From 8fd6e9e01789bd458dd420bab8284f92af305698 Mon Sep 17 00:00:00 2001
From: ziang663 <119752791+ziang663@users.noreply.github.com>
Date: Fri, 26 Jun 2026 12:49:23 +0800
Subject: [PATCH] Fix fixed-size HiCache capacity under PP (#29258)

---
 .../sglang/srt/mem_cache/memory_pool_host.py  |  5 +-
 python/sglang/srt/mem_cache/pool_host/base.py | 46 ++++++++++++++++++-
 2 files changed, 49 insertions(+), 2 deletions(-)

diff --git a/python/sglang/srt/mem_cache/memory_pool_host.py b/python/sglang/srt/mem_cache/memory_pool_host.py
index cbbd07b3b..5f0618cc5 100644
--- a/python/sglang/srt/mem_cache/memory_pool_host.py
+++ b/python/sglang/srt/mem_cache/memory_pool_host.py
@@ -73,6 +73,7 @@ logger = logging.getLogger(__name__)
 from sglang.srt.mem_cache.pool_host import HostKVCache
 from sglang.srt.mem_cache.pool_host.base import (
     HICACHE_HOST_MEMORY_RESERVE_BYTES,
+    sync_fixed_hicache_size,
     synchronized,
 )
 from sglang.srt.mem_cache.pool_host.common import (
@@ -1435,7 +1436,9 @@ class MambaPoolHost(HostKVCache):
         self.size_per_token = self.get_size_per_token()
 
         if host_size > 0:
-            self.size = int(host_size * 1e9 // self.size_per_token)
+            self.size = sync_fixed_hicache_size(
+                int(host_size * 1e9 // self.size_per_token), host_size
+            )
         else:
             self.size = int(device_pool.size * host_to_device_ratio)
 
diff --git a/python/sglang/srt/mem_cache/pool_host/base.py b/python/sglang/srt/mem_cache/pool_host/base.py
index 0329fc7ca..0ca962225 100644
--- a/python/sglang/srt/mem_cache/pool_host/base.py
+++ b/python/sglang/srt/mem_cache/pool_host/base.py
@@ -25,6 +25,48 @@ _is_hip = is_hip()
 HICACHE_HOST_MEMORY_RESERVE_BYTES: int = 10 * (1024**3)
 
 
+def sync_fixed_hicache_size(size: int, host_size: int) -> int:
+    """Sync fixed-size HiCache token capacity across PP ranks.
+
+    A fixed --hicache-size is specified in GB, but each PP stage may have a
+    different bytes/token because it owns different layers. Use the global
+    minimum token capacity within the PP group so all stages expose the same
+    host-cache capacity.
+    Ratio-based sizing already derives from the synced device pool size.
+    """
+    if host_size <= 0 or not torch.distributed.is_available():
+        return size
+
+    if not torch.distributed.is_initialized():
+        return size
+
+    try:
+        from sglang.srt.distributed.parallel_state import get_pp_group
+
+        pp_group = get_pp_group()
+    except AssertionError:
+        return size
+
+    if pp_group.world_size <= 1:
+        return size
+
+    tensor = torch.tensor(size, dtype=torch.int64)
+    torch.distributed.all_reduce(
+        tensor,
+        op=torch.distributed.ReduceOp.MIN,
+        group=pp_group.cpu_group,
+    )
+    synced_size = int(tensor.item())
+
+    if synced_size != size:
+        logger.info(
+            "Sync fixed-size HiCache host token capacity from %d to %d.",
+            size,
+            synced_size,
+        )
+    return synced_size
+
+
 def synchronized(func):
     @wraps(func)
     def wrapper(self, *args, **kwargs):
@@ -58,7 +100,9 @@ class HostKVCache(abc.ABC):
         self.dtype = device_pool.store_dtype
         self.size_per_token = self.get_size_per_token()
         if host_size > 0:
-            self.size = int(host_size * 1e9 // self.size_per_token)
+            self.size = sync_fixed_hicache_size(
+                int(host_size * 1e9 // self.size_per_token), host_size
+            )
         else:
             self.size = int(device_pool.size * host_to_device_ratio)
         # Align up the host memory pool size to the page size