ROCm AMD Infinity Context

Caution

This release is an early-access software technology preview. Running production workloads is not recommended.

ROCm(tm) AMD Infinity Context (AIC) is a disaggregated KV-cache inference stack for large language models on AMD Instinct GPUs. It combines an LLM serving framework with a KV Cache block manager to offload KV tensors across a tiered memory hierarchy. GPU VRAM, CPU DRAM, local NVMe and NFS-over-RDMA.

It focuses on enabling a low-latency, shared level of KV Cache that can be accessed by O(1000) GPUs.

ROCm AIC enables long-context serving at scale without recomputation. This repository provides a Docker image build and test harness, benchmarking harness, and Slurm/Spur automation used to validate and demonstrate the platform.

Stack overview

Ubuntu 24.04  (rocm/dev-ubuntu-24.04:7.2.4-complete, ROCm 7.2.4, Python 3.12)
  └── vLLM v0.25.0+rocm723  (pre-built wheel — bundles torch/triton/flash-attn)
        └── LMCacheMPConnector (ZMQ)
              └── LMCache server (standalone MP mode)  [v0.5.2 + 8 AMD patches]
                    ├── L1:  GPU / CPU DRAM   (--l1-size-gb)
                    │    or  hipFile NVMe slab (GDS L1 mode)
                    ├── L2a: NIXL AIS_MT → local NVMe   (hipFile P2PDMA, GDS)
                    └── L2b: NIXL POSIX  → NFS-over-RDMA

Component versions (pinned in docker/Dockerfile — update ARGs there before each release):

Component	Source	Ref
Base OS	rocm/dev-ubuntu-24.04:7.2.4-complete	Ubuntu 24.04, ROCm 7.2.4, Python 3.12
vLLM	wheels.vllm.ai/rocm/0.25.0/rocm723	v0.25.0+rocm723 (pre-built wheel, bundles torch)
LMCache	LMCache/LMCache (upstream)	v0.5.2 + 8 AMD patches
NIXL	ai-dynamo/nixl (upstream)	v1.3.2 + nixl-rocm-ais-mt.patch
hipFile	ROCm/rocm-systems	develop @ 6901b670

Pip-installable nightly wheels

See docs/PIP_WHEELS.md for installation instructions and compatibility notes. Wheels are rebuilt nightly from main and published to the nightly release.

Prerequisites

• ROCm-capable host (MI300X recommended; gfx942). The default build is multi-arch — it bakes in every gfx the vLLM wheel supports (gfx90a, gfx942, gfx950, and the RDNA gfx11xx/gfx12xx line), so one image runs on any of them. Narrow with ROCM_ARCH=gfx942 for a faster single-arch build.
• Docker with BuildKit and the docker compose (v2) plugin (Docker 23+). On a node that lacks it, make ensure-compose installs the plugin user-locally (~/.docker/cli-plugins); the Slurm cliff / smoke / tiny-test jobs self-install it automatically. The whole stack is docker compose only — there is no docker-compose v1 or docker-run fallback.
• Host mounts: local NVMe (NVME_DATA) and NFS-over-RDMA (NFS_DATA) pre-mounted
• HuggingFace token with access to the target model
• Python 3.10+ for host-side benchmarks

Quick start

See docs/QUICK_START.md for the full step-by-step guide (build, start, benchmark, plot). For Slurm / SPUR cluster usage see docs/SLURM_SPUR.md.

Metrics & observability

See docs/METRICS_TELEMETRY.md for the full Prometheus scrape config, exporter details, NIXL telemetry, hsa-snoop, and NFS caveats.

Key environment variables

See docs/ENVIRONMENT.md for the full reference.

License

MIT