Modal Labs

Modal LabsSince 2022-11

Modal Labs is a high-performance serverless cloud platform that enables AI engineers and developers to run Python code in the cloud with instant access to thousands of CPUs, GPUs, and persistent network volumes. Founded by former Spotify CTO Erik Bernhardsson, Modal reimagines cloud computing with sub-second cold starts and zero infrastructure configuration. With Modal, you define your container image, dependencies, and GPU hardware directly inside standard Python code using simple decorators (e.g. `@app.function(gpu="H100")`). Modal handles container building, volume mounting, GPU scheduling, and automatic scaling down to zero in milliseconds, making it the premier choice for running generative AI models, ComfyUI video pipelines, and massive parallel batch jobs.

Modal operates a custom container runtime built in Rust that bypasses standard Docker daemon overhead, allowing container images to spawn in under 900 milliseconds. Its distributed filesystem mounts shared NetworkFileSystem (NFS) volumes across thousands of simultaneous workers with near-local NVMe read speeds. Modal supports NVIDIA T4, L4, A10G, A100 (40GB/80GB), and H100 SXM5 GPUs. Developers can attach web endpoints (`@app.web_endpoint`), schedule recurring cron tasks, execute distributed map-reduce jobs across tens of thousands of cores, and monitor live streaming logs via the interactive web console.

vLLMSince 2023-06

vLLM is the industry-standard open-source LLM serving and inference engine designed for ultra-high throughput and minimal memory waste. Developed by UC Berkeley researchers, vLLM introduced PagedAttention—a revolutionary memory management algorithm that manages attention key-value (KV) cache like virtual memory in operating systems, virtually eliminating memory fragmentation. Capable of delivering 2x to 4x higher throughput than Hugging Face TGI and standard PyTorch runtimes, vLLM powers production AI inference infrastructure across enterprise cloud clusters and high-volume API providers worldwide.

vLLM features state-of-the-art inference optimizations including continuous request batching, Chunked Prefill, speculative decoding, prefix caching, and native quantization support (AWQ, GPTQ, FP8, INT4, SqueezeLLM). It provides drop-in OpenAI-compatible REST API endpoints, supports multi-GPU distributed tensor parallelism with Ray/NCCL, and serves all major model architectures including DeepSeek-V3, Llama 3.3, Mistral, Qwen 2.5, and Command R+.