Add Agentic-Aware Tail-Optimized LRU eviction to the unified radix cache (#34012)

Co-authored-by: Shuwen Wang <47200617+alphabetc1@users.noreply.github.com>

docs/docs/advanced_features/radix_eviction_policy.mdx

@@ -24,6 +24,7 @@ Select one with `--radix-eviction-policy`. All of them fall back to least-recent
 | `lfu` | The prefix with the fewest cache hits, then the least recent. | A small set of prompts is reused far more than the rest, and you want those to survive bursts of one-off traffic. |
 | `slru` | Prefixes still in the probationary segment, then the least recent within a segment. | Like `lfu`, but you want a hard floor on how much a one-off prefix can displace a proven one. See [`slru` parameters](#slru-parameters). |
 | `priority` | The prefix belonging to the lowest-priority request, then the least recent. | You run [priority scheduling](/docs/advanced_features/server_arguments) and want cache retention to follow the same ranking as admission. |
+| `tlru` | The tail of a conversation beyond what its next prefill needs to meet the TTFT budget ("TEL-safe" tokens), then the least recent. | Agentic / multi-turn workloads where tail TTFT matters more than raw hit rate. See [`tlru` parameters](#tlru-parameters). Requires the unified radix cache (the default tree). |
 
 Notes on the scoring inputs:
 
@@ -55,7 +56,7 @@ TypeError: SLRUStrategy.__init__() got an unexpected keyword argument 'protected
 
 ### Policy parameters
 
-Only `slru` currently takes a parameter. `lru`, `lfu`, and `priority` take none, so `--radix-eviction-policy-config` has no effect with them and any key is an error.
+`slru` and `tlru` take parameters. `lru`, `lfu`, and `priority` take none, so `--radix-eviction-policy-config` has no effect with them and any key is an error.
 
 #### `slru` parameters
 
@@ -67,6 +68,24 @@ Only `slru` currently takes a parameter. `lru`, `lfu`, and `priority` take none,
 
 Raising it makes promotion harder, so the protected set stays small and closer to your genuinely hot prefixes; a prefix hit three times stays probationary at `4` but is protected at the default `2`. Lowering it to `1` promotes any prefix that is reused even once, which approaches `lru` with a one-hit grace period.
 
+#### `tlru` parameters
+
+`tlru` (Tail-Optimized LRU, [arXiv:2510.15152](https://arxiv.org/abs/2510.15152)) protects only the cached history a conversation's next prefill needs to stay under a TTFT budget; the rest of its tail is evicted first, and eviction continues in plain recency order once the TEL-safe tokens run out.
+
+| Key | Type | Meaning |
+|---|---|---|
+| `threshold` | int | Tail-latency threshold ξ, in tokens: a conversation only keeps enough cache to hold its next prefill under ξ uncached tokens. Convert from a TTFT target by dividing it by the measured ms per uncached token. The paper states ξ in blocks, so multiply its values by `--page-size`. |
+| `next_prompt_estimate` | int | Estimated tokens the next turn of a conversation will add; use the trace's empirical mean. |
+
+Both keys are required in practice: `threshold` must be greater than `next_prompt_estimate` (only the difference affects behaviour), and values at or above it reduce T-LRU to plain LRU, which startup rejects.
+
+```bash Command
+python3 -m sglang.launch_server \
+  --model-path MODEL_PATH \
+  --radix-eviction-policy tlru \
+  --radix-eviction-policy-config '{"threshold": 4096, "next_prompt_estimate": 512}'
+```
+
 ## Which policy to pick
 
 Start with `lru` and change it only against a measured cache hit rate — the counters are exposed under `--enable-metrics`. `lfu` and `slru` help when your traffic has a stable hot set that a recency-only policy keeps flushing; they hurt when prefix popularity shifts over time, because a prefix that earned a high hit count keeps its advantage after it stops being useful.

