From 0eff0f74608fef93a19a7ab058f1cba5e94cab57 Mon Sep 17 00:00:00 2001
From: cctry <csycfl@gmail.com>
Date: Fri, 4 Sep 2026 15:05:57 -0700
Subject: [PATCH] Add num_prealloc_ready_tokens to decode load snapshot
 (#38065)

---
 python/sglang/srt/managers/load_snapshot.py            |  1 +
 .../srt/managers/scheduler_components/load_inquirer.py | 10 +++++++---
 2 files changed, 8 insertions(+), 3 deletions(-)

diff --git a/python/sglang/srt/managers/load_snapshot.py b/python/sglang/srt/managers/load_snapshot.py
index 26e306e4d..968f438f8 100644
--- a/python/sglang/srt/managers/load_snapshot.py
+++ b/python/sglang/srt/managers/load_snapshot.py
@@ -202,6 +202,7 @@ class LoadSnapshot(msgspec.Struct, omit_defaults=True):
     # num_total_tokens minus tokens still awaiting a KV transfer (equal to it
     # outside disaggregated decode).
     num_active_tokens: int = 0
+    num_prealloc_ready_tokens: int = 0
     max_total_num_tokens: int = 0
     max_running_requests: int = 0
     token_usage: float = 0.0
diff --git a/python/sglang/srt/managers/scheduler_components/load_inquirer.py b/python/sglang/srt/managers/scheduler_components/load_inquirer.py
index e470a40ae..bb1f5d47a 100644
--- a/python/sglang/srt/managers/scheduler_components/load_inquirer.py
+++ b/python/sglang/srt/managers/scheduler_components/load_inquirer.py
@@ -173,6 +173,7 @@ class SchedulerLoadInquirer:
         prefill_bootstrap = prefill_inflight = 0
         decode_prealloc = decode_transfer = decode_retracted = 0
         decode_prealloc_ready = 0
+        num_prealloc_ready_tokens = 0
         if self.disaggregation_mode == DisaggregationMode.PREFILL:
             mode_str = "prefill"
             prefill_bootstrap = len(self.get_disagg_prefill_bootstrap_queue().queue)
@@ -184,11 +185,13 @@ class SchedulerLoadInquirer:
             decode_retracted = len(
                 self.get_disagg_decode_prealloc_queue().retracted_queue
             )
-            decode_prealloc_ready = sum(
-                1
+            ready_reqs = [
+                decode_req.req
                 for decode_req in self.get_disagg_decode_prealloc_queue().queue
                 if decode_req.waiting_for_input
-            )
+            ]
+            decode_prealloc_ready = len(ready_reqs)
+            num_prealloc_ready_tokens = sum(req.seqlen for req in ready_reqs)
         disaggregation = DisaggregationMetrics(
             mode=mode_str,
             prefill_bootstrap_queue_reqs=prefill_bootstrap,
@@ -220,6 +223,7 @@ class SchedulerLoadInquirer:
             num_used_tokens=num_used_tokens,
             num_total_tokens=num_total_tokens,
             num_active_tokens=num_active_tokens,
+            num_prealloc_ready_tokens=num_prealloc_ready_tokens,
             max_total_num_tokens=self.max_total_num_tokens,
             max_running_requests=self.max_running_requests,
             token_usage=round(kv_token_usage, 4),