|
Download README.md from SketchSSM/README: direct link, hf CLI and curl.
- Browser
- Download file 8.83 kB
-
https://huggingface.co/spaces/SketchSSM/README/resolve/main/README.md
- Command line
-
hf download hf://spaces/SketchSSM/README/README.md
-
curl -L -o README.md https://huggingface.co/spaces/SketchSSM/README/resolve/main/README.md
8.83 kB
metadata
title: README
emoji: 📐
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
SketchSSM calibrations
SketchSSM speeds up the decode of hybrid
linear-attention models: the recurrent state is updated exactly once per window,
and the other decode steps read a compact per-request sketch of it. Each
repository here holds the calibration file (calibration.pt) for one model
checkpoint; it contains no model weights.
vllm serve <base model> --sketchssm SketchSSM/<calibration repo> \
--sketchssm-mean-rank 8 --mamba-ssm-cache-dtype float32
Use the calibration of the exact checkpoint (precision included) you serve. To calibrate another model, see the calibration guide.
Mamba-2 (NemotronH)
Gated DeltaNet
KDA
| Calibration | Base checkpoint | Weights |
|---|---|---|
SketchSSM/GLM-5.3-Flash-NVFP4 |
RedHatAI/GLM-5.3-Flash-NVFP4 |
NVFP4 |