[HiCache] buffer mode: anchor-lock staged prefetches by default (#37464)

python/sglang/srt/environ.py

@@ -723,9 +723,9 @@ class Envs:
     SGLANG_HICACHE_FILE_BACKEND_ENABLE_METADATA_CACHE = EnvBool(False)
     # Positive cache TTL for filesystem metadata lookups (-1 disables positive expiration)
     SGLANG_HICACHE_FILE_BACKEND_METADATA_TTL = EnvFloat(5.0)
-    # Buffer mode: pin a staged prefetch's device anchor from IO commit to
-    # consumption so eviction cannot waste the fetch; cap = fraction of pool.
-    SGLANG_ENABLE_HICACHE_BUFFER_ANCHOR_LOCK = EnvBool(False)
+    # Buffer mode: staged prefetches pin their device anchor from IO commit
+    # to consumption so eviction cannot waste the fetch. Cap = fraction of
+    # the pool the pins may hold; 0 disables pinning.
     SGLANG_HICACHE_BUFFER_ANCHOR_LOCK_CAP = EnvFloat(0.5)
     SGLANG_HICACHE_NIXL_BACKEND_STORAGE_DIR = EnvStr(None)
     # Enable O_DIRECT when opening NIXL POSIX backend files (bypasses OS page cache).
@@ -1765,6 +1765,10 @@ _DEPRECATED_ENVS: Dict[str, _DeprecatedEnv] = {
     "SGLANG_FLASHINFER_PR4266_SOURCE": _DeprecatedEnv(),
     # DSV4 compressor V2 is always used.
     "SGLANG_OPT_USE_COMPRESSOR_V2": _DeprecatedEnv(),
+    "SGLANG_ENABLE_HICACHE_BUFFER_ANCHOR_LOCK": _DeprecatedEnv(
+        note="Buffer-mode anchor pinning is always on; set "
+        "SGLANG_HICACHE_BUFFER_ANCHOR_LOCK_CAP=0 to disable it."
+    ),
     # Replaced by CLI flags.
     "SGLANG_ENABLE_GRPC": _DeprecatedEnv(
         note="Please use '--grpc-port' to enable the native gRPC server."

python/sglang/srt/mem_cache/buffer_mode/pipeline.py

@@ -222,8 +222,6 @@ class BufferModePipeline:
         # Metadata-only pending-write backlog cap; beyond it new intents
         # are dropped at admission (re-trigger on a later hit).
         self.write_backlog_cap = write_backlog_cap
-        # Anchor-lock knobs; the cap keeps queued holds from pinning the pool.
-        self.anchor_lock_enabled = envs.SGLANG_ENABLE_HICACHE_BUFFER_ANCHOR_LOCK.get()
         from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
 
         kvcache = cache.token_to_kv_pool_allocator.get_kvcache()
@@ -239,8 +237,7 @@ class BufferModePipeline:
             ),
         )
         logger.info(
-            "BufferModePipeline anchor_lock_enabled=%s cap_tokens=%d",
-            self.anchor_lock_enabled,
+            "BufferModePipeline anchor_lock_cap_tokens=%d",
             self.anchor_lock_cap_tokens,
         )
         self.reset()
@@ -690,8 +687,6 @@ class BufferModePipeline:
         O(prefix path)). Returns "locked", "no_anchor" (nothing to pin),
         "cap_skip" (over cap; launches unlocked), or "anchor_lost" (splice
         base gone — the caller cancels the storage IO)."""
-        if not self.anchor_lock_enabled:
-            return "no_anchor"
         if req_id in self.anchor_locks:
             return "locked"
         prefix_ctx = self._prefetch_prefix_ctx.get(req_id)