LLM-42: Enabling Determinism in LLM Inference with Verified Speculation

_{For SOSP 2026 artifact evaluation, please see the sosp-artifact/ README.}

Raja Gond‡, Aditya K Kamath†, Ramachandran Ramjee‡, Ashish Panwar‡
‡Microsoft Research †University of Washington

---

LLM-42 enables deterministic LLM inference via a decode–verify–rollback protocol, without rewriting GPU kernels. Built on SGLang v0.5.3.

How it works

Standard LLM serving is non-deterministic: dynamic batching changes GPU reduction orders, producing different outputs across runs. LLM-42 fixes this with a lightweight verify-rollback loop:

1. Decode — generate tokens using fast, unmodified kernels with dynamic batching.
2. Verify — replay a window of tokens under a fixed-shape schedule to check consistency.
3. Rollback — on mismatch, discard inconsistent tokens and resume from the last verified position.

Only requests marked is_deterministic=True incur verification; the rest run at full speed.

Quick start

# Create and attach to a GPU-enabled Docker container (uses lmsysorg/sglang:v0.5.4)
./run_container.sh create && ./run_container.sh attach
# Inside the container: workspace is mounted at /workspacecd /workspace
apt update; apt upgrade -y
git config --global --add safe.directory /workspace
# Build sgl-kernel and install sglang in editable mode
./build_all.sh
# Authenticate with Hugging Face to download gated models (e.g., Llama)
huggingface-cli login --token <HF_TOKEN># Terminal 1: Start the LLM-42 server (waits for model to load)
bash llm42_benchmarks/basic/launch_server.sh
# Terminal 2: Once the server is ready, run the determinism-check client
python3 llm42_benchmarks/basic/client.py

Configuration

Flag	Default	Description
--enable-llm42	0	Set to 3 to enable LLM-42 DVR
--llm42-window-size	64	Tokens decoded before verification
--llm42-verify-batch-size	8	Requests per verification batch (grouped verification)

Additional flags for benchmarking:--enable-deterministic-inference 2 (global batch-invariant baseline), --llm42-skip-mismatch (mismatch rate control / synthetic mismatch injection).

Hardware

4× NVIDIA H100 PCIe (80 GB HBM3), 64-core CPU, ~1.65 TB DRAM.

Project Structure

├── python/sglang/
│ ├── srt/
│ │ ├── llm42/ # Core LLM-42 decode–verify–rollback logic
│ │ ├── batch_invariant_ops/ # Batch-invariant kernel wrappers
│ │ ├── layers/ # Model layers (attention, MoE, etc.)
│ │ ├── models/ # Supported model architectures
│ │ ├── managers/ # Request scheduling & memory management
│ │ └── sampling/ # Sampling strategies
│ ├── launch_server.py # Server entry point
│ └── bench_serving.py # Serving benchmark client
├── sgl-kernel/ # Custom CUDA/Triton kernels
│ ├── csrc/ # C++/CUDA sources
│ └── python/ # Python bindings
├── llm42_benchmarks/ # LLM-42 benchmark scripts
├── llm42-plots/ # Plotting scripts for paper figures
├── benchmark/ # Upstream SGLang benchmarks
├── docker/ # Dockerfiles & Kubernetes manifests
├── scripts/ # CI, utility, and helper scripts
├── build_all.sh # Build sgl-kernel + install sglang
└── run_container.sh # Create/attach to dev container

---

Citation

@article{gond2025llm42,
title = {{LLM-42}: Enabling Determinism in {LLM} Inference with Verified Speculation},
author = {Gond, Raja and Kamath, Aditya K and Ramjee, Ramachandran and Panwar, Ashish},
journal = {arXiv preprint arXiv:2601.17768},
year = {2026},
url = {https://arxiv.org/abs/2601.17768}
}

License

This project is licensed under the terms in the LICENSE file. It is built on SGLang, which is licensed under the Apache License 2.0.

Trademark Notice

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.