The right way to Replace LLM Weights with No Downtime

Think about attempting to renovate the inspiration of a towering skyscraper with out asking its occupants to depart or pause their work. That’s precisely what MoonshotAI’s Checkpoint Engine does for AI fashions. It permits large language fashions to replace their brains, the weights, whereas nonetheless operating, so there’s no downtime. This breakthrough lets builders enhance their AI shortly and effectively, even on fashions with over a trillion parameters operating on hundreds of GPUs. It’s quick, dependable, and designed to maintain AI techniques operating easily whereas evolving in real-time, making it an important software for cutting-edge AI functions. This text goes over what it’s, the way it works, and why it issues for the way forward for large-scale AI techniques.

What’s Moonshot AI’s Checkpoint engine?

Moonshot AI’s Checkpoint Engine is a specialised middleware designed to replace the weights of giant language fashions (LLMs) in real-time throughout inference with out interrupting ongoing operations. This functionality is important in Reinforcement studying eventualities the place mannequin weights must be up to date incessantly. The Checkpoint Engine at present integrates seamlessly with vLLM inference frameworks and affords optimized efficiency via pipelining and reminiscence administration methods. It additionally gives options like reusing weights from present situations to scale back overhead in scaling eventualities.

Structure 

The core of the Checkpoint is the ParameterServer class, which handles the burden replace logic and orchestrates the info movement.

1. H2D(Host to Machine): Strikes up to date weights from CPU reminiscence or storage to GPU reminiscence, utilizing optimized switch pipelines.
2. Broadcast: Distributes the burden throughout all inference engine situations effectively, leveraging CUDA IPC buffers for shared reminiscence communication.
3. Reload: Every inference engine then selectively reloads related weight shards from the broadcasted knowledge in keeping with its sharding sample.

These three-stage pipelines guarantee environment friendly, overlapping communication and copying for pace.

When GPU reminiscence is restricted, the system can fall again to serial execution to take care of reliability.

Strategies Used

The Checkpoint Engine makes use of two primary strategies to replace mannequin weights throughout inference.

1. Broadcast Methodology: That is the quickest and the default strategy. That is ultimate when a lot of inference situations must be up to date concurrently. It broadcasts the up to date weights from CPU reminiscence to all inference GPUs synchronously, making certain all situations keep completely in sync with minimal delay. 
2. P2P (Peer-to-Peer) Methodology: It’s used when inference situations are added or eliminated dynamically throughout runtime. It avoids disrupting present inference workloads by sending weights instantly from CPUs in present situations to GPUs in new situations via a peer-to-peer switch system, permitting easy and versatile updates.

Working 

The Checkpoint Engine orchestrates all the switch course of. It first gathers mandatory metadata to create a plan, together with deciding the correct bucket dimension for knowledge switch. Then, it executes the switch, controlling the inference engine via a ZeroMQ socket to maximise efficiency. It organizes knowledge switch into pipelines with overlapped communication and duplicate, enabling quick and environment friendly weight updates even underneath heavy workload.

By implementing the above-mentioned strategies and structure, the Checkpoint Engine permits reside weight updates for LLMs throughout hundreds of GPUs with minimal latency and repair disruption.

Set up and Utilization

Set up

To make use of the quickest broadcast 

Use Code:

pip set up checkpoint-engine

To make use of the versatile P2P implementation:

Use Code:

pip set up 'checkpoint-engine[p2p]'

It will set up mooncake-transfer-engine to help RDMA switch between completely different ranks.

Instance Use case

Step 1:

Put together an H800 or H20 machine with 8 GPUs with the most recent vLLM. You should definitely embrace /collective_rpc API endpoint commit (out there in the principle department) since checkpoint-engine will use this endpoint to replace weights.

Step 2:

set up checkpoint-engine

Code:

uv pip set up 'checkpoint-engine[p2p]'

Step 3:

For our use case, we’re gonna use Qwen/Qwen3-235B-A22B-Instruct-2507 because the take a look at mannequin.

