中文Sign up

n-Arno/FLUX.2-klein-base-9B-GGUF

🤗 Hugging Face sourceapache-2.09B activated6.0 GBGGUF✓ 1 checksumupdated today
Submit in one command

Run it next to your model folder. It makes the torrent, checks your files against Hugging Face, and submits it. You just start seeding and paste your key from your account. It only reads your files and never changes them. Read the script first if you like.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo n-Arno/FLUX.2-klein-base-9B-GGUF ./model-folder
Needs a seeder →

Quantized version of black-forest-labs/FLUX.2-klein-9B made using sd.cpp.

Brain-dead mixed tensor selection of quantization based on the amount of element of the associated tensor:

Finding the distribution of layer profile:

from safetensors.torch import load_file
k = load_file("flux-2-klein-9b.safetensors")
keys = list(k.keys())

distrib = []
for key in keys:
  n = k[key].nelement()
  if n not in distrib:
    distrib.append(n)

distrib.sort()

print(distrib)

Making the CLI parameter for sd-cli.exe:

repart = {}
repart["524288"] = "q4_K"
repart["1048576"] = "q4_K"
repart["16777216"] = "q3_K"
repart["33554432"] = "q3_K"
repart["50331648"] = "q3_K"
repart["100663296"] = "q2_K"
repart["150994944"] = "q2_K"

tensor_type_rules = []
for key in keys:
  n = str(k[key].nelement())
  if n in repart.keys():
    tensor_type_rules.append(f"{key}={repart[n]}")

print(",".join(tensor_type_rules))

Resulting command:

sd-cli.exe --mode convert --model "flux-2-klein-9b.safetensors" --output "flux-2-klein-9b-Q3_K_M.gguf" --tensor-type-rules "double_blocks.0.img_attn.proj.weight=q3_K,double_blocks.0.img_attn.qkv.weight=q3_K,double_blocks.0.img_mlp.0.weight=q2_K,double_blocks.0.img_mlp.2.weight=q3_K,double_blocks.0.txt_attn.proj.weight=q3_K,double_blocks.0.txt_attn.qkv.weight=q3_K,double_blocks.0.txt_mlp.0.weight=q2_K,double_blocks.0.txt_mlp.2.weight=q3_K,double_blocks.1.img_attn.proj.weight=q3_K,double_blocks.1.img_attn.qkv.weight=q3_K,double_blocks.1.img_mlp.0.weight=q2_K,double_blocks.1.img_mlp.2.weight=q3_K,double_blocks.1.txt_attn.proj.weight=q3_K,double_blocks.1.txt_attn.qkv.weight=q3_K,double_blocks.1.txt_mlp.0.weight=q2_K,double_blocks.1.txt_mlp.2.weight=q3_K,double_blocks.2.img_attn.proj.weight=q3_K,double_blocks.2.img_attn.qkv.weight=q3_K,double_blocks.2.img_mlp.0.weight=q2_K,double_blocks.2.img_mlp.2.weight=q3_K,double_blocks.2.txt_attn.proj.weight=q3_K,double_blocks.2.txt_attn.qkv.weight=q3_K,double_blocks.2.txt_mlp.0.weight=q2_K,double_blocks.2.txt_mlp.2.weight=q3_K,double_blocks.3.img_attn.proj.weight=q3_K,double_blocks.3.img_attn.qkv.weight=q3_K,double_blocks.3.img_mlp.0.weight=q2_K,double_blocks.3.img_mlp.2.weight=q3_K,double_blocks.3.txt_attn.proj.weight=q3_K,double_blocks.3.txt_attn.qkv.weight=q3_K,double_blocks.3.txt_mlp.0.weight=q2_K,double_blocks.3.txt_mlp.2.weight=q3_K,double_blocks.4.img_attn.proj.weight=q3_K,double_blocks.4.img_attn.qkv.weight=q3_K,double_blocks.4.img_mlp.0.weight=q2_K,double_blocks.4.img_mlp.2.weight=q3_K,double_blocks.4.txt_attn.proj.weight=q3_K,double_blocks.4.txt_attn.qkv.weight=q3_K,double_blocks.4.txt_mlp.0.weight=q2_K,double_blocks.4.txt_mlp.2.weight=q3_K,double_blocks.5.img_attn.proj.weight=q3_K,double_blocks.5.img_attn.qkv.weight=q3_K,double_blocks.5.img_mlp.0.weight=q2_K,double_blocks.5.img_mlp.2.weight=q3_K,double_blocks.5.txt_attn.proj.weight=q3_K,double_blocks.5.txt_attn.qkv.weight=q3_K,double_blocks.5.txt_mlp.0.weight=q2_K,double_blocks.5.txt_mlp.2.weight=q3_K,double_blocks.6.img_attn.proj.weight=q3_K,double_blocks.6.img_attn.qkv.weight=q3_K,double_blocks.6.img_mlp.0.weight=q2_K,double_blocks.6.img_mlp.2.weight=q3_K,double_blocks.6.txt_attn.proj.weight=q3_K,double_blocks.6.txt_attn.qkv.weight=q3_K,double_blocks.6.txt_mlp.0.weight=q2_K,double_blocks.6.txt_mlp.2.weight=q3_K,double_blocks.7.img_attn.proj.weight=q3_K,double_blocks.7.img_attn.qkv.weight=q3_K,double_blocks.7.img_mlp.0.weight=q2_K,double_blocks.7.img_mlp.2.weight=q3_K,double_blocks.7.txt_attn.proj.weight=q3_K,double_blocks.7.txt_attn.qkv.weight=q3_K,double_blocks.7.txt_mlp.0.weight=q2_K,double_blocks.7.txt_mlp.2.weight=q3_K,double_stream_modulation_img.lin.weight=q2_K,double_stream_modulation_txt.lin.weight=q2_K,final_layer.adaLN_modulation.1.weight=q3_K,final_layer.linear.weight=q4_K,img_in.weight=q4_K,single_blocks.0.linear1.weight=q2_K,single_blocks.1.linear1.weight=q2_K,single_blocks.10.linear1.weight=q2_K,single_blocks.11.linear1.weight=q2_K,single_blocks.12.linear1.weight=q2_K,single_blocks.13.linear1.weight=q2_K,single_blocks.14.linear1.weight=q2_K,single_blocks.15.linear1.weight=q2_K,single_blocks.16.linear1.weight=q2_K,single_blocks.17.linear1.weight=q2_K,single_blocks.18.linear1.weight=q2_K,single_blocks.19.linear1.weight=q2_K,single_blocks.2.linear1.weight=q2_K,single_blocks.20.linear1.weight=q2_K,single_blocks.21.linear1.weight=q2_K,single_blocks.22.linear1.weight=q2_K,single_blocks.23.linear1.weight=q2_K,single_blocks.3.linear1.weight=q2_K,single_blocks.4.linear1.weight=q2_K,single_blocks.5.linear1.weight=q2_K,single_blocks.6.linear1.weight=q2_K,single_blocks.7.linear1.weight=q2_K,single_blocks.8.linear1.weight=q2_K,single_blocks.9.linear1.weight=q2_K,single_stream_modulation.lin.weight=q3_K,time_in.in_layer.weight=q4_K,time_in.out_layer.weight=q3_K,txt_in.weight=q3_K"