GEAK compatibility matrix

GEAK compatibility matrix#

2026-07-20

2 min read time

Applies to Linux

Verified hardware, software, runtime, and backend combinations for GEAK 4.0.0 (a Claude Code + JS-Workflow GPU optimizer; no pip package, no CLI). Only tested configurations are listed.

Use the following matrix to view the compatibility and system requirements:

AMD Instinct GPU

ROCm version

Python

Ubuntu

MI300X, MI325X, MI355X

6.4x, 7.0x, 7.1x, 7.2x

3.8, 3.12

22.04, 24.04

Note

  • The on-box card is auto-detected (rocminfo / rocm_agent_enumerator); PYTORCH_ROCM_ARCH is pinned to the local gfx at build time.

  • The on-box GPU arch, backend, profiler, and op backends are re-checked at runtime by the Setup preflight, which writes env_report.{md,json}.

  • Only tested configurations are listed. To report a verified configuration not listed here, open a pull request.

  • For the Python version, the compiled artifacts in the tree are cpython-312.

Runtime — Claude Code#

Component

Version

Notes

Claude Code

≥ 2.1.177

The workflows use the dynamic Workflow (JS orchestration) feature, available only from this version. Check with claude --version.

Launch mode

IS_SANDBOX=1 claude --dangerously-skip-permissions

Workflows spawn sub-agents and run profiling, benchmark, and build commands on the box, so permissions must be auto-approved.

Default model

claude-opus-4-8

Default used by the external-orchestrator entry point (interface/run_e2e.py).

Effort

ultracode

Default effort for interface/run_e2e.py.

Invocation mode#

Mode

Notes

Natural language → Workflow tool

Describe the task to Claude Code; it maps the prompt onto Workflow({ scriptPath, args }).

Direct Workflow call (e2e)

scriptPath: "<repo>/e2e_workflow/e2e_workflow.js"

Direct Workflow call (single kernel)

scriptPath: "<repo>/kernel_workflow/kernel_workflow.js"

External orchestrator (Hyperloom)

python interface/run_e2e.py <handoff.json> <result.json>

Profilers#

The profiler is auto-detected using a degrade ladder (PROFILER_PRIORITY="rocprof-compute rocprofv3 rocprof metrix").

  • rocprof-compute

  • rocprofv3

  • rocprof

  • metrix

Serving backends (e2e_workflow)#

The serving stack is not baked in; args.backend selects scripts/adapters/<backend>.sh.

Backend

Default port

sglang

30000

vllm

8000

Bench client

Notes

Backend-native (bench_e2e.sh)

Default dispatcher.

inferencex

Opt-in using BENCH_CLIENT=inferencex (Hyperloom or Magpie parity); needs $INFERENCEX_PATH.

Kernel languages (kernel_workflow)#

target_language for author mode; also the languages the head-kernel bake-off can win with.

Kernel language

Notes

Triton

Always a viable author target.

FlyDSL

Preferred author target for dense / quantized GEMM (aiter’s SOTA GEMM DSL, JIT, no build). Probed using aiter.ops.flydsl.is_flydsl_available().

HIP

Used when headroom justifies it.

CK (Composable Kernel)

Used when headroom justifies it; FP8 GEMM tuning.

Op-backend bake-off (op_bench.py head kernels)#

  • hipblaslt

  • rocblas or TunableOp

  • aiter

  • triton (with --triton-autotune)

Precision and data types#

Data type

Notes

FP16 / BF16

General kernel optimization.

FP8

Head-GEMM tuning + author target; gsm8k accuracy gate recommended for quantized kernels.

FP4 / MXFP

Quantized GEMM (A4W4); accuracy gate recommended.

Accuracy gate (e2e_workflow)#

Gate

Requirement

none (default)

Throughput delta + greedy output parity.

gsm8k

Sampled gsm8k (5-shot, greedy, fixed seed) using scripts/gsm8k_eval.py against an OpenAI-compatible /v1 endpoint.
Contents