python/sglang/srt/arg_groups/choices.py

@@ -187,7 +187,7 @@ FP4_GEMM_RUNNER_BACKEND_CHOICES = [
     "marlin",
 ]
 
-RADIX_EVICTION_POLICY_CHOICES = ["lru", "lfu", "slru", "priority"]
+RADIX_EVICTION_POLICY_CHOICES = ["lru", "lfu", "slru", "priority", "tlru"]
 
 RL_ON_POLICY_TARGET_CHOICES = ["fsdp"]
 

python/sglang/srt/arg_groups/fields/memory.py

@@ -35,8 +35,11 @@ class Memory(msgspec.Struct):
             help=(
                 "The eviction policy of radix trees. 'lru' stands for Least "
                 "Recently Used, 'lfu' stands for Least Frequently Used, 'slru' "
-                "stands for Segmented Least Recently Used, and 'priority' evicts "
-                "lower-priority requests first. See "
+                "stands for Segmented Least Recently Used, 'priority' evicts "
+                "lower-priority requests first, and 'tlru' stands for "
+                "Tail-Optimized LRU (arXiv:2510.15152), which evicts the part of "
+                "a conversation that cannot affect tail TTFT before falling back "
+                "to LRU. See "
                 "https://docs.sglang.io/docs/advanced_features/radix_eviction_policy "
                 "for what each policy optimizes for."
             ),
@@ -52,10 +55,13 @@ class Memory(msgspec.Struct):
         Arg(
             help=(
                 "Tuning parameters for --radix-eviction-policy, as a json object "
-                "passed to the policy as keyword arguments. Only 'slru' takes any "
-                "today: protected_threshold (int, default 2), e.g. "
-                "'{\"protected_threshold\": 4}'. An unrecognized key fails at "
-                "startup, naming the key and the policy. See "
+                "passed to the policy as keyword arguments. 'slru' takes "
+                "protected_threshold (int, default 2), e.g. "
+                "'{\"protected_threshold\": 4}'; 'tlru' takes threshold and "
+                "next_prompt_estimate (ints, tokens), e.g. "
+                '\'{"threshold": 4096, "next_prompt_estimate": 512}\'. An '
+                "unrecognized key fails at startup, naming the key and the "
+                "policy. See "
                 "https://docs.sglang.io/docs/advanced_features/radix_eviction_policy#policy-parameters "
                 "for the full parameter list."
             ),

python/sglang/srt/arg_groups/kv_cache_hook.py

@@ -163,6 +163,24 @@ def handle_cache_compatibility(server_args: Any) -> None:
             "--disable-priority-preemption when priority scheduling is enabled."
         )
 
+    if cfg.radix_eviction_policy == "tlru":
+        tlru_config = cfg.radix_eviction_policy_config or {}
+        threshold = tlru_config.get("threshold", 0)
+        next_prompt_estimate = tlru_config.get("next_prompt_estimate", 0)
+        if threshold < 0 or next_prompt_estimate < 0:
+            raise ValueError(
+                "--radix-eviction-policy tlru requires non-negative 'threshold' and "
+                "'next_prompt_estimate' in --radix-eviction-policy-config, got "
+                f"{threshold} and {next_prompt_estimate}."
+            )
+        if threshold <= next_prompt_estimate:
+            raise ValueError(
+                "--radix-eviction-policy tlru needs 'threshold' greater than "
+                f"'next_prompt_estimate' in --radix-eviction-policy-config, got "
+                f"{threshold} <= {next_prompt_estimate}; otherwise no tokens are "
+                "ever TEL-safe and T-LRU is exactly LRU."
+            )
+
     if cfg.enable_hierarchical_cache and cfg.disable_radix_cache:
         raise ValueError(
             "The arguments enable-hierarchical-cache and disable-radix-cache are mutually exclusive "

python/sglang/srt/mem_cache/evict_policy.py

@@ -46,6 +46,41 @@ class PriorityStrategy(EvictionStrategy):
         return (node.priority, node.last_access_time)
 
