+13% tokens/sec on AMD MI300X.
6 months ahead of AMD's perf team.

graphx is the cross-vendor Triton config-transfer layer for AMD MI300X. We automatically translate every NVIDIA-tuned kernel config into a competitive AMD config the day vLLM ships it — so your inference fleet never runs stale.

Install in 5 minutes

Drop the sidecar into your vLLM-AMD cluster:

docker run -d \
  --name graphx-agent \
  --network host \
  -v $VLLM_CONFIGS_DIR:/vllm-configs \
  -e GRAPHX_API_KEY=$YOUR_KEY \
  -e GRAPHX_CLUSTER_ID=$YOUR_CLUSTER \
  ghcr.io/graphx/agent:latest \
  --vllm-configs-root=/vllm-configs

Talk to founder →

Proof — 269 commits from vLLM's public git history

Every config we transfer is a public commit you can verify. Each row links to the actual NVIDIA-committed and AMD-committed file pair.

What's in the paper

• Cross-Vendor Triton Config Transfer for AMD MI300X (preprint) — full 269-triple analysis, bootstrap CIs, mfma-block ablation
• graphx on GitHub — S1-S8 strategies as a Python package
• Phase-validation report — every claim backed by raw measurements