# Foundry: Template-Based CUDA Graph Context Materialization for Fast LLM Serving Cold Start

arXiv: https://arxiv.org/abs/2604.06664 · HTML: https://mlsanity.com/paper/2604.06664/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Serving systems and request scheduling, one of 12 directions in Efficient LLM inference and serving.
- Builds on 24 of its 52 references that are in the map.
- First posted on arXiv 8 Apr 2026.

## Placed in

- Efficient LLM inference and serving / [Serving systems and request scheduling](https://mlsanity.com/inference/serving-systems-and-request-scheduling/)

## Builds on (most cited first, 24 of 24)

- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Efficient Memory Management for Large Language Model Serving with PagedAttention](https://mlsanity.com/paper/2309.06180/) (2023)
- [Gemma 3 Technical Report](https://mlsanity.com/paper/2503.19786/) (2025)
- [DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving](https://mlsanity.com/paper/2401.09670/) (2024)
- [Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve](https://mlsanity.com/paper/2403.02310/) (2024)
- [LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism](https://mlsanity.com/paper/2404.09526/) (2024)
- [vAttention: Dynamic Memory Management for Serving LLMs without PagedAttention](https://mlsanity.com/paper/2405.04437/) (2024)
- [BurstGPT: A Real-world Workload Dataset to Optimize LLM Serving Systems](https://mlsanity.com/paper/2401.17644/) (2024)
- [Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures](https://mlsanity.com/paper/2505.09343/) (2025)
- [BLITZSCALE: Fast and Live Large Model Autoscaling with O(1) Host Caching](https://mlsanity.com/paper/2412.17246/) (2024)
- [Prism: Cost-Efficient Multi-LLM Serving via GPU Memory Ballooning](https://mlsanity.com/paper/2505.04021/) (2025)
- [DeepServe: Serverless Large Language Model Serving at Scale](https://mlsanity.com/paper/2501.14417/) (2025)
- [HydraServe: Minimizing Cold Start Latency for Serverless LLM Serving in Public Clouds](https://mlsanity.com/paper/2502.15524/) (2025)
- [Expert-as-a-Service: Towards Efficient, Scalable, and Robust Large-scale MoE Serving](https://mlsanity.com/paper/2509.17863/) (2025)
- [Elastic MoE: Unlocking the Inference-Time Scalability of Mixture-of-Experts](https://mlsanity.com/paper/2509.21892/) (2025)
- [λScale: Enabling Fast Scaling for Serverless Large Language Model Inference](https://mlsanity.com/paper/2502.09922/) (2025)
- [PhoenixOS: Concurrent OS-level GPU Checkpoint and Restore with Validated Speculation](https://mlsanity.com/paper/2405.12079/) (2024)
- [FLYING SERVING: On-the-Fly Parallelism Switching for Large Language Model Serving](https://mlsanity.com/paper/2602.22593/) (2026)
- [ConServe: Fine-Grained GPU Harvesting for LLM Online and Offline Co-Serving](https://mlsanity.com/paper/2410.01228/) (2024)
- [CRIUgpu: Transparent Checkpointing of GPU-Accelerated Workloads](https://mlsanity.com/paper/2502.16631/) (2025)
- [WarmServe: Enabling One-for-Many GPU Prewarming for Multi-LLM Serving](https://mlsanity.com/paper/2512.09472/) (2025)
- [CRAC: Checkpoint-Restart Architecture for CUDA with Streams and UVM](https://mlsanity.com/paper/2008.10596/) (2020)
- [Amoeba: Runtime Tensor Parallel Transformation for LLM Inference Services](https://mlsanity.com/paper/2509.19729/) (2025)
- [CRUM: Checkpoint-Restart Support for CUDA's Unified Memory](https://mlsanity.com/paper/1808.00117/) (2018)

## Built on by (most cited first, 2 of 2)

- [Fast Recovery for LLM Serving via Decoupled Device Memory Lifetime in Dynamo](https://mlsanity.com/paper/2609.25451/) (2026)
- [FastEEP: Fast Intra-Instance Elastic Scaling for Expert-Parallel MoE Serving](https://mlsanity.com/paper/2610.18451/) (2026)

## Abstract (arXiv)

Modern LLM service providers increasingly rely on autoscaling and parallelism reconfiguration to respond to rapidly changing workloads, but cold-start latency remains a major bottleneck. While recent systems have reduced model weight loading to seconds, CUDA graph capture still takes tens of seconds to minutes and often dominates startup. Unfortunately, CUDA graphs cannot be naively serialized: beyond graph topology, they are tightly coupled to execution context, including device addresses embedded in kernel arguments and kernel code lazily loaded during warmup. Existing approaches either rely on brittle kernel-specific patching or heavyweight process-level checkpoint/restore that are inflexible to dynamic parallelism switching. We present Foundry, a template-based CUDA graph context materialization system that persists both graph topology and execution context during an offline processing stage, and reconstructs executable graphs online with negligible overhead. Foundry enforces deterministic memory layouts, automatically extracts and reloads kernel binaries required by captured graphs, and reduces online reconstruction costs through topology-based templating. For distributed serving, Foundry further enables a single-GPU offline capture to generate templates for multi-GPU deployments by patching only rank-dependent communication state. Across dense and MoE models up to 235B parameters, Foundry reduces cold-start latency by up to 99%, cutting the initialization time of Qwen3-235B-A22B from 10 minutes to 3.9 seconds while preserving the throughput gains of CUDA graphs.