Code:

hf obtain Qwen/Qwen3-235B-A22B-Instruct-2507 --local-dir /choose/fashions/Qwen/Qwen3-235B-A22B-Instruct-2507/

Step 4:

Begin vLLM in dev mode and set –load-format dummy. Make certain to set –worker-extension-cls=checkpoint_engine.employee.VllmColocateWorkerExtension

Code:

VLLM_SERVER_DEV_MODE=1 python3 -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 19730 --trust-remote-code 

    --tensor-parallel-size=8 --max-model-len 4096 --load-format dummy 

    --served-model-name checkpoint-engine-demo --model /choose/fashions/Qwen/Qwen3-235B-A22B-Instruct-2507/ 

    --worker-extension-cls checkpoint_engine.employee.VllmColocateWorkerExtension

To replace weights by checkpoint-engine. No want to attend for vLLM to prepare. Use the code under.

Code:

torchrun --nproc-per-node 8 examples/replace.py --update-method all --checkpoint-path /choose/fashions/Qwen/Qwen3-235B-A22B-Instruct-2507/

To reuse weights from present situations

New checkpoint-engine situations can be a part of present situations and reuse their weights.

Utilizing the tactic under:

Step 1: Begin the prevailing occasion with –save-metas-file global_metas.pkl to avoid wasting international metas to a file.

Step 2: Use –sleep-time 300 to ensure they keep alive.

Code:

torchrun --nproc-per-node 8 examples/replace.py --checkpoint-path $MODEL_PATH 

    --sleep-time 300 --save-metas-file global_metas.pkl

Step 3: After a checkpoint is registered, new situations can receive a replica of the checkpoint by setting –load-metas-file global_metas.pkl

Code:

torchrun --nproc-per-node 8 examples/replace.py --load-metas-file global_metas.pkl

FP8 quantization

At present, FP8 quantization doesn’t work in vLLM when updating weights. It makes use of a easy patch in patches/vllm_fp8.patch to deal with the proper weight replace. Additionally ,this patch is simply examined in DeepSeek-V3.1 and Kimi-K2. So there are possibilities of having some compatibility points with different fashions.

Take a look at

Run a easy correctness take a look at for checkpoint_engine

Code:

torchrun --nproc-per-node 8 checks/test_update.py

Benchmark

Insights

Listed below are a number of observations that I’ve made:

1. The printed methodology typically affords the quickest replace time, optimized for synchronous weight updates throughout many inference situations.
2. The P2P methodology takes longer however permits dynamic updates when situations be a part of or depart throughout runtime.
3. These benchmark exhibits the scalability of Checkpoint Engine, dealing with a trillion parameter fashions effectively on clusters starting from 8 to 256 GPUs

Limitations of Checkpoint Engine

Whereas Checkpoint Engine is a strong answer for reside weight updates in LLMs, it at present has some limitations.

• Works Greatest with vLLM for Now: The engine is especially examined with the vLLM framework. For those who’re hoping to make use of it with different AI frameworks or customized setups, you may want some further work to get it operating easily.
• Pipeline Nonetheless Bettering: The perfect seamless pipeline that overlaps knowledge strikes completely isn’t absolutely completed but. This implies there’s nonetheless potential to make the updates even sooner.
• P2P Replace Might Be Smoother: The peer-to-peer methodology sends knowledge via a bottleneck at one primary node earlier than sharing it with others, which may gradual issues down when you have got numerous GPUs.
• Wants Further GPU Reminiscence: The intelligent broadcast system makes use of extra GPU reminiscence to hurry issues up. On machines with much less reminiscence, it switches to a slower, much less environment friendly course of.
• Restricted Help for FP8 Fashions: For those who’re working with the newer FP8 quantized fashions, you’ll want some experimental patches. And even then, not all fashions play properly, but past a few examined ones.

Conclusion

Moonshot AI’s Checkpoint Engine is a game-changer for updating enormous AI fashions with out stopping them. It retains every thing operating easily, even whereas the mannequin’s “mind” is getting smarter in real-time. Whereas it nonetheless has a number of areas to enhance, the potential is large. For those who’re working with giant AI techniques, this software is unquestionably price watching. It’s serving to make the way forward for AI sooner and extra environment friendly, with none downtime.

Continuously Requested Questions

I’m a Knowledge Science Trainee at Analytics Vidhya, passionately engaged on the event of superior AI options reminiscent of Generative AI functions, Giant Language Fashions, and cutting-edge AI instruments that push the boundaries of expertise. My position additionally entails creating participating instructional content material for Analytics Vidhya’s YouTube channels, creating complete programs that cowl the complete spectrum of machine studying to generative AI, and authoring technical blogs that join foundational ideas with the most recent improvements in AI. By means of this, I purpose to contribute to constructing clever techniques and share information that conjures up and empowers the AI group.