This model was compressed using kfac_svd with uniform rank search starting from unsloth/mistral-7b 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	`unsloth/mistral-7b`
Method	`kfac_svd`
Search Method	`uniform`
Target Ratio	0.7
Compression Metric	`params`
Recommended Dtype	`float16`
Compressed Layers	224
Total Parameters	5,145,628,896

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_uniform_mistral-7b_0.7",
    trust_remote_code=True,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("MoritzMo123/kfac-svd_uniform_mistral-7b_0.7")

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	9.29
ptb	110.72
c4	34.64

Rank Allocation

Per-layer ranks (click to expand)

Layer	Rank
`model.layers.0.mlp.down_proj`	2229
`model.layers.0.mlp.gate_proj`	2229
`model.layers.0.mlp.up_proj`	2229
`model.layers.0.self_attn.k_proj`	573
`model.layers.0.self_attn.o_proj`	1433
`model.layers.0.self_attn.q_proj`	1433
`model.layers.0.self_attn.v_proj`	573
`model.layers.1.mlp.down_proj`	2229
`model.layers.1.mlp.gate_proj`	2229
`model.layers.1.mlp.up_proj`	2229
`model.layers.1.self_attn.k_proj`	573
`model.layers.1.self_attn.o_proj`	1433
`model.layers.1.self_attn.q_proj`	1433
`model.layers.1.self_attn.v_proj`	573
`model.layers.10.mlp.down_proj`	2229
`model.layers.10.mlp.gate_proj`	2229
`model.layers.10.mlp.up_proj`	2229
`model.layers.10.self_attn.k_proj`	573
`model.layers.10.self_attn.o_proj`	1433
`model.layers.10.self_attn.q_proj`	1433
`model.layers.10.self_attn.v_proj`	573
`model.layers.11.mlp.down_proj`	2229
`model.layers.11.mlp.gate_proj`	2229
`model.layers.11.mlp.up_proj`	2229
`model.layers.11.self_attn.k_proj`	573
`model.layers.11.self_attn.o_proj`	1433
`model.layers.11.self_attn.q_proj`	1433
`model.layers.11.self_attn.v_proj`	573
`model.layers.12.mlp.down_proj`	2229
`model.layers.12.mlp.gate_proj`	2229
`model.layers.12.mlp.up_proj`	2229
`model.layers.12.self_attn.k_proj`	573
`model.layers.12.self_attn.o_proj`	1433
`model.layers.12.self_attn.q_proj`	1433
`model.layers.12.self_attn.v_proj`	573
`model.layers.13.mlp.down_proj`	2229
`model.layers.13.mlp.gate_proj`	2229
`model.layers.13.mlp.up_proj`	2229
`model.layers.13.self_attn.k_proj`	573
`model.layers.13.self_attn.o_proj`	1433
`model.layers.13.self_attn.q_proj`	1433
`model.layers.13.self_attn.v_proj`	573
`model.layers.14.mlp.down_proj`	2229
`model.layers.14.mlp.gate_proj`	2229
`model.layers.14.mlp.up_proj`	2229
`model.layers.14.self_attn.k_proj`	573
`model.layers.14.self_attn.o_proj`	1433
`model.layers.14.self_attn.q_proj`	1433
`model.layers.14.self_attn.v_proj`	573
`model.layers.15.mlp.down_proj`	2229
`model.layers.15.mlp.gate_proj`	2229
`model.layers.15.mlp.up_proj`	2229
`model.layers.15.self_attn.k_proj`	573
`model.layers.15.self_attn.o_proj`	1433
`model.layers.15.self_attn.q_proj`	1433
`model.layers.15.self_attn.v_proj`	573
`model.layers.16.mlp.down_proj`	2229
`model.layers.16.mlp.gate_proj`	2229
`model.layers.16.mlp.up_proj`	2229
`model.layers.16.self_attn.k_proj`	573
`model.layers.16.self_attn.o_proj`	1433
`model.layers.16.self_attn.q_proj`	1433
`model.layers.16.self_attn.v_proj`	573
`model.layers.17.mlp.down_proj`	2229
`model.layers.17.mlp.gate_proj`	2229
`model.layers.17.mlp.up_proj`	2229
`model.layers.17.self_attn.k_proj`	573
`model.layers.17.self_attn.o_proj`	1433
`model.layers.17.self_attn.q_proj`	1433
`model.layers.17.self_attn.v_proj`	573
`model.layers.18.mlp.down_proj`	2229
`model.layers.18.mlp.gate_proj`	2229
`model.layers.18.mlp.up_proj`	2229
`model.layers.18.self_attn.k_proj`	573
`model.layers.18.self_attn.o_proj`	1433
`model.layers.18.self_attn.q_proj`	1433
`model.layers.18.self_attn.v_proj`	573
`model.layers.19.mlp.down_proj`	2229
`model.layers.19.mlp.gate_proj`	2229
`model.layers.19.mlp.up_proj`	2229
`model.layers.19.self_attn.k_proj`	573
`model.layers.19.self_attn.o_proj`	1433
`model.layers.19.self_attn.q_proj`	1433
`model.layers.19.self_attn.v_proj`	573
`model.layers.2.mlp.down_proj`	2229
`model.layers.2.mlp.gate_proj`	2229
`model.layers.2.mlp.up_proj`	2229
`model.layers.2.self_attn.k_proj`	573
`model.layers.2.self_attn.o_proj`	1433
`model.layers.2.self_attn.q_proj`	1433
`model.layers.2.self_attn.v_proj`	573
`model.layers.20.mlp.down_proj`	2229
`model.layers.20.mlp.gate_proj`	2229
`model.layers.20.mlp.up_proj`	2229
`model.layers.20.self_attn.k_proj`	573
`model.layers.20.self_attn.o_proj`	1433
`model.layers.20.self_attn.q_proj`	1433
`model.layers.20.self_attn.v_proj`	573
`model.layers.21.mlp.down_proj`	2229
`model.layers.21.mlp.gate_proj`	2229
`model.layers.21.mlp.up_proj`	2229
`model.layers.21.self_attn.k_proj`	573
`model.layers.21.self_attn.o_proj`	1433
`model.layers.21.self_attn.q_proj`	1433
`model.layers.21.self_attn.v_proj`	573
`model.layers.22.mlp.down_proj`	2229
`model.layers.22.mlp.gate_proj`	2229
`model.layers.22.mlp.up_proj`	2229
`model.layers.22.self_attn.k_proj`	573
`model.layers.22.self_attn.o_proj`	1433
`model.layers.22.self_attn.q_proj`	1433
`model.layers.22.self_attn.v_proj`	573
`model.layers.23.mlp.down_proj`	2229
`model.layers.23.mlp.gate_proj`	2229
`model.layers.23.mlp.up_proj`	2229
`model.layers.23.self_attn.k_proj`	573
`model.layers.23.self_attn.o_proj`	1433
`model.layers.23.self_attn.q_proj`	1433
`model.layers.23.self_attn.v_proj`	573
`model.layers.24.mlp.down_proj`	2229
`model.layers.24.mlp.gate_proj`	2229
`model.layers.24.mlp.up_proj`	2229
`model.layers.24.self_attn.k_proj`	573
`model.layers.24.self_attn.o_proj`	1433
`model.layers.24.self_attn.q_proj`	1433
`model.layers.24.self_attn.v_proj`	573
`model.layers.25.mlp.down_proj`	2229
`model.layers.25.mlp.gate_proj`	2229
`model.layers.25.mlp.up_proj`	2229
`model.layers.25.self_attn.k_proj`	573
`model.layers.25.self_attn.o_proj`	1433
`model.layers.25.self_attn.q_proj`	1433
`model.layers.25.self_attn.v_proj`	573
`model.layers.26.mlp.down_proj`	2229
`model.layers.26.mlp.gate_proj`	2229
`model.layers.26.mlp.up_proj`	2229
`model.layers.26.self_attn.k_proj`	573
`model.layers.26.self_attn.o_proj`	1433
`model.layers.26.self_attn.q_proj`	1433
`model.layers.26.self_attn.v_proj`	573
`model.layers.27.mlp.down_proj`	2229
`model.layers.27.mlp.gate_proj`	2229
`model.layers.27.mlp.up_proj`	2229
`model.layers.27.self_attn.k_proj`	573
`model.layers.27.self_attn.o_proj`	1433
`model.layers.27.self_attn.q_proj`	1433
`model.layers.27.self_attn.v_proj`	573
`model.layers.28.mlp.down_proj`	2229
`model.layers.28.mlp.gate_proj`	2229
`model.layers.28.mlp.up_proj`	2229
`model.layers.28.self_attn.k_proj`	573
`model.layers.28.self_attn.o_proj`	1433
`model.layers.28.self_attn.q_proj`	1433
`model.layers.28.self_attn.v_proj`	573
`model.layers.29.mlp.down_proj`	2229
`model.layers.29.mlp.gate_proj`	2229
`model.layers.29.mlp.up_proj`	2229
`model.layers.29.self_attn.k_proj`	573
`model.layers.29.self_attn.o_proj`	1433
`model.layers.29.self_attn.q_proj`	1433
`model.layers.29.self_attn.v_proj`	573
`model.layers.3.mlp.down_proj`	2229
`model.layers.3.mlp.gate_proj`	2229
`model.layers.3.mlp.up_proj`	2229
`model.layers.3.self_attn.k_proj`	573
`model.layers.3.self_attn.o_proj`	1433
`model.layers.3.self_attn.q_proj`	1433
`model.layers.3.self_attn.v_proj`	573
`model.layers.30.mlp.down_proj`	2229
`model.layers.30.mlp.gate_proj`	2229
`model.layers.30.mlp.up_proj`	2229
`model.layers.30.self_attn.k_proj`	573
`model.layers.30.self_attn.o_proj`	1433
`model.layers.30.self_attn.q_proj`	1433
`model.layers.30.self_attn.v_proj`	573
`model.layers.31.mlp.down_proj`	2229
`model.layers.31.mlp.gate_proj`	2229
`model.layers.31.mlp.up_proj`	2229
`model.layers.31.self_attn.k_proj`	573
`model.layers.31.self_attn.o_proj`	1433
`model.layers.31.self_attn.q_proj`	1433
`model.layers.31.self_attn.v_proj`	573
`model.layers.4.mlp.down_proj`	2229
`model.layers.4.mlp.gate_proj`	2229
`model.layers.4.mlp.up_proj`	2229
`model.layers.4.self_attn.k_proj`	573
`model.layers.4.self_attn.o_proj`	1433
`model.layers.4.self_attn.q_proj`	1433
`model.layers.4.self_attn.v_proj`	573
`model.layers.5.mlp.down_proj`	2229
`model.layers.5.mlp.gate_proj`	2229
`model.layers.5.mlp.up_proj`	2229
`model.layers.5.self_attn.k_proj`	573
`model.layers.5.self_attn.o_proj`	1433
`model.layers.5.self_attn.q_proj`	1433
`model.layers.5.self_attn.v_proj`	573
`model.layers.6.mlp.down_proj`	2229
`model.layers.6.mlp.gate_proj`	2229
`model.layers.6.mlp.up_proj`	2229
`model.layers.6.self_attn.k_proj`	573
`model.layers.6.self_attn.o_proj`	1433
`model.layers.6.self_attn.q_proj`	1433
`model.layers.6.self_attn.v_proj`	573
`model.layers.7.mlp.down_proj`	2229
`model.layers.7.mlp.gate_proj`	2229
`model.layers.7.mlp.up_proj`	2229
`model.layers.7.self_attn.k_proj`	573
`model.layers.7.self_attn.o_proj`	1433
`model.layers.7.self_attn.q_proj`	1433
`model.layers.7.self_attn.v_proj`	573
`model.layers.8.mlp.down_proj`	2229
`model.layers.8.mlp.gate_proj`	2229
`model.layers.8.mlp.up_proj`	2229
`model.layers.8.self_attn.k_proj`	573
`model.layers.8.self_attn.o_proj`	1433
`model.layers.8.self_attn.q_proj`	1433
`model.layers.8.self_attn.v_proj`	573
`model.layers.9.mlp.down_proj`	2229
`model.layers.9.mlp.gate_proj`	2229
`model.layers.9.mlp.up_proj`	2229
`model.layers.9.self_attn.k_proj`	573
`model.layers.9.self_attn.o_proj`	1433
`model.layers.9.self_attn.q_proj`	1433
`model.layers.9.self_attn.v_proj`	573

Hydra Configuration Summary

Config Field	Value
Model	`unsloth/mistral-7b`
SVD Method	`kfac_svd`
Search Method	`uniform`
Compression Target	`0.7`
Target Metric	`params`
Calibration Dataset	`wikitext2`
Sequence Length	`2048`
Seed	`42`