nvidia

/

Muse-Glimmer-30B-NVFP4

like 2

Follow

NVIDIA 67.3k

[by Codex]

Adds an SGLang DFlash serving command for the NVIDIA B200 directly below the existing vLLM example.

The command currently depends on pending SGLang PR #37510. Once that SGLang PR is merged, the temporary pending-PR note in the model card should be removed.

Validation performed on one B200 with the 200-example GSM8K subset:

• Accuracy: 96.5%
• DFlash average acceptance length: 5.25
• Stop rate: 100% (no truncations or request errors)