diff --git a/python/sglang/srt/managers/cache_controller.py b/python/sglang/srt/managers/cache_controller.py
index 87e5edbc6..c3b5df40b 100644
--- a/python/sglang/srt/managers/cache_controller.py
+++ b/python/sglang/srt/managers/cache_controller.py
@@ -23,7 +23,7 @@ from typing import TYPE_CHECKING, Dict, List, NamedTuple, Optional, Set
 import torch
 
 from sglang.srt.environ import envs
-from sglang.srt.mem_cache.cp_hicache_trace import cptrace, rng, trace_enabled
+from sglang.srt.mem_cache.cp_hicache_trace import cptrace, khash, rng, trace_enabled
 from sglang.srt.mem_cache.hicache_storage import HiCacheStorageConfig
 
 if TYPE_CHECKING:
@@ -1489,6 +1489,23 @@ class HiCacheController:
                     state.draft_host_indices, self.write_stream
                 )
 
+        # Value fingerprint of the EXACT device KV being backed up, taken on the
+        # DEFAULT stream (outside the write_stream block) so it reflects the
+        # correct post-store value -- NOT the racy write_stream copy's view. If a
+        # per-layer store/copy ordering race exists, the host copy will diverge
+        # from THIS hash, and the reload hash (post-H2D) will not match it.
+        if trace_enabled(2):
+            for state in target_transfer_states:
+                kv = self.mem_pool_device.get_key_buffer(layer_id)
+                cptrace(
+                    2,
+                    "backup_kv_hash",
+                    node_id=state.reservation.node_id,
+                    layer=layer_id,
+                    kvhash=khash(kv[state.physical_device_indices]),
+                    nrows=int(state.physical_device_indices.numel()),
+                )
+
         for state in final_states:
             self._append_layer_write_ack(state)
 
@@ -1951,6 +1968,21 @@ class HiCacheController:
                                 self.io_backend,
                             )
                             submitted_target = True
+                            # Value fingerprint of what actually landed on device
+                            # after H2D (on load_stream, in-order after the load).
+                            # Compare to backup_kv_hash(node,layer): inequality =
+                            # round-trip delivered wrong KV. node_ids is the merged
+                            # op's nodes (single node in the c=1 rehit pass).
+                            if trace_enabled(2):
+                                kv = self.mem_pool_device.get_key_buffer(i)
+                                cptrace(
+                                    2,
+                                    "reload_kv_hash",
+                                    node_ids=node_ids,
+                                    layer=i,
+                                    kvhash=khash(kv[device_indices]),
+                                    nrows=int(device_indices.numel()),
+                                )
                         producer_event.complete(i)
                     elif op is None and i < self.layer_num:
                         producer_event.complete(i)
diff --git a/python/sglang/srt/mem_cache/cp_hicache_trace.py b/python/sglang/srt/mem_cache/cp_hicache_trace.py
index 21338b07b..e24af4e92 100644
--- a/python/sglang/srt/mem_cache/cp_hicache_trace.py
+++ b/python/sglang/srt/mem_cache/cp_hicache_trace.py
@@ -57,6 +57,28 @@ def _fmt(v) -> str:
     return str(v).replace(" ", "")
 
 
+def khash(t) -> int:
+    """Full-tensor, position-weighted int64 fingerprint of a KV slice.
+
+    Sensitive to BOTH value and position (catches permutation/scatter, unlike a
+    plain sum), covers the whole tensor (not a byte sample), and reduces to one
+    int. Debug-only: the .item() sync is acceptable when tracing is on. Used to
+    compare the KV bytes a node's pages hold at backup (correct, post-store, on
+    the default stream) vs what comes back at reload (post-H2D); inequality =
+    the round-trip delivered wrong KV (store-vs-copy race or transfer corruption)."""
+    try:
+        import torch
+
+        if t is None or t.numel() == 0:
+            return 0
+        b = t.detach().contiguous().view(torch.uint8).reshape(-1).to(torch.int64)
+        n = b.numel()
+        idx = torch.arange(1, n + 1, device=b.device, dtype=torch.int64)
+        return int((((b * idx).sum() ^ (b.sum() * 1000003) ^ n) & 0x7FFFFFFFFFFFFFFF).item())
+    except Exception:
+        return -1
+
+
 def cptrace(level: int, tag: str, **fields) -> None:
     """Emit one ``[CPTRACE <tag>] k=v ...`` line if the gate >= level.