SketchSSM/Qwen3.6-35B-A3B-BF16

SketchSSM calibration: Qwen3.6 35B-A3B (BF16 weights)

calibration.pt is a portable SketchSSM calibration file for Qwen/Qwen3.6-35B-A3B. 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, and the model layer index of each calibrated recurrent layer. It contains no model weights.

Calibration weights

Collected with BF16 weights: Qwen/Qwen3.6-35B-A3B at revision 995ad96eacd98c81ed38be0c5b274b04031597b0. 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	Qwen3.6 35B-A3B (Gated DeltaNet)
Recurrent layers	30 (model layers 0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14, 16, 17, 18, 20, 21, 22, 24, 25, 26, 28, 29, 30, 32, 33, 34, 36, 37, 38)
State heads per layer	32
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 (30, 16, 78, 128)
File size	19,779,525 bytes

The file loads with torch.load(..., weights_only=True). The calibration guide documents its keys.

Verified mean ranks

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

Mean rank	Dense heads	Sketch heads
2	0	960
4	0	960
8	1	959
16	32	928

Other mean ranks are allocated with the same rule. manifest.json lists the SHA-256 of calibration.pt and these results.

Usage

With vLLM's SketchSSM support (--sketchssm), the ranks for the requested mean rank are allocated when the model loads:

vllm serve Qwen/Qwen3.6-35B-A3B \
  --sketchssm SketchSSM/Qwen3.6-35B-A3B-BF16 --sketchssm-mean-rank 8 \
  --mamba-ssm-cache-dtype float32

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

hf download SketchSSM/Qwen3.6-35B-A3B-BF16 calibration.pt --local-dir calibration
python -m sketchssm.calibration export --calibration calibration/calibration.pt \
  --mean-rank 8 --out frames.pt

How it was produced

With the public calibration pipeline of the SketchSSM repository (release 0.1.15) and collect.yaml from this repository:

python -m sketchssm.calibration calibrate --config collect.yaml --out out
python -m sketchssm.calibration package --bundle out --out calibration.pt
python -m sketchssm.calibration manifest --bundle out --calibration calibration.pt --out manifest.json

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.