From 9057db94175e58ce59a234c60aea167c436ca5c2 Mon Sep 17 00:00:00 2001
From: Raghavendra Vedula <vedula@fireworks.ai>
Date: Tue, 21 Jul 2026 15:28:23 -0700
Subject: [PATCH] Gate Mamba slot-donation debug asserts behind
 SGLANG_MAMBA_DEBUG_ASSERTS (#31982)

---
 .../sglang/srt/mem_cache/mamba_radix_cache.py | 23 +++++++++++++------
 python/sglang/srt/mem_cache/memory_pool.py    | 21 ++++++++++++-----
 2 files changed, 31 insertions(+), 13 deletions(-)

diff --git a/python/sglang/srt/mem_cache/mamba_radix_cache.py b/python/sglang/srt/mem_cache/mamba_radix_cache.py
index 921f5af2c..20905dd87 100644
--- a/python/sglang/srt/mem_cache/mamba_radix_cache.py
+++ b/python/sglang/srt/mem_cache/mamba_radix_cache.py
@@ -19,6 +19,7 @@ limitations under the License.
 The radix tree data structure for managing the hybrid (full and Mamba) KV cache.
 """
 
+import os
 from array import array
 from collections import defaultdict
 from typing import TYPE_CHECKING, List, Optional, Tuple
@@ -61,6 +62,12 @@ from sglang.srt.runtime_context import get_parallel
 
 logger = logging.getLogger(__name__)
 
+# Debug-only invariant checks in the Mamba slot-donation path call tensor.item(),
+# which forces a per-request cudaStreamSynchronize on the scheduler thread. Under
+# load this can serialize/stall the scheduler. Gate them off by default; set
+# SGLANG_MAMBA_DEBUG_ASSERTS=1 to re-enable for debugging.
+_MAMBA_DEBUG_ASSERTS = os.environ.get("SGLANG_MAMBA_DEBUG_ASSERTS", "0") == "1"
+
 
 class TreeNode:
 
@@ -592,13 +599,15 @@ class MambaRadixCache(KVCacheEventMixin, BasePrefixCache):
                 src_active = req.mamba_ping_pong_track_buffer[
                     mamba_ping_pong_track_buffer_to_keep
                 ].unsqueeze(-1)
-                assert src_active.item() != -1, (
-                    f"Cached mamba slot is -1: keep_idx={mamba_ping_pong_track_buffer_to_keep}, "
-                    f"buf={req.mamba_ping_pong_track_buffer.tolist()}, "
-                    f"next_track_idx={req.mamba_next_track_idx}, "
-                    f"last_track_seqlen={req.mamba_last_track_seqlen}, "
-                    f"rid={req.rid}"
-                )
+                if _MAMBA_DEBUG_ASSERTS:
+                    # .item() forces a cudaStreamSynchronize; only pay it when debugging.
+                    assert src_active.item() != -1, (
+                        f"Cached mamba slot is -1: keep_idx={mamba_ping_pong_track_buffer_to_keep}, "
+                        f"buf={req.mamba_ping_pong_track_buffer.tolist()}, "
+                        f"next_track_idx={req.mamba_next_track_idx}, "
+                        f"last_track_seqlen={req.mamba_last_track_seqlen}, "
+                        f"rid={req.rid}"
+                    )
                 if self.int8_ckpt_pool is not None:
                     mamba_value = self._commit_int8_checkpoint(src_active)
                     # quantized -> no ping-pong slot needs keeping
diff --git a/python/sglang/srt/mem_cache/memory_pool.py b/python/sglang/srt/mem_cache/memory_pool.py
index 5c951f811..0140f9daa 100644
--- a/python/sglang/srt/mem_cache/memory_pool.py
+++ b/python/sglang/srt/mem_cache/memory_pool.py
@@ -27,6 +27,7 @@ import copy
 import dataclasses
 import logging
 import math
+import os
 from contextlib import contextmanager, nullcontext
 from dataclasses import dataclass, fields
 from functools import cached_property
@@ -93,6 +94,12 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+# Debug-only invariant in the Mamba slot-donation path calls tensor.item(), which
+# forces a per-request cudaStreamSynchronize on the scheduler thread and can stall
+# the scheduler under load. Off by default; set SGLANG_MAMBA_DEBUG_ASSERTS=1 to
+# re-enable for debugging.
+_MAMBA_DEBUG_ASSERTS = os.environ.get("SGLANG_MAMBA_DEBUG_ASSERTS", "0") == "1"
+
 GB = 1024 * 1024 * 1024
 _is_cuda = is_cuda()
 _is_npu = is_npu()
@@ -1404,12 +1411,14 @@ class HybridReqToTokenPool(ReqToTokenPool):
         mamba_value_donated = (
             req.mamba_ping_pong_track_buffer[donate_idx].unsqueeze(-1).clone()
         )
-        assert mamba_value_donated.item() != -1, (
-            f"Donated mamba slot is -1: donate_idx={donate_idx}, "
-            f"buf={req.mamba_ping_pong_track_buffer.tolist()}, "
-            f"next_track_idx={req.mamba_next_track_idx}, "
-            f"rid={req.rid}"
-        )
+        if _MAMBA_DEBUG_ASSERTS:
+            # .item() forces a cudaStreamSynchronize; only pay it when debugging.
+            assert mamba_value_donated.item() != -1, (
+                f"Donated mamba slot is -1: donate_idx={donate_idx}, "
+                f"buf={req.mamba_ping_pong_track_buffer.tolist()}, "
+                f"next_track_idx={req.mamba_next_track_idx}, "
+                f"rid={req.rid}"
+            )
         self.set_mamba_ping_pong_slot(req, donate_idx, new_slot[0])
         return mamba_value_donated