 
+class TLRUStrategy(EvictionStrategy):
+    """Tail-Optimized LRU (Zhang et al., arXiv:2510.15152).
+
+    A conversation with history length L whose next prompt is expected to add
+    Q_hat tokens only has to keep L + Q_hat - threshold tokens cached to hold its
+    next prefill under the TTFT budget; tokens past that budget cannot improve
+    tail latency and are "TEL-safe", i.e. free to evict. Such nodes are reported
+    as infinitely old, which is the implementation the paper suggests: the
+    existing eviction driver then drains them before anything else (the paper's
+    phase 1) and continues in plain recency order once they run out (phase 2),
+    so neither eviction loop needs to know about T-LRU.
+
+    threshold and next_prompt_estimate are token counts, whereas the paper states
+    both in blocks; multiply the paper's values by page_size to convert.
+    """
+
+    def __init__(self, threshold: int = 0, next_prompt_estimate: int = 0):
+        self.threshold = threshold
+        self.next_prompt_estimate = next_prompt_estimate
+
+    def get_priority(self, node: TreeNode) -> Tuple[int, float]:
+        # node._tlru_history_len is the branch's high-water depth, i.e. the
+        # paper's L, and deliberately does not shrink when the tail is trimmed.
+        # Deriving L from what is still resident instead would leave the
+        # shortened conversation over budget on the next pass too, and T-LRU
+        # would walk it down to nothing rather than stopping after
+        # (threshold - Q_hat) tokens.
+        budget = max(
+            node._tlru_history_len + self.next_prompt_estimate - self.threshold, 0
+        )
+        cached_without_this_node = node._tlru_cached_prefix_len - len(node.key)
+        tel_safe = cached_without_this_node >= budget
+        return (-1 if tel_safe else 0, node.last_access_time)
+
+
 class SLRUStrategy(EvictionStrategy):
     def __init__(self, protected_threshold: int = 2):
         self.protected_threshold = protected_threshold

python/sglang/srt/mem_cache/registry.py

@@ -263,6 +263,18 @@ def create_tree_cache(ctx: TreeCacheBuildContext) -> BasePrefixCache:
             "option that selected another tree cache for this model."
         )
 
+    if get_memory().radix_eviction_policy == "tlru":
+        from sglang.srt.mem_cache.unified_radix_cache import UnifiedRadixCache
+
+        # T-LRU's per-node tail bookkeeping only exists on the unified tree;
+        # any other cache would silently fall back to LRU ordering.
+        if not isinstance(cache, UnifiedRadixCache):
+            raise ValueError(
+                "--radix-eviction-policy tlru requires UnifiedRadixCache, but "
+                f"tree_cache is {type(cache).__name__}. Drop the flag or the "
+                "option that selected another tree cache for this model."
+            )
+
     hicache_attached = cache.cache_controller is not None
     streaming_wrapped = False
     if (

python/sglang/srt/mem_cache/unified_cache/unified_tree_core.py

@@ -127,6 +127,10 @@ class UnifiedTreeNode:
         # Namespace-aware hashes used only for external KV events.
         self.event_hash_value: Optional[list[str]] = None
         self.hit_count = 0
+        # T-LRU only (0 under other policies): tokens root -> self, and the
+        # branch's high-water depth, which survives tail trimming.
+        self._tlru_cached_prefix_len = 0
+        self._tlru_history_len = 0
         self.external_cache_stored = False
         self.priority = priority
         self.lru_prev: list[UnifiedTreeNode | None] = [None] * (
@@ -182,6 +186,28 @@ class UnifiedTreeNode:
         return [value for chunk in reversed(chunks) for value in chunk]
 
