From 9271fc4575ef77db2bc65e6002a6f5f2607a1841 Mon Sep 17 00:00:00 2001
From: foraxe <1055696449@qq.com>
Date: Sun, 26 Apr 2026 11:10:01 +0800
Subject: [PATCH 1/2] Avoid DSv4 compressed mixed multi-prefill crash

---
 python/sglang/srt/managers/scheduler.py | 25 +++++++++++++++++++++++--
 1 file changed, 23 insertions(+), 2 deletions(-)

diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index 7fb06598feb4..ddc9a622f1fc 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -33,7 +33,7 @@
 from torch.cuda import StreamContext as CudaStreamContext
 from torch.distributed import barrier
 
-from sglang.srt.configs.model_config import ModelConfig
+from sglang.srt.configs.model_config import ModelConfig, is_deepseek_compressed
 from sglang.srt.constrained.grammar_manager import GrammarManager
 from sglang.srt.disaggregation.decode import (
     DecodePreallocQueue,
@@ -329,6 +329,7 @@ def __init__(
 
         # Init model configs
         self.init_model_config()
+        self._dsv4_compressed_prefill_guard_warned = False
 
         # Init metrics stats
         self.init_metrics(tp_rank, pp_rank, dp_rank)
@@ -1950,6 +1951,24 @@ def get_num_allocatable_reqs(self, running_bs):
         res = min(res, self.req_to_token_pool.available_size())
         return res
 
+    def _get_effective_prefill_max_requests(self, running_bs: int) -> Optional[int]:
+        prefill_max_requests = self.server_args.prefill_max_requests
+        if (
+            running_bs > 0
+            and is_deepseek_compressed(self.model_config.hf_config)
+            and self.server_args.get_attention_backends() == ("compressed", "compressed")
+            and (prefill_max_requests is None or prefill_max_requests > 1)
+        ):
+            if not self._dsv4_compressed_prefill_guard_warned:
+                logger.warning(
+                    "DSv4 compressed prefill guard active: running_bs=%s, "
+                    "effective_prefill_max_requests=1",
+                    running_bs,
+                )
+                self._dsv4_compressed_prefill_guard_warned = True
+            return 1
+        return prefill_max_requests
+
     def get_new_batch_prefill(self) -> Optional[ScheduleBatch]:
         prefill_delayer_single_pass = None
         if self.prefill_delayer:
@@ -2019,6 +2038,8 @@ def _get_new_batch_prefill_raw(
             if dynamic_size is not None:
                 chunked_prefill_size = dynamic_size
 
+        prefill_max_requests = self._get_effective_prefill_max_requests(running_bs)
+
         # Prefill policy
         adder = PrefillAdder(
             self.page_size,
@@ -2030,7 +2051,7 @@ def _get_new_batch_prefill_raw(
             chunked_prefill_size,
             running_bs if self.is_mixed_chunk else 0,
             self.priority_scheduling_preemption_threshold,
-            prefill_max_requests=self.server_args.prefill_max_requests,
+            prefill_max_requests=prefill_max_requests,
             prefill_delayer_single_pass=prefill_delayer_single_pass,
             dllm_config=self.dllm_config,
         )

From 27467df7f0892e45da6be8ae7129e470440b40dc Mon Sep 17 00:00:00 2001
From: foraxe <1055696449@qq.com>
Date: Sun, 26 Apr 2026 11:21:29 +0800
Subject: [PATCH 2/2] Cache DSv4 compressed prefill guard predicate

---
 python/sglang/srt/managers/scheduler.py | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index ddc9a622f1fc..a942ce0ba577 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -330,6 +330,10 @@ def __init__(
         # Init model configs
         self.init_model_config()
         self._dsv4_compressed_prefill_guard_warned = False
+        self._is_dsv4_compressed_with_compressed_backend = (
+            is_deepseek_compressed(self.model_config.hf_config)
+            and self.server_args.get_attention_backends() == ("compressed", "compressed")
+        )
 
         # Init metrics stats
         self.init_metrics(tp_rank, pp_rank, dp_rank)
@@ -1955,8 +1959,7 @@ def _get_effective_prefill_max_requests(self, running_bs: int) -> Optional[int]:
         prefill_max_requests = self.server_args.prefill_max_requests
         if (
             running_bs > 0
-            and is_deepseek_compressed(self.model_config.hf_config)
-            and self.server_args.get_attention_backends() == ("compressed", "compressed")
+            and self._is_dsv4_compressed_with_compressed_backend
             and (prefill_max_requests is None or prefill_max_requests > 1)
         ):
             if not self._dsv4_compressed_prefill_guard_warned: