Instructions to use MoritzMo123/kfac-svd_lems_llama-3-8b_0.8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MoritzMo123/kfac-svd_lems_llama-3-8b_0.8 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="MoritzMo123/kfac-svd_lems_llama-3-8b_0.8", trust_remote_code=True)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("MoritzMo123/kfac-svd_lems_llama-3-8b_0.8", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use MoritzMo123/kfac-svd_lems_llama-3-8b_0.8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MoritzMo123/kfac-svd_lems_llama-3-8b_0.8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MoritzMo123/kfac-svd_lems_llama-3-8b_0.8",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/MoritzMo123/kfac-svd_lems_llama-3-8b_0.8

SGLang

How to use MoritzMo123/kfac-svd_lems_llama-3-8b_0.8 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "MoritzMo123/kfac-svd_lems_llama-3-8b_0.8" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MoritzMo123/kfac-svd_lems_llama-3-8b_0.8",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "MoritzMo123/kfac-svd_lems_llama-3-8b_0.8" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MoritzMo123/kfac-svd_lems_llama-3-8b_0.8",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use MoritzMo123/kfac-svd_lems_llama-3-8b_0.8 with Docker Model Runner:
```
docker model run hf.co/MoritzMo123/kfac-svd_lems_llama-3-8b_0.8
```

unsloth/llama-3-8b — lems search — 0.8 target ratio

This model was compressed using kfac_svd with lems rank search starting from unsloth/llama-3-8b 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	`unsloth/llama-3-8b`
Method	`kfac_svd`
Search Method	`lems`
Target Ratio	0.8
Compression Metric	`params`
Recommended Dtype	`float16`
Compressed Layers	108
Total Parameters	6,633,918,572

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.

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "MoritzMo123/kfac-svd_lems_llama-3-8b_0.8",
    trust_remote_code=True,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("MoritzMo123/kfac-svd_lems_llama-3-8b_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	8.04
ptb	25.14
c4	17.66

Rank Allocation

Per-layer ranks (click to expand)