 
+def _set_tlru_lens_and_raise_history(
+    node: UnifiedTreeNode, parent: UnifiedTreeNode
+) -> None:
+    """Record a new node's path depth and raise the branch high-water mark.
+
+    The walk stops early only at an ancestor whose subtree already reached this
+    depth (e.g. under a deeper sibling); a conversation that keeps setting a new
+    high-water mark walks its whole root path, so an insert costs O(path nodes)
+    in the worst case. The path length is bounded by the number of node segments
+    (roughly extend/split operations, not tokens), and the walk only runs under
+    --radix-eviction-policy tlru.
+    """
+    assert node.key is not None
+    node._tlru_cached_prefix_len = parent._tlru_cached_prefix_len + len(node.key)
+    if node._tlru_history_len < node._tlru_cached_prefix_len:
+        node._tlru_history_len = node._tlru_cached_prefix_len
+    cur = parent
+    while cur is not None and cur._tlru_history_len < node._tlru_cached_prefix_len:
+        cur._tlru_history_len = node._tlru_cached_prefix_len
+        cur = cur.parent
+
+
 class UnifiedLRUList:
     def __init__(
         self,
@@ -425,6 +451,9 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
         self.eviction_strategy = get_eviction_strategy(
             params.eviction_policy.lower(), params.eviction_policy_config
         )
+        # The node _tlru_* lens are read only by TLRUStrategy.get_priority;
+        # every other policy skips the per-insert bookkeeping entirely.
+        self.tlru_bookkeeping = params.eviction_policy.lower() == "tlru"
 
         # ``device`` is derived from the construction-time allocator; the
         # allocator/pool themselves are owned by the cache, not the tree.
@@ -1355,6 +1384,14 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
         new_node.hit_count = child.hit_count
         new_node.external_cache_stored = child.external_cache_stored
         new_node.creation_time = child.creation_time
+        if self.tlru_bookkeeping:
+            # A split adds no depth to the branch: the new parent sits at
+            # split_len tokens and inherits the branch's high-water mark, while
+            # child keeps its own depth because its path length is unchanged.
+            new_node._tlru_cached_prefix_len = (
+                new_node.parent._tlru_cached_prefix_len + split_len
+            )
+            new_node._tlru_history_len = child._tlru_history_len
         # Split fragments stay on the anchor's root path for the ack's walk.
         new_node.load_back_pending_id = child.load_back_pending_id
         # The rotation base is constant along a chain (position-page P keeps
@@ -1413,6 +1450,8 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
         new_node = self._new_node(priority=priority)
         new_node.parent = parent
         new_node.key = key
+        if self.tlru_bookkeeping:
+            _set_tlru_lens_and_raise_history(new_node, parent)
         # Chain-constant under sharding: the pre-flight decline in
         # begin_insert() guarantees this tail continues the matched prefix's
         # rotation, so stamping the inserting request's base keeps every node
@@ -2165,6 +2204,8 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
         new_node = self._new_node(priority=node.priority)
         new_node.parent = node
         new_node.key = key
+        if self.tlru_bookkeeping:
+            _set_tlru_lens_and_raise_history(new_node, node)
         new_node.hash_value = hash_value
         new_node.component_data[BASE_COMPONENT_TYPE].host_value = host_value.clone()
         node.children[child_key] = new_node

python/sglang/srt/mem_cache/utils.py

@@ -54,6 +54,7 @@ from sglang.srt.mem_cache.evict_policy import (
     MRUStrategy,
     PriorityStrategy,
     SLRUStrategy,
+    TLRUStrategy,
 )
 
 _EVICTION_POLICY_FACTORIES: dict[str, Callable[..., EvictionStrategy]] = {
@@ -64,6 +65,7 @@ _EVICTION_POLICY_FACTORIES: dict[str, Callable[..., EvictionStrategy]] = {
     "filo": FILOStrategy,
     "priority": PriorityStrategy,
     "slru": SLRUStrategy,
+    "tlru": TLRUStrategy,
 }
 
