DX-596-together-gpu-clusters: sync with mintlify-docs#1190

[zainhas]

Sync the together-gpu-clusters skill with the merged mintlify-docs PR that added the TorchTitan training benchmark active health check.

Source PR: togethercomputer/mintlify-docs#1190 -- [TCL-6830]docs: add TorchTitan training health check

Changed docs files that triggered this sync:

• docs/health-checks.mdx

Skill changes:

• Add a new Training Performance (preview) subsection to references/cluster-management.md under Available Health Check Tests.
• Document the TorchTitan Training test: short TorchTitan benchmark measuring steady-state median MFU after skipping warmup.
• Capture the B200-only dependency (Blackwell, sm_100) and the built-in CUDA + PyTorch (FlashAttention 4) stack, ~15 min runtime, on-demand-only status, and that it is not part of automatic acceptance testing.
• Record the default pass threshold: steady-state median MFU >= 30% for Llama 3 8B on a single B200 node (reference ~36.5%), with per-run references for larger models/multi-node runs (e.g. Llama 3 70B ~38% on 4 nodes).
• No changes to SKILL.md, other references, or scripts -- health check test enumeration lives in references/cluster-management.md.

Generated by the Sync Skills Cursor Automation. Please review before merging.

[broly-code-security-scanner]

Broly Security Scan

Note

✅ Clean scan
No vulnerabilities detected in this PR.

Note

Re-scan this PR anytime with /broly scan — useful after /broly undismiss, or to refresh findings without a new push.

Broly — SAST (zai-org/GLM-5.2) · Secrets · SCA · GH Actions (zizmor) · Containers · SBOM · Powered by Together AI