kv_cache_routing.md

<!--
SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
SPDX-License-Identifier: Apache-2.0
-->

# KV Cache Routing
This documentation explains how Key-Value (KV) cache routing works in Dynamo, providing optimized inference for large language models by intelligently directing requests to workers with the most relevant cached data while simultaneously load balancing based on utilization metrics sent by the workers.

To enable KV cache aware routing start the frontend node like this:
```
python -m dynamo.frontend --router-mode kv
```

The engine announces when a KV block is created or removed. The Dynamo router run finds the worker with the best match for those KV blocks and directs the traffic to that node.

For performance testing, compare a typical workload with `--router-mode random|round-robin` to see if it can benefit from KV-aware routing.

The KV-aware routing arguments:

- `--kv-overlap-score-weight`: Sets the amount of weighting on overlaps with prefix caches, which directly contributes to the prefill cost. A large weight is expected to yield a better TTFT (at the expense of worse ITL). When set to 0, prefix caches are not considered at all (falling back to pure load balancing behavior on the active blocks). Defaults to 1.

- `--router-temperature`: Sets the temperature when randomly selecting workers to route to via softmax sampling on the router cost logits. Setting it to 0 (default) recovers the deterministic behavior where the min logit is picked.

- `--use-kv-events`/`--no-kv-events`: Sets whether to listen to KV events for maintaining the global view of cached blocks. If true (default), then we use the `KvIndexer` to listen to the block creation and deletion events. If false, `ApproxKvIndexer`, which assumes the kv cache of historical prompts exists for fixed time durations (hard-coded to 120s), is used to predict the kv cache hit ratio in each engine. Set false if your backend engine does not emit KV events.

- `--router-replica-sync`: Enables state synchronization between multiple router replicas via NATS. Disabled by default, and can be enabled by passing the flag in. When enabled, router replicas share their view of KV cache distribution and active sequences, allowing all routers to make optimal routing decisions even when requests are distributed across multiple router instances. This improves fault tolerance and routing accuracy in multi-router deployments.

## Architecture

Colloquially, we refer to a Dynamo component that serves an endpoint for LLM inference as a **worker**.

## Basic Routing
Dynamo supports several routing strategies when sending requests from one component to another component's endpoint.

First, we must create a client tied to a components endpoint, we can do this using the labels defined above. Here we are getting a client tied to the `generate` endpoint of the `VllmWorker` component.

```python
client = namespace('dynamo').component('VllmWorker').endpoint('generate').client()
```

We can then use the default routing methods exposed by the client class to send requests to the `VllmWorker` component.

- **Random routing**: Default strategy, available via `client.generate()` or `client.random()`
- **Round-robin routing**: Cycles through available workers via `client.round_robin()`
- **Direct routing**: Explicitly targets a specific worker via `client.direct(input, component_id)`

KV Cache routing uses direct routing with a special worker selection algorithm.

## Serving Two Router Replicas

For improved fault tolerance, you can launch two frontend + router replicas. Since the frontend and router are currently tied together, you'll need to use two different HTTP ports for each instance.

To enable state sharing between the router replicas (which provides more accurate routing decisions), use the `--router-replica-sync` flag when starting the frontend:

```bash
# Router replica 1
python -m dynamo.frontend --router-mode kv --port 8000 --router-replica-sync

# Router replica 2
python -m dynamo.frontend --router-mode kv --port 8001 --router-replica-sync
```

When `--router-replica-sync` is enabled, the router replicas will communicate with each other via NATS to maintain consistent state across instances. This allows both routers to have a complete view of the KV cache distribution and make optimal routing decisions, even when requests are distributed across multiple router instances.

## Understanding KV Cache
The leading Large Language Models (LLMs) today are auto-regressive and based off of the [transformer architecture](https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf). One key inference optimization technique is to cache the already computed keys and values and to reuse them for the future tokens. This is called the [KV Cache](https://developer.nvidia.com/blog/mastering-llm-techniques-inference-optimization/#key-value_caching).

### KV Cache Optimizations
Every inference framework will have a KV Cache for each worker. A popular inference framework library is [vLLM](https://github.com/vllm-project/vllm) where a key contribution was [PagedAttention](https://arxiv.org/abs/2309.06180), which allowed them to manage KV Cache in an efficient way by chunking requests into blocks.

Another popular inference framework, [SGLang](https://github.com/sgl-project/sglang), contributed [RadixAttention](https://arxiv.org/abs/2312.07104) which introduced a
prefix tree which allows for efficient matching, inserting and eviction of KV Cache blocks. The prefix tree structure popularized KV Cache reuse.

In Dynamo, we introduce a KVPublisher which emits KV Cache events that occur at each worker and a KVIndexer which keeps track of these events globally.

To get a feel for how KV Cache management works on a single worker with KV Cache reuse turned on and where the KVPublisher gets plugged in, we can walk through the KV Block management flow:
1. Request tokenization: The incoming prompt is converted into tokens
2. Block partitioning: The token sequence is divided into fixed-size blocks (e.g., 16 or 64 tokens per block)
3. Block hashing: Each block of tokens is hashed to create a unique identifier
4. Cache lookup:
    - For each block, the system checks if a matching block already exists in the KV cache
    - If a match is found, the existing KV cache block is reused
    - If no match is found, the system proceeds to the next step
5. Resource allocation:
    - For blocks without matches, the system attempts to allocate new memory space
    - If sufficient memory is available, allocate memory space and proceed to step 7
    - If memory is constrained, proceed to step 6
6. Cache eviction (when necessary):
    - The system applies an eviction policy (e.g., LRU, LFU) to identify blocks for removal
    - Selected blocks are evicted from the cache
    - **KVPublisher emits a KV removed event notifying KVIndexer about the removed block.**
    - Alternatively, some systems may offload less-frequently used blocks to CPU memory.
7. KV computation:
    - For new blocks, the model computes key and value tensors
    - These tensors are stored in the newly allocated cache blocks
    - **KVPublisher emits a kv stored event notifying KVIndexer about newly stored blocks**.