 

test/registered/unit/mem_cache/test_tlru_eviction_policy.py

@@ -0,0 +1,208 @@
+"""Unit tests for the Tail-Optimized LRU eviction strategy (arXiv:2510.15152).
+
+T-LRU reports a node as infinitely old once the conversation holding it is above
+its TEL-safe budget, so the ordinary eviction driver drains those nodes first
+(the paper's phase 1) and then continues in recency order (phase 2). These tests
+exercise that ordering against stub nodes, so they need no GPU, model or CUDA
+build and run in well under a second.
+
+evict_policy is loaded straight from its file: importing sglang as a package
+pulls in the engine's runtime dependencies, which would make a pure-logic test
+require a full install.
+"""
+
+import importlib.util
+import math
+import os
+from dataclasses import dataclass, field
+
+try:
+    from sglang.test.ci.ci_register import register_cpu_ci
+except ImportError:  # standalone run without an sglang install; CI parses the
+    # registration below from the AST, so the stub changes nothing for CI.
+    def register_cpu_ci(**kwargs):
+        pass
+
+
+register_cpu_ci(est_time=1, suite="base-a-test-cpu")
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_EVICT_POLICY = os.path.normpath(
+    os.path.join(_HERE, "../../../../python/sglang/srt/mem_cache/evict_policy.py")
+)
+_spec = importlib.util.spec_from_file_location("evict_policy", _EVICT_POLICY)
+_mod = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(_mod)
+LRUStrategy, TLRUStrategy = _mod.LRUStrategy, _mod.TLRUStrategy
+
+PAGE = 256
+XI = 4096
+Q_HAT = 1024
+DELTA = XI - Q_HAT  # tokens of tail the policy may free
+
+
+@dataclass
+class FakeKey:
+    n: int
+
+    def __len__(self):
+        return self.n
+
+
+@dataclass
+class FakeNode:
+    """A radix node on one conversation's path; only the fields T-LRU reads."""
+
+    _tlru_cached_prefix_len: int
+    key_len: int
+    _tlru_history_len: int
+    last_access_time: float = 0.0
+    key: FakeKey = field(init=False)
+
+    def __post_init__(self):
+        self.key = FakeKey(self.key_len)
+
+
+def chain(node_lens, convo_length=None, t0=0.0):
+    """Build a root->leaf chain, deepest last.
+
+    convo_length defaults to the full depth, which is the state right after the
+    conversation's newest turn was inserted.
+    """
+    convo = sum(node_lens) if convo_length is None else convo_length
+    nodes, depth = [], 0
+    for i, n in enumerate(node_lens):
+        depth += n
+        nodes.append(
+            FakeNode(
+                _tlru_cached_prefix_len=depth,
+                key_len=n,
+                _tlru_history_len=convo,
+                last_access_time=t0 + i,
+            )
+        )
+    return nodes
+
+
+def strategy(threshold=XI, next_prompt_estimate=Q_HAT):
+    return TLRUStrategy(threshold=threshold, next_prompt_estimate=next_prompt_estimate)
+
+
+def is_tel_safe(s, node):
+    return s.get_priority(node)[0] < 0
+
+
+def freed_tokens(s, nodes):
+    return sum(n.key_len for n in nodes if is_tel_safe(s, n))
+
+
+def test_fresh_conversation_frees_exactly_the_tail_budget():
+    s = strategy()
+    assert freed_tokens(s, chain([PAGE] * 40)) == DELTA
+
+
+def test_trimming_stops_after_the_budget():
+    """The survivors of a trim must be protected.
+
+    This is the regression guard for deriving the history length from what is
+    still resident: that would leave the shortened conversation over budget on
+    every subsequent pass and walk it down to nothing.
+    """
+    s = strategy()
+    nodes = chain([PAGE] * 40)
+    survivors = [n for n in nodes if not is_tel_safe(s, n)]
+    assert freed_tokens(s, survivors) == 0
