SketchSSM/Qwen3.8-Flash-Next-NVFP4

SketchSSM calibration: Qwen3.8 Flash-Next (NVFP4 weights)

calibration.pt is a portable SketchSSM calibration file for Qwen3.8 Flash-Next. It contains the group-shared sketch basis and the Full-Gram allocation scores from which the per-head rank table and the ordered frames for any mean rank are derived. It contains no model weights.

Calibration weights

Collected with NVFP4 weights: RadixArk/Qwen3.8-Flash-Next-NVFP4 at revision 7b719225242aacd3dbd3f9407468c2ee9a9d2594 (an NVFP4 checkpoint quantized with Model Optimizer). The basis and the allocation scores depend on the weights, so use this file with these weights; for another precision or checkpoint, calibrate with that checkpoint.

Contents

Field	Value
Base model	Qwen3.8 Flash-Next (Gated DeltaNet)
Recurrent layers	36
State heads per layer	48
Key dim K / value dim V	128 / 128
Basis groups per layer	16
Window W	16
Erase factor	yes
Allocation rank cap	60
Basis omega	float32, shape (36, 16, 78, 128)
File size	24,098,949 bytes

Layers are stored in the order of the model's recurrent layers. The file loads with torch.load(..., weights_only=True). The offline calibration guide documents its keys.

Verified mean ranks

For these mean ranks, the derived tables equal the published bundle tables and the exported frames equal those of the bundle export path:

Mean rank	Dense heads	Sketch heads
3	0	1728
4	0	1728
7	1	1727
11	2	1726
26	287	1441

Other mean ranks are allocated with the same rule but have no stored table to compare with. manifest.json lists the SHA-256 of calibration.pt and these results.

Usage

With the SketchSSM repository, export the frames for a mean rank:

hf download SketchSSM/Qwen3.8-Flash-Next-NVFP4 calibration.pt --local-dir calibration
python -m offline_calibration export --calibration calibration/calibration.pt \
  --mean-rank 7 --out frames.pt

With vLLM (requires the SketchSSM vLLM fork with calibration-file support):

vllm serve RadixArk/Qwen3.8-Flash-Next-NVFP4 --sketchssm SketchSSM/Qwen3.8-Flash-Next-NVFP4 --sketchssm-mean-rank 7

How it was produced

From the public calibration bundle offline_calibration/example/qwen_flash_next in the SketchSSM repository:

python -m offline_calibration package --bundle offline_calibration/example/qwen_flash_next --out calibration.pt

package re-allocates every configured mean rank from the packaged contents and fails unless each table equals the bundle table.

License

This calibration file is released under the Apache License 2.0, like the SketchSSM repository. It is derived from the base model's weights, so use it under the base model's license as well.