Remove stale fork patch: superseded by xianbaoqian/vllm#1 (fix-spec-decode), which the card now points to

Fix install step: plain editable install, not --no-deps (matches the validated setup)

Add SGLang instructions via cloudnathan5/sglang W4A16 fork: 111-207 tok/s with DFlash

Make ignore list portable: name the unquantized vision stack explicitly (SGLang compatibility)

Document engine support (SGLang cannot serve W4A16 yet) and the related Preyazz checkpoint

Record that the quant is verified on the upstream PR + fix-spec-decode route

Point install instructions at upstream PR #51655 + xianbaoqian/vllm#1 rather than this fork

GPTQ-rounded NVFP4A16: PPL 5.5631 (was 5.5965 RTN), 139.3 tok/s with DFlash

Update README.md

Replace W4A4 with NVFP4A16: faster (74.7 vs 54.3 tok/s) and better PPL (5.5965 vs 5.7660)

Update vllm-muse-glimmer-fork.patch: point at live fork branch, refresh patch

Update README.md: point at live fork branch, refresh patch

Upload folder using huggingface_hub

initial commit