Hyperloom compatibility matrix

Hyperloom compatibility matrix#

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)
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
1.0.0a1	TraceLens 0.1.0	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	22.04, 24.04	3.8, 3.12
	IntelliKit 0.1.0	MI300X, MI325X, MI355X	7.2.x	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	22.04, 24.04	>= 3.10

Note

MI325X shares the gfx942/CDNA3 runner family with MI300X. Hyperloom keeps the resolved GPU type distinct (mi325x), 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	7.2.0	Default framework
vLLM	7.2.0	Do not mix frameworks within one session
Atom	7.2.0	Single-node only (multi-node rejected by the IR-8 guard)
xDiT (diffusion)	7.2.0	Scriptable diffusion pipeline (no serving server). Internal throughput is tracked in img/s, but the primary session-facing metric is end-to-end latency `e2el_mean_ms` (ms).

Container images#

Pick the image that matches your environment. Public Docker Hub refs (primussafe/sglang:<tag>) are used on your own GPU machine. If your deployment uses a private registry mirror, set the registry prefix accordingly.

Image	GPU
primussafe/sglang:v0.5.12-rocm720-mi30x-profilerfix	MI300X / MI325X
primussafe/sglang:v0.5.12-rocm720-mi35x-profilerfix	MI355X
primussafe/vllm-openai-rocm:v0.21.0-rocm720-profilerfix	MI300X / MI325X / MI355X

Browse all available SGLang tags at hub.docker.com/r/primussafe/sglang/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	Bare-metal patch level differs per framework: the vLLM stack uses ROCm 7.2.2 and the SGLang stack uses ROCm 7.2.0 (see the note below). `docker` mode uses ROCm 7.2.0 for both frameworks (the `-rocm720-` images).
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	v0.5.12 (rocm720)	Installed in `shared` mode (reuses the host torch). Uses the ROCm 7.2.0 AMD wheel index (`SGLANG_ROCM_EXTRA=rocm720`), so the SGLang ROCm layer is 7.2.0.
vLLM	v0.21.0 (rocm722), isolated venv	Installs `vllm==0.21.0+rocm722` 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. The vLLM version matches the v0.21.0 Docker image, but the pip index only publishes the rocm722 variant (ROCm 7.2.2), so the bare-metal vLLM ROCm layer is 7.2.2 rather than the image’s rocm720. The SGLang stack installs from the ROCm 7.2.0 AMD wheel index, so its ROCm layer is 7.2.0. In docker mode both frameworks use ROCm 7.2.0 (the -rocm720- images). For a fully validated, pre-aligned vLLM stack with rocm720, prefer docker mode with primussafe/vllm-openai-rocm:v0.21.0-rocm720-profilerfix.

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.