How-toSelected10月19日 09:22

vLLM cuts startup time 14–18% with precompiled Python bytecode

官方账号vLLM@vllm_project

Post

Why we picked it

Image-level bytecode caching means faster pod spin-up and less wasted GPU time for anyone running vLLM at scale, especially on autoscaling or serverless deployments where cold starts hurt. Engineering teams deploying inference can pick it up by using release images from v0.32.0 or the CUDA nightlies, with no workflow change beyond the image tag.

The vLLM project now ships precompiled Python bytecode in its container images, reducing startup time by 14–18%, about 18 seconds on an H100. The speedup comes from a single uv flag applied at build time, so no recompilation happens at launch. The change is already live in CUDA nightly images and official release images from v0.32.0. Contributor Nils Matteson built the improvement.

Original · vLLM

Cold start, meet build time. 📷

Shipping precompiled Python bytecode in our images cuts vLLM startup by 14–18% (~18s on H100) — one uv flag at build time, zero recompiles at launch.

Already in CUDA nightlies, official release images from v0.32.0. Great work @nilsmatteson 📷

#vLLM #uv #CUDA #inference

Keep reading

Recent stories with the same tag or section

Products BuildersMulti-sourceSelected04:07

PyTorchCon Demo Theater: 13 Sessions on Training, Inference and Agents

The lineup reveals where the PyTorch ecosystem is concentrating effort: inference speed for video models, MoE expert parallelism on NVL72, and fault-tolerant training that survives link and node failures live. Engineers choosing accelerator backends can use these demos to compare AWS Trainium, Intel AI PCs, TPUs and Rebellions hardware on real workloads. The Huawei CRCR CI relay is also a practical pattern for teams maintaining forks against upstream PyTorch.

#PyTorch #PyTorchCon #Wan 2.2 #vLLM

Event

官方账号PyTorch@PyTorch11 sourcesSource

Products BuildersSelected02:54

vLLM-TorchTPU brings native PyTorch serving to Google TPUs

The practical effect is that vLLM's optimizations — prefix caching and speculative decoding — now apply to TPU hardware, giving teams a second accelerator option beyond NVIDIA GPUs for self-hosted inference. Organisations already locked into the PyTorch and vLLM stack can evaluate TPU economics without rewriting their serving code. It also signals Google's continued push to open TPU inference beyond JAX.

#vLLM #TPU #PyTorch #Google

官方账号vLLM@vllm_projectSource

Products BuildersSelected02:01

FlyDSL kernel DSL now supports PyTorch on AMD MI350 GPUs

AMD MI350 support matters because GPU kernel tooling has historically centred on NVIDIA's stack, and this gives teams serving open models a Python-native path to tune kernels without CUDA-only tools. Benchmark and deployment engineers running vLLM on AMD accelerators can adopt the DSL directly from PyTorch's ecosystem and measure the 3.01x serving gain against their own baselines.

#FlyDSL #PyTorch #AMD #vLLM

官方账号PyTorch@PyTorchSource

Industry BuildersSelected01:07

Cohere's Rachad Alao on open science for foundation models

The talk links open model releases to concrete security practices like independent auditing and dynamic safeguards, rather than treating openness and safety as opposing goals. Researchers and ML engineers can follow the named stack — PyTorch, TorchTitan and vLLM — to reproduce or audit open models themselves. Cohere's stance is notable as a US-listed enterprise lab that publicly backs open science. origin

#Cohere #PyTorch #TorchTitan #vLLM

官方账号PyTorch@PyTorchSource

How-to BuildersMulti-sourceSelected01:01

Anyscale session covers Ray Serve LLM deployment and inference scaling

The session bundles inference infrastructure decisions most teams learn the hard way: when disaggregated prefill/decode and KV-cache-aware routing pay off, and whether self-hosting beats token-based API pricing. Engineers deploying open-source models with vLLM or SGLang can register to get a hands-on walkthrough of Ray Serve LLM and agentic-platform integration. Registration is open now via the Anyscale link.

#Ray Serve #vLLM #SGLang #inference

Event

Anyscale@anyscalecompute10 sourcesSource

Models BuildersMulti-sourceSelected00:47

vLLM adds Day-0 support for Recursive's Flint model

A sparse 2.5B-active architecture hitting 500 TPS on one GB200 shows how far inference efficiency now goes without speculative decoding, a benchmark point worth tracking against dense models of similar quality. Teams serving long-context workloads can run Flint in vLLM immediately in either NVFP4 or BF16. Builders evaluating open-weight model serving stacks should compare its throughput on their own GPUs.

#vLLM #Flint #NVFP4 #open-source models

Event

官方账号vLLM@vllm_project8 sourcesSource