Skip to content

Instantly share code, notes, and snippets.

@bigsnarfdude
Created August 16, 2026 14:51
Show Gist options
  • Select an option

  • Save bigsnarfdude/d354a432d5fefbf9b07ddac4f6c50d7b to your computer and use it in GitHub Desktop.

Select an option

Save bigsnarfdude/d354a432d5fefbf9b07ddac4f6c50d7b to your computer and use it in GitHub Desktop.
Every 1.0s: nvidia-smi 192-222-51-153: Sun Aug 16 14:51:34 2026
Sun Aug 16 14:51:34 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.105.08 Driver Version: 580.105.08 CUDA Version: 13.0 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GH200 480GB On | 00000000:DD:00.0 Off | 0 |
| N/A 34C P0 122W / 700W | 88842MiB / 97871MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 64382 C VLLM::EngineCore 88830MiB |
+-----------------------------------------------------------------------------------------+
@bigsnarfdude

Copy link
Copy Markdown
Author

The tokens/sec answer. vLLM BF16 + DFlash2 on GH200: 178 tok/s single-request, 1,030 tok/s
aggregate, 1,262 talks/hr on the real corrector. That's 4.6× against its own no-speculation
control — better than inco's published 2.7–3.4×, because your JSON output drafts unusually
well (acceptance 0.71–0.99). llama.cpp with the drafter only matches vLLM without it.

A build recipe that skips the 40-minute ARM source build — PR #52816 is pure Python/Triton,
so the released wheel's kernels plus the PR's Python gets you serving in ~15 minutes.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment