This model was compressed using kfac_svd with lems rank search starting from Qwen/Qwen3-8B as base model. You may check out our publication and project page for details on kfac-svd and our LEMS rank search.

Compression Details

Metric	Value
Base Model	`Qwen/Qwen3-8B`
Method	`kfac_svd`
Search Method	`lems`
Target Ratio	0.8
Compression Metric	`params`
Recommended Dtype	`float16`
Compressed Layers	103
Total Parameters	6,801,536,103

Usage

The checkpoint records its recommended dtype in config.json; no explicit torch_dtype argument should be needed with this remote-code wrapper. For standard Transformers models, torch_dtype="auto" is the portable fallback.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "MoritzMo123/kfac-svd_lems_Qwen3-8B_0.8",
    trust_remote_code=True,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("MoritzMo123/kfac-svd_lems_Qwen3-8B_0.8")

inputs = tokenizer('Hello, ', return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation Results

Dataset	Perplexity
wikitext2	10.38
ptb	18.32
c4	18.32

Rank Allocation

Per-layer ranks (click to expand)

Layer	Rank
`model.layers.0.mlp.gate_proj`	2192
`model.layers.0.mlp.up_proj`	1952
`model.layers.1.mlp.down_proj`	1456
`model.layers.1.mlp.gate_proj`	1368
`model.layers.1.mlp.up_proj`	2184
`model.layers.1.self_attn.o_proj`	848
`model.layers.1.self_attn.q_proj`	624
`model.layers.10.self_attn.o_proj`	1352
`model.layers.11.self_attn.o_proj`	1328
`model.layers.11.self_attn.q_proj`	1216
`model.layers.12.self_attn.o_proj`	1320
`model.layers.12.self_attn.q_proj`	896
`model.layers.13.self_attn.o_proj`	1168
`model.layers.13.self_attn.q_proj`	1128
`model.layers.14.self_attn.q_proj`	792
`model.layers.15.mlp.down_proj`	2144
`model.layers.15.mlp.gate_proj`	1672
`model.layers.15.mlp.up_proj`	2144
`model.layers.15.self_attn.o_proj`	1376
`model.layers.15.self_attn.q_proj`	640
`model.layers.16.mlp.gate_proj`	1736
`model.layers.16.mlp.up_proj`	2256
`model.layers.16.self_attn.o_proj`	1456
`model.layers.16.self_attn.q_proj`	768
`model.layers.17.mlp.down_proj`	1888
`model.layers.17.mlp.gate_proj`	1304
`model.layers.17.mlp.up_proj`	1840
`model.layers.17.self_attn.o_proj`	1432
`model.layers.17.self_attn.q_proj`	760
`model.layers.18.mlp.down_proj`	1856
`model.layers.18.mlp.gate_proj`	1440
`model.layers.18.mlp.up_proj`	1928
`model.layers.18.self_attn.o_proj`	1528
`model.layers.18.self_attn.q_proj`	824
`model.layers.19.mlp.down_proj`	1800
`model.layers.19.mlp.gate_proj`	1360
`model.layers.19.mlp.up_proj`	1688
`model.layers.19.self_attn.o_proj`	1432
`model.layers.19.self_attn.q_proj`	704
`model.layers.2.mlp.down_proj`	1248
`model.layers.2.mlp.gate_proj`	1488
`model.layers.2.mlp.up_proj`	936
`model.layers.2.self_attn.o_proj`	688
`model.layers.2.self_attn.q_proj`	688
`model.layers.20.mlp.down_proj`	1736
`model.layers.20.mlp.gate_proj`	1368
`model.layers.20.mlp.up_proj`	1648
`model.layers.20.self_attn.o_proj`	1240
`model.layers.20.self_attn.q_proj`	736
`model.layers.21.mlp.down_proj`	1832
`model.layers.21.mlp.gate_proj`	1536
`model.layers.21.mlp.up_proj`	1664
`model.layers.21.self_attn.o_proj`	1304
`model.layers.21.self_attn.q_proj`	992
`model.layers.22.mlp.gate_proj`	1816
`model.layers.22.mlp.up_proj`	1800
`model.layers.22.self_attn.o_proj`	1192
`model.layers.22.self_attn.q_proj`	704
`model.layers.23.mlp.gate_proj`	2088
`model.layers.23.mlp.up_proj`	1928
`model.layers.23.self_attn.q_proj`	760
`model.layers.24.mlp.gate_proj`	2032
`model.layers.24.mlp.up_proj`	1912
`model.layers.24.self_attn.q_proj`	696
`model.layers.25.self_attn.o_proj`	1048
`model.layers.25.self_attn.q_proj`	616
`model.layers.26.self_attn.o_proj`	1136
`model.layers.26.self_attn.q_proj`	624
`model.layers.27.self_attn.o_proj`	760
`model.layers.27.self_attn.q_proj`	616
`model.layers.28.self_attn.o_proj`	864
`model.layers.28.self_attn.q_proj`	640
`model.layers.29.self_attn.o_proj`	832
`model.layers.29.self_attn.q_proj`	672
`model.layers.3.mlp.down_proj`	2232
`model.layers.3.mlp.gate_proj`	2240
`model.layers.3.mlp.up_proj`	1464
`model.layers.3.self_attn.k_proj`	480
`model.layers.3.self_attn.o_proj`	968
`model.layers.3.self_attn.q_proj`	616
`model.layers.30.self_attn.o_proj`	1024
`model.layers.30.self_attn.q_proj`	616
`model.layers.31.self_attn.o_proj`	944
`model.layers.31.self_attn.q_proj`	680
`model.layers.32.self_attn.o_proj`	944
`model.layers.32.self_attn.q_proj`	624
`model.layers.33.self_attn.q_proj`	616
`model.layers.34.self_attn.o_proj`	1168
`model.layers.34.self_attn.q_proj`	624
`model.layers.35.self_attn.o_proj`	824
`model.layers.35.self_attn.q_proj`	632
`model.layers.4.self_attn.o_proj`	1208
`model.layers.4.self_attn.q_proj`	672
`model.layers.5.self_attn.o_proj`	984
`model.layers.5.self_attn.q_proj`	864
`model.layers.6.self_attn.o_proj`	1344
`model.layers.6.self_attn.q_proj`	768
`model.layers.7.self_attn.o_proj`	1128
`model.layers.7.self_attn.q_proj`	656
`model.layers.8.self_attn.o_proj`	1312
`model.layers.8.self_attn.q_proj`	848
`model.layers.9.self_attn.o_proj`	1088
`model.layers.9.self_attn.q_proj`	1176

Hydra Configuration Summary

Config Field	Value
Model	`Qwen/Qwen3-8B`
SVD Method	`kfac_svd`
Search Method	`lems`
Compression Target	`0.8`
Target Metric	`params`
Calibration Dataset	`wikitext2`
Sequence Length	`2048`
Seed	`42`