AITOPAI热报
How-toSelected

vLLM KV Cache Offloading Poster at PyTorch Conference

官方账号PyTorch@PyTorch
Post
View the post
Why we picked it

The routing layer is the interesting part: simply offloading KV cache to CPU or storage is not enough, since a request must land on the node whose cache already holds its blocks. Engineers running multi-node inference can use this to decide between cheaper CPU offload and shared-storage tiers. The poster session is open to PyTorch Conference attendees on 20-21 October in San Jose.

vLLM offloads its KV cache to CPU or shared storage when GPU memory fills up. The llm.d project tracks where every cache block lives and routes each request to the cache holding it. Bongwoo Bak from Moreh will present a poster titled "Where the Blocks Live" at PyTorch Conference North America in San Jose on 20-21 October, comparing when CPU offloading suffices and when shared storage pays off.

Original · PyTorch

When GPU memory fills up, @vllm_project offloads the KV cache to CPU or storage. @_llm_d_ tracks where every block lives, and sends each request to its cache. At PyTorch Conference North America, Bongwoo Bak from Moreh will present a poster on "Where the Blocks Live." You'll see when CPU is enough and when shared storage pays off.

Join us at #PyTorchCon in San Jose, October 20-21: https://t.co/jBApW8ocHQ

#vLLM#KV cache#llm.d#PyTorch#inference

Keep reading

Recent stories with the same tag or section
ProductsBuildersMulti-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
ProductsBuildersSelected02:54
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
ProductsBuildersSelected02:01
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
IndustryBuildersSelected01:07
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