+
+
+def test_conversation_under_threshold_is_entirely_free():
+    s = strategy()
+    short = chain([PAGE] * 4)  # 1024 + Q_hat <= xi, so no caching is needed
+    assert freed_tokens(s, short) == sum(n.key_len for n in short)
+
+
+def test_phase_one_spreads_across_conversations_then_falls_back_to_lru():
+    s = strategy()
+    old = chain([PAGE] * 40, t0=0.0)
+    new = chain([PAGE] * 40, t0=100.0)
+    order = sorted(old + new, key=s.get_priority)
+    n_safe = 2 * (DELTA // PAGE)
+
+    assert all(is_tel_safe(s, n) for n in order[:n_safe])
+    # Both conversations donate their tail, which is what the paper's
+    # per-conversation loop exists to produce.
+    assert any(n in old for n in order[:n_safe])
+    assert any(n in new for n in order[:n_safe])
+    # Phase 2 is plain recency.
+    assert order[n_safe].last_access_time == old[0].last_access_time
+    safe_times = [n.last_access_time for n in order[:n_safe]]
+    assert safe_times == sorted(safe_times)
+
+
+def test_degenerates_to_lru_when_estimate_reaches_threshold():
+    pool = chain([PAGE] * 40, t0=0.0) + chain([PAGE] * 40, t0=100.0)
+    degenerate = strategy(next_prompt_estimate=XI)
+    lru = LRUStrategy()
+    assert [id(n) for n in sorted(pool, key=degenerate.get_priority)] == [
+        id(n) for n in sorted(pool, key=lru.get_priority)
+    ]
+    assert freed_tokens(degenerate, chain([PAGE] * 40)) == 0
+
+
+def test_oversized_tail_node_is_protected_rather_than_partially_freed():
+    """Node granularity under-trims instead of over-trimming.
+
+    The paper trims one block at a time; a radix tree can only drop whole leaves,
+    so a turn larger than the budget stays put and phase 2 decides its fate.
+    """
+    s = strategy()
+    assert freed_tokens(s, chain([PAGE * 40])) == 0
+
+
+def test_compacted_branch_keeps_shared_prefix_protected():
+    """Context compaction shortens a conversation instead of extending it.
+
+    It occurs in about 2% of turns in the agentic traces we benchmark, and forks a
+    shallow branch off a shared ancestor that still carries the deeper branch's
+    high-water mark. The ancestor is then measured against a history longer than
+    what hangs below it, which must stay conservative: over-protect the shared
+    prefix, never free it early.
+    """
+    s = strategy()
+    assert freed_tokens(s, chain([PAGE] * 8, convo_length=200_000)) == 0
+
+
+def test_budget_clamps_at_zero():
+    """A budget below zero must mean nothing needs caching, not wrap around."""
+    s = strategy()
+    node = FakeNode(_tlru_cached_prefix_len=PAGE, key_len=PAGE, _tlru_history_len=0)
+    assert is_tel_safe(s, node)
+
+
+def test_priority_is_finite_and_orderable():
+    """The driver pushes (priority, node) onto a heap, so the keys must compare
+    without falling through to comparing nodes."""
+    s = strategy()
+    a, b = chain([PAGE] * 4)[:2]
+    for node in (a, b):
+        flag, when = s.get_priority(node)
+        assert flag in (-1, 0)
+        assert math.isfinite(when)
+    assert (s.get_priority(a) < s.get_priority(b)) or (
+        s.get_priority(b) < s.get_priority(a)
+    )
+
+
+if __name__ == "__main__":
+    failures = 0
+    for name, fn in sorted(globals().items()):
+        if name.startswith("test_") and callable(fn):
+            try:
+                fn()
+                print(f"OK   {name}")
+            except AssertionError as e:
+                failures += 1
+                print(f"FAIL {name}: {e}")
+    print(
+        "\n" + ("TLRU_TESTS_OK" if not failures else f"TLRU_TESTS_FAILED ({failures})")
+    )
+    raise SystemExit(1 if failures else 0)