `false` |
 | `SGLANG_DISABLE_FA4_WARMUP` | Disable Flash Attention 4 warmup passes (set to `1`, `true`, `yes`, or `on` to disable) | `false` |
+| `SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_DEFAULT` | Default weight value for scheduler recv skipper counter (used when forward mode doesn't match specific modes). Only active when `--scheduler-recv-interval > 1`. The counter accumulates weights and triggers request polling when reaching the interval threshold. | `1000` |
+| `SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_DECODE` | Weight increment for decode forward mode in scheduler recv skipper. Works with `--scheduler-recv-interval` to control polling frequency during decode phase. | `1` |
+| `SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_VERIFY` | Weight increment for target verify forward mode in scheduler recv skipper. Works with `--scheduler-recv-interval` to control polling frequency during verification phase. | `1` |
+| `SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_NONE` | Weight increment when forward mode is None in scheduler recv skipper. Works with `--scheduler-recv-interval` to control polling frequency when no specific forward mode is active. | `1` |
+
 
 ## DeepGEMM Configuration (Advanced Optimization)
 

python/sglang/srt/environ.py

@@ -168,6 +168,13 @@ class Envs:
     SGLANG_RETRACT_DECODE_STEPS = EnvInt(20)
     SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION = EnvInt(4096)
 
+    # Scheduler: recv interval
+    SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_DEFAULT = EnvInt(1000)
+    SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_DECODE = EnvInt(1)
+    SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_TARGET_VERIFY = EnvInt(1)
+    SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_NONE = EnvInt(1)
+
+
     # Scheduler: others:
     SGLANG_EMPTY_CACHE_INTERVAL = EnvFloat(-1)  # in seconds. Set if you observe high memory accumulation over a long serving period.
     SGLANG_DISABLE_CONSECUTIVE_PREFILL_OVERLAP = EnvBool(False)

python/sglang/srt/managers/scheduler_recv_skipper.py

@@ -1,3 +1,4 @@
+from sglang.srt.environ import envs
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
 from sglang.srt.server_args import ServerArgs
 
@@ -14,11 +15,20 @@ class SchedulerRecvSkipper:
         assert not server_args.enable_dp_attention
         self._counter = 0
         self._threshold = server_args.scheduler_recv_interval
+        # All can be tuned if needed
+        self._default_weight = envs.SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_DEFAULT.get()
+        self._weight_of_forward_mode = {
+            ForwardMode.DECODE: envs.SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_DECODE.get(),
+            ForwardMode.TARGET_VERIFY: envs.SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_TARGET_VERIFY.get(),
+            None: envs.SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_NONE.get(),
+        }
 
     def handle(self, last_forward_mode: ForwardMode):
         should_recv = False
 
-        last_weight = _WEIGHT_OF_FORWARD_MODE.get(last_forward_mode, _DEFAULT_WEIGHT)
+        last_weight = self._weight_of_forward_mode.get(
+            last_forward_mode, self._default_weight
+        )
         self._counter += last_weight
 
         if self._counter >= self._threshold:
@@ -26,12 +36,3 @@ class SchedulerRecvSkipper:
             should_recv = True
 
         return should_recv
-
-
-# All can be tuned if needed
-_DEFAULT_WEIGHT = 1000
-_WEIGHT_OF_FORWARD_MODE = {
-    ForwardMode.DECODE: 1,
-    ForwardMode.TARGET_VERIFY: 1,
-    None: 1,
-}