Layer	Rank
`model.layers.0.self_attn.k_proj`	272
`model.layers.0.self_attn.o_proj`	1384
`model.layers.0.self_attn.q_proj`	616
`model.layers.1.self_attn.k_proj`	736
`model.layers.1.self_attn.q_proj`	616
`model.layers.10.mlp.down_proj`	1800
`model.layers.10.mlp.gate_proj`	1448
`model.layers.10.mlp.up_proj`	1936
`model.layers.10.self_attn.k_proj`	704
`model.layers.10.self_attn.o_proj`	1216
`model.layers.10.self_attn.q_proj`	656
`model.layers.11.mlp.down_proj`	1728
`model.layers.11.mlp.gate_proj`	1320
`model.layers.11.mlp.up_proj`	1792
`model.layers.11.self_attn.o_proj`	1160
`model.layers.11.self_attn.q_proj`	760
`model.layers.12.mlp.down_proj`	2000
`model.layers.12.mlp.gate_proj`	1368
`model.layers.12.mlp.up_proj`	1960
`model.layers.12.self_attn.o_proj`	1384
`model.layers.12.self_attn.q_proj`	616
`model.layers.13.mlp.down_proj`	2128
`model.layers.13.mlp.gate_proj`	1368
`model.layers.13.mlp.up_proj`	2168
`model.layers.13.self_attn.q_proj`	696
`model.layers.14.mlp.gate_proj`	1728
`model.layers.14.self_attn.o_proj`	1392
`model.layers.14.self_attn.q_proj`	728
`model.layers.15.mlp.gate_proj`	1840
`model.layers.15.self_attn.q_proj`	656
`model.layers.16.mlp.gate_proj`	2064
`model.layers.16.self_attn.k_proj`	640
`model.layers.16.self_attn.o_proj`	1320
`model.layers.16.self_attn.q_proj`	640
`model.layers.17.self_attn.o_proj`	1160
`model.layers.17.self_attn.q_proj`	624
`model.layers.18.self_attn.o_proj`	1160
`model.layers.18.self_attn.q_proj`	616
`model.layers.19.self_attn.k_proj`	608
`model.layers.19.self_attn.o_proj`	776
`model.layers.19.self_attn.q_proj`	616
`model.layers.2.self_attn.o_proj`	1096
`model.layers.2.self_attn.q_proj`	816
`model.layers.20.self_attn.k_proj`	408
`model.layers.20.self_attn.o_proj`	616
`model.layers.20.self_attn.q_proj`	616
`model.layers.21.self_attn.k_proj`	648
`model.layers.21.self_attn.o_proj`	632
`model.layers.21.self_attn.q_proj`	616
`model.layers.22.mlp.gate_proj`	2184
`model.layers.22.self_attn.k_proj`	464
`model.layers.22.self_attn.o_proj`	696
`model.layers.22.self_attn.q_proj`	616
`model.layers.23.mlp.gate_proj`	2136
`model.layers.23.self_attn.k_proj`	336
`model.layers.23.self_attn.o_proj`	688
`model.layers.23.self_attn.q_proj`	616
`model.layers.24.mlp.gate_proj`	2152
`model.layers.24.mlp.up_proj`	2024
`model.layers.24.self_attn.k_proj`	424
`model.layers.24.self_attn.o_proj`	632
`model.layers.24.self_attn.q_proj`	616
`model.layers.25.mlp.gate_proj`	2120
`model.layers.25.mlp.up_proj`	2032
`model.layers.25.self_attn.k_proj`	392
`model.layers.25.self_attn.o_proj`	664
`model.layers.25.self_attn.q_proj`	624
`model.layers.26.mlp.down_proj`	2104
`model.layers.26.mlp.up_proj`	1928
`model.layers.26.self_attn.k_proj`	376
`model.layers.26.self_attn.o_proj`	616
`model.layers.26.self_attn.q_proj`	632
`model.layers.27.mlp.down_proj`	1912
`model.layers.27.self_attn.k_proj`	440
`model.layers.27.self_attn.o_proj`	632
`model.layers.27.self_attn.q_proj`	632
`model.layers.28.mlp.down_proj`	2008
`model.layers.28.mlp.up_proj`	2040
`model.layers.28.self_attn.k_proj`	600
`model.layers.28.self_attn.o_proj`	648
`model.layers.28.self_attn.q_proj`	616
`model.layers.29.mlp.up_proj`	1992
`model.layers.29.self_attn.o_proj`	616
`model.layers.29.self_attn.q_proj`	616
`model.layers.3.self_attn.q_proj`	792
`model.layers.30.self_attn.k_proj`	336
`model.layers.30.self_attn.o_proj`	648
`model.layers.30.self_attn.q_proj`	616
`model.layers.31.self_attn.k_proj`	408
`model.layers.31.self_attn.o_proj`	792
`model.layers.31.self_attn.q_proj`	616
`model.layers.4.self_attn.o_proj`	1160
`model.layers.4.self_attn.q_proj`	768
`model.layers.5.self_attn.o_proj`	1064
`model.layers.5.self_attn.q_proj`	680
`model.layers.6.self_attn.o_proj`	1336
`model.layers.6.self_attn.q_proj`	680
`model.layers.7.self_attn.o_proj`	1352
`model.layers.7.self_attn.q_proj`	640
`model.layers.8.mlp.down_proj`	2016
`model.layers.8.mlp.gate_proj`	1584
`model.layers.8.self_attn.o_proj`	1168
`model.layers.8.self_attn.q_proj`	808
`model.layers.9.mlp.down_proj`	2016
`model.layers.9.mlp.gate_proj`	1680
`model.layers.9.mlp.up_proj`	2184
`model.layers.9.self_attn.o_proj`	1288
`model.layers.9.self_attn.q_proj`	664

Hydra Configuration Summary

Config Field	Value
Model	`unsloth/llama-3-8b`
SVD Method	`kfac_svd`
Search Method	`lems`
Compression Target	`0.8`
Target Metric	`params`
Calibration Dataset	`wikitext2`
Sequence Length	`2048`
Seed	`42`