Further details can be found for: [TRT-LLM](https://developer.nvidia.com/blog/introducing-new-kv-cache-reuse-optimizations-in-nvidia-tensorrt-llm/), [vLLM](https://docs.vllm.ai/en/latest/design/automatic_prefix_caching.html#design-automatic-prefix-caching) and [SGLang](https://lmsys.org/blog/2024-01-17-sglang/).

## KV Cache Routing and Load Balancing
```text
+---------+          +------------------+           +---------+
|  Tokens |--------->| KV Aware Router  |---------> | Worker 2|
+---------+          +------------------+           +---------+
                             |
          +------------------+------------------+
          |                  |                  |
          | Cached: 2 blocks | Cached: 5 blocks | Cached: 8 blocks
          | Prefill: 8 blks  | Prefill: 5 blks  | Prefill: 2 blks
          | Decode: 10 blks  | Decode: 7 blks   | Decode: 9 blks
          v                  v                  v
   +----------------+  +----------------+  +----------------+
   |   Worker 1     |  |   Worker 2     |  |   Worker 3     |
   +----------------+  +----------------+  +----------------+
```

Load balancing in LLM serving becomes complex when enabling KV Cache reuse. While KV Cache reuse can save significant computation, if the routing strategy is not aware of the unique KV states of each worker we can:
- miss opportunities for KV Cache reuse if routing to the "wrong" node
- get into an imbalanced state where a few workers are processing many requests, lowering throughput of entire system

The router uses a cost function that considers both the prefill cost (influenced by cached blocks) and the decode load to make optimal routing decisions:

### Cost Calculation

1. **Prefill blocks**: The number of tokens that need to be processed during prefill is predicted based on the request's input tokens and the cached blocks available on each worker. This is divided by the block size to get the effective "prefill blocks". This prediction is updated when the first output token is produced, signaling prefill completion.

2. **Decode blocks**: The number of blocks needed during the decode phase is predicted based on the request's input tokens and the current active sequences on each worker. This is updated when the request is freed (blocks are dereferenced or freed).

3. **Cost formula**: `cost = overlap_score_weight * prefill_blocks + decode_blocks`
   - Lower cost is better
   - The `overlap_score_weight` parameter controls the importance of cache hits vs. load balancing
   - A higher weight prioritizes cache reuse (better TTFT) while a lower weight prioritizes load distribution (better ITL)

### Worker Selection

The router selects the worker with the lowest cost. When `router_temperature` is set to a non-zero value, the router uses softmax sampling on the normalized cost logits to introduce randomness in the selection, which can help with load distribution.

Example calculation with `overlap_score_weight = 1.0`:
- Worker 1: cost = 1.0 * 8 + 10 = 18
- **Worker 2: cost = 1.0 * 5 + 7 = 12** (selected - lowest cost)
- Worker 3: cost = 1.0 * 2 + 9 = 11

## Events

In Dynamo, we support KV Cache Routing for many backends that have different implementations of KV Cache. To enable this, we built a KVPublisher that can be plugged into any framework to publish KV Events.

On the receiving side we have a KVIndexer which accepts events from the KVPublisher and puts them into a global prefix tree for tracking cached blocks across all workers.

```text
+----------------+                         +-----------------+
|                |                         | KV Aware Router |
|     Worker     |                         |                 |
|                | create_kv_block()       | +-------------+ |
| +------------+ | remove_kv_block()       | |  KVIndexer  | |
| |KVPublisher | |------------------------>| +-------------+ |
| +------------+ |                         |                 |
|                |                         |                 |
+----------------+                         +-----------------+
```

### KVPublisher
The KVPublisher can be initialized and then called in the inference framework where blocks are allocated and removed.

The two types of events are:
- KV stored event
- KV removed event

The publisher can be initialized and used through C bindings or Python bindings.

### KVIndexer
The KVIndexer builds and maintains a global view of cached blocks in a prefix tree. We modify the original prefix tree by also storing the worker id on each node. This is so we can return the number of matched blocks for each worker.

The KVIndexer has a method `find_matches_for_request`, which takes in tokens and returns a dictionary with keys of worker id and values of the number of matched KV Blocks.

### Inter-Router Communication

In multi-router deployments, each router only observes a subset of requests. To maintain a consistent global view of active sequences and KV cache states, routers broadcast their local actions to other replicas through three synchronization events:

1. **AddRequest**: Published when assigning a request to a worker, containing the request ID, worker ID, token sequence blocks, and overlap score. This updates other routers' tracking of which blocks are in use.

2. **MarkPrefillCompleted**: Published when a request transitions from prefill to decode phase, signaling that prefill tokens should no longer count toward the worker's active prefill load.

3. **Free**: Published when a request completes and its resources are released, allowing other routers to update their block reference counts.

Each event includes a unique router ID to prevent processing of self-generated events. This asynchronous communication ensures all routers maintain synchronized KV cache state for optimal routing decisions despite handling different request streams.