Hyperloom compatibility matrix

2026-08-27

4 min read time

Applies to Linux

This topic lists the hardware, inference frameworks, and container images that Hyperloom is validated against.

Note

ROCm versions or framework builds not listed in this matrix might work, but are not regularly tested.

Hyperloom support matrix

The following table lists the minimum requirements for running Hyperloom.

Requirement	Support
AMD Instinct™ GPU	MI300X, MI325X, MI355X
Operating System	Ubuntu 22.04, Ubuntu 24.04
ROCm Version	7.2.x
Python	>= 3.10
Inference Framework	SGLang (>= 0.5.12), vLLM (>= 0.21.0), plus custom for your own benchmark script
Kernel Languages	HIP, Triton, FlyDSL

Component support matrix

The following table lists the validated Hyperloom version and component combinations.

Hyperloom version	Component	GPU	ROCm version	Ubuntu	Python	GitHub
1.0.0	TraceLens 1.0.0	Hardware-agnostic	No dependency	OS-independent	>= 3.6
	GEAK 4.0.0	MI300X, MI325X, MI355X	6.4.x, 7.0.x, 7.1.x, 7.2.x, 10.0.0	22.04, 24.04	3.8, 3.12
	IntelliKit 0.1.1	MI300X, MI325X, MI355X	7.2.x, 10.0.0	22.04, 24.04	>= 3.10
	AgentKernelArena 0.2.0	MI300X, MI325X, MI355X	7.2.x	22.04, 24.04	>= 3.10
	Magpie 0.2.0	MI300X, MI325X, MI355X	7.0.x, 7.1.x, 7.2.x, 10.0.0	22.04, 24.04	>= 3.10

Note

TraceLens does not have hard requirements for the GPU, ROCm version, or the OS; it has scripts to verify whether a trace is valid/parseable. TraceLens is:

• OS-independent and runs anywhere Python does.

• Not limited to MI300X/MI325X/MI355X; it’s hardware-agnostic.

See the TraceLens documentation for more information.

Note

MI325X shares the gfx942/CDNA3 runner family with MI300X. Hyperloom keeps the resolved GPU types distinct, but Magpie benchmark rendering reuses the MI300X runner scripts and image family unless a dedicated image is supplied.

Inference frameworks

The following inference frameworks are supported:

Framework	ROCm version	Notes
SGLang	7.2.4	Default framework; recommended docker/bare-metal stack uses rocm724 (see below)
vLLM	7.2.3	Do not mix frameworks within one session
custom	Host-defined	Escape hatch for your own benchmark script; Hyperloom does not manage the server lifecycle. Requires HYPERLOOM_BENCHMARK_BACKEND=bypass plus --framework-path (or FRAMEWORK_REPO_PATH) and --benchmark-scripts-dir (or HYPERLOOM_BYPASS_SCRIPTS_DIR); the CLI exits with status 2 when any of the three is missing.

Container images

Pick the image that matches your environment. Public Docker Hub refs are used on your own GPU machine: the official upstream lmsysorg/sglang-rocm:<tag> for SGLang and vllm/vllm-openai-rocm:<tag> for vLLM. If your deployment uses a private registry mirror, set the registry prefix accordingly.

Image	GPU
lmsysorg/sglang-rocm:v0.5.17-rocm724-mi30x-20260821	MI300X / MI325X
lmsysorg/sglang-rocm:v0.5.17-rocm724-mi35x-20260821	MI355X
vllm/vllm-openai-rocm:v0.27.1	MI300X / MI325X / MI355X

The vLLM image entrypoint is vllm serve, so override it (for example --entrypoint tail) when starting a long-running Hyperloom container.

Browse all available tags at hub.docker.com/r/lmsysorg/sglang-rocm/tags and hub.docker.com/r/vllm/vllm-openai-rocm/tags.

Bare-metal recommended environment

For baremetal setup, align the host to this combination before running setup. Hyperloom does not install ROCm or torch itself.

Item	Recommended	Notes
ROCm	7.2.x	The patch level differs per framework and is the same in both setup modes: the vLLM stack uses ROCm 7.2.3 and the SGLang stack uses ROCm 7.2.4 (see the note below).
Python	3.12	Required by the vLLM ROCm wheel.
ROCm torch	ROCm build matching the host ROCm	Preinstalled by the operator; not managed by Hyperloom.
SGLang	v0.5.17 (rocm724)	Installed in shared mode (reuses the host torch). Uses the ROCm 7.2.4 AMD wheel index (SGLANG_ROCM_EXTRA=rocm724), so the SGLang ROCm layer is 7.2.4. Note: SGLANG_REF (v0.5.17) only pins the version on the source-install branch (non-3.10 Python); on Python 3.10 the AMD wheel index installs amd-sglang unpinned, which might resolve to a different patch release.
vLLM	v0.27.1 (rocm723), isolated venv	Installs vllm==0.27.1+rocm723 from the wheels.vllm.ai pip index. vLLM’s ROCm wheel pins its own torch, so it installs into a dedicated venv (--framework-env isolated, the default for vLLM) and never touches the host torch.

Bare-metal ROCm patch levels differ per framework, and each one matches its container image. The vLLM stack installs the rocm723 variant (ROCm 7.2.3), matching vllm/vllm-openai-rocm:v0.27.1; the SGLang stack installs from the ROCm 7.2.4 AMD wheel index, matching the two lmsysorg/sglang-rocm:v0.5.17-rocm724-* images. docker mode is still the preferred route for a pre-validated stack, since the images also pin the surrounding torch, Triton, and AITER builds.

These are recommended defaults, not hard pins. Framework and ROCm versions are overridable via env (SGLANG_REF, SGLANG_ROCM_EXTRA, VLLM_VERSION, VLLM_ROCM_VARIANT) for hosts that need a different pinned stack.