Checkpoint-engine is a simple middleware to update model weights in LLM inference engines -- a critical step in reinforcement learning. We provide an efficient and lightweight implementation for inplace weight update: updating our Kimi-K2 model (1 Trillion parameters) across thousands of GPUs takes about 20s.

The core weight update logic is in ParameterServer class, a service colocated with inference engines. It provides two implementations of weight update: Broadcast and P2P.

• Broadcast: Used when a large number of inference instances need to update weights in synchronous. This is the fastest implementation and should be used as the default update method. See _update_per_bucket with ranks == None or [].
• P2P: Used when new inference instances are dynamically added (due to restarts or dynamic availability) while the existing instances are already serving requests. Under this scenario, to avoid affecting the workloads on existing instances, we use the mooncake-transfer-engine to P2P send weights from CPUs in existing instances to GPUs in new instances. See _update_per_bucket with ranks specified.

In the Broadcast implementation, the checkpoint-engine holds references to sharded weights in CPU memory, and need to efficiently broadcast them to a cluster of inference instances, often under a different sharding pattern. We arrange the data transfer into 3 stages:

1. H2D: moving weights to GPU memory. These weights may come from disk or the training engine.
2. broadcast: broadcast among checkpoint engine workers; the data results in a CUDA IPC buffer shared with inference engine.
3. reload: inference engine decides what subset of weights to copy from the broadcasted data.

Checkpoint-engine orchestrates the entire transfer process. It first gathers necessary metadata to create a plan, including deciding the proper bucket size for data transfer. It then executes the transfer, where it controls the inference engine through a ZeroMQ socket. To maximize performance, it organizes the data transfers into a pipeline with overlapped communication and copy, illustrated below. The details can be found in Kimi-K2 Technical Report.

Pipelining naturally requires more GPU memory. When memory is not enough, checkpoint-engine will fallback to serial execution.

In the P2P implementation, checkpoint-engine needs to send weights from existing instances to new instances. To minimize the overall transfer time, checkpoint-engine optimizes the bucket assignment for each sender-receiver pair. The optimization goal is to make full use of the available network bandwidth for each sender and receiver. See issue #25

All results above are tested by examples/update.py and use vLLM v0.10.2rc1 as inference engine. Some notes:

• FP8 test needs additional vLLM patches, see FP8 quantization.
• Device Info: we tested various combination of devices and parallelism setups. For example, a 256-GPU TP16 setup means that we deploy 16 vLLM instances, each with 16-way tensor parallelism.
• Since update duration is related to IPC bucket size, we provide the bucket size in the table.
• The P2P time were tested for updating no more than two nodes (16 GPUs) (ParameterServer.update(ranks=range(0, 16))) out of the entire cluster.
• We bind each GPU to its corresponding NUMA node to ensure stable H2D transfer speeds.

Use the fastest broadcast implementation

Use the flexible P2P implementation, notice this will install mooncake-transfer-engine to support RDMA transfer between different ranks.

Prepare an H800 or H20 machine with 8 GPUs with vLLM. Be sure to include /collective_rpc API endpoint commit (available in main branch) since checkpoint-engine will use this endpoint to update weights. vLLM version v0.10.2 is fully tested and recommended.

Install checkpoint-engine

We use Qwen/Qwen3-235B-A22B-Instruct-2507 (BF16) as the test model

Start vLLM in dev mode and set --load-format dummy. Notice that we also set --worker-extension-cls=checkpoint_engine.worker.VllmColocateWorkerExtension

Meanwhile, use this command to update weights by checkpoint-engine. No need to wait for vLLM to get ready.

New checkpoint-engine instances can join existing instances and reuse their weights. This is simple to achieve.

First, start the existing instances with --save-metas-file global_metas.pkl to save global metas to a file and use --sleep-time 300 to make sure they stay alive.

After a checkpoint is registered, new instances can obtain a copy of the checkpoint by setting --load-metas-file global_metas.pkl.

FP8 quantization currently do not natively work in vLLM when updating weights. We provide a simple patch in patches/vllm_fp8.patch to handle the correct weight update. Notice this patch is only tested in DeepSeek-V3.1 and Kimi-K2. Other models may meet some compatible issues.

A PR is opened to the vLLM project and waiting to discuss and review.

Run a simple correctness test for checkpoint_engine

test_update.py are only designed to run with pytest. Please don't run it directly with torchrun.

Other unit tests can also be done with pytest. Only test_update.py requires GPUs, other tests can be run on CPUs. Only to run CPU tests, use:

• PS_MAX_BUCKET_SIZE_GB: An integer is used to set the maximum bucket size for checkpoint-engine. If not set, 8GB is used as default.
• PS_P2P_STORE_RDMA_DEVICES: Comma-separated RDMA devices' names for P2P transfer. If not set, checkpoint-engine will fall back to use NCCL_IB_HCA to detect RDMA devices.
• NCCL_IB_HCA: Available patterns can be found from NCCL documentation. If also not set, all RDMA devices will be used and divided evenly among the ranks.

Checkpoint Engine provides efficient distributed checkpoint loading for SGLang inference servers, significantly reducing model loading time for large models and multi-node setups.

1. Install checkpoint-engine:

2. Launch SGLang server:

3. Run checkpoint engine:

For 2-node setup, run the same commands on both nodes with appropriate --host and distributed training parameters.

SGLang Server:

• --wait-for-initial-weights: Wait for checkpoint engine before becoming ready
• --load-format dummy: Enable overlapping initialization tasks

Checkpoint Engine:

• --update-method: Choose broadcast, p2p, or all
• --inference-parallel-size: Number of parallel processes
• --checkpoint-path: Model checkpoint directory

• This project is currently tested with vLLM and SGLang. Integration with other frameworks is planned for future releases.
• The perfect three-stage pipeline mentioned in our paper is currently not implemented. This could be useful for architectures where H2D and broadcast do not conflict in PCIE.

This open source project uses the same vLLM interface in vllm-project/vllm#24295 . Thanks for the comments and insights from youkaichao.