모종의 음모/ai 프로그램2026. 6. 1. 18:28

torchvision model

pytorch 껀진 모르겠는데 아무튼, 

torchvision에 model을 뒤지다 보니

mobilenet v3 관련 QAT 가 있어서 정리

 

[링크 : https://github.com/gdwang08/STFPM]

[링크 : https://github.com/gdwang08/STFPM/blob/main/main.py]

[링크 : https://docs.pytorch.org/vision/main/models/mobilenetv3_quant.html#model-builders]

[링크 : https://github.com/pytorch/vision/blob/main/torchvision/models/quantization/mobilenetv3.py]

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

nvidia tesla t4 16GB  (0) 2026.06.02
llama.cpp reasoning 옵션  (0) 2026.06.01
safetensors to gguf 일단 실패  (0) 2026.06.01
antigravity 2  (0) 2026.05.28
telegram bot api 로 기능 변경하기  (0) 2026.05.25
Posted by 구차니
모종의 음모/ai 프로그램2026. 6. 1. 13:52

safetensors to gguf 일단 실패

한번 해봐야지. 그나저나 1080 에서 F16만 된다고 하면.. 그냥 직접 변환하는것도 방법이 될려나?

/llama.cpp$ ls convert_*
convert_hf_to_gguf.py  convert_hf_to_gguf_update.py  convert_llama_ggml_to_gguf.py  convert_lora_to_gguf.py

 

[링크 : https://huggingface.co/google/gemma-4-E4B/tree/main]

[링크 : https://velog.io/@choonsik_mom/llama.cpp로-gguf-모델-서빙하기-ul02hone]

 

+

실행해보면 다음과 같이 나오고

$ /home/minimonk/src/llama.cpp/convert_hf_to_gguf.py 
usage: convert_hf_to_gguf.py [-h] [--vocab-only] [--outfile OUTFILE] [--outtype {f32,f16,bf16,q8_0,tq1_0,tq2_0,auto}]
                             [--bigendian] [--use-temp-file] [--no-lazy] [--model-name MODEL_NAME] [--verbose]
                             [--split-max-tensors SPLIT_MAX_TENSORS] [--split-max-size SPLIT_MAX_SIZE] [--dry-run]
                             [--no-tensor-first-split] [--metadata METADATA] [--print-supported-models] [--remote]
                             [--mmproj] [--mistral-format] [--disable-mistral-community-chat-template]
                             [--sentence-transformers-dense-modules] [--fuse-gate-up-exps]
                             [model]
convert_hf_to_gguf.py: error: the following arguments are required: model

 

huggingface를 통해서 받아도 되고, 최소한의 파일 저 3개를 받아줘도 된다.

/mnt/Downloads/model/gemma4-e4b/google$ ls
config.json  model.safetensors  tokenizer.json

 

GPU를 사용하지 않고, 한번에 메모리에 전부 올려서 하는건지 메모리 터져서 한번 리부팅.. 크흡

그 와중에 옵션 안주니까 BF16 으로 설정해버리는 센스(!) 하지만 난 BF16 가속을 못받는 하드웨어.. ㅠ

$ /home/minimonk/src/llama.cpp/convert_hf_to_gguf.py ./google/ --outfile gguf.gguf
INFO:hf-to-gguf:Loading model: google
INFO:hf-to-gguf:Model architecture: Gemma4ForConditionalGeneration
INFO:hf-to-gguf:gguf: indexing model part 'model.safetensors'
INFO:hf-to-gguf:heuristics detected bfloat16 tensor dtype, setting --outtype bf16
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:hf-to-gguf:Exporting model...

Writing:   9%|██████▎                                                                | 1.34G/15.0G [00:20<01:16, 180Mbyte/s]

Writing: 100%|███████████████████████████████████████████████████████████████████████| 15.0G/15.0G [01:54<00:00, 131Mbyte/s]
INFO:hf-to-gguf:Model successfully exported to gguf.gguf

 

exporting model.. 과 wrting 사이 로그

더보기

INFO:hf-to-gguf:rope_freqs.weight,                 torch.float32 --> F32, shape = {256}
INFO:hf-to-gguf:token_embd.weight,                 torch.bfloat16 --> BF16, shape = {2560, 262144}
INFO:hf-to-gguf:per_layer_token_embd.weight,       torch.bfloat16 --> BF16, shape = {10752, 262144}
INFO:hf-to-gguf:blk.0.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.0.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.0.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.0.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.0.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.0.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.0.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.0.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.0.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.0.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.0.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.0.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.1.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.1.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.1.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.1.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.1.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.1.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.1.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.1.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.1.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.1.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.1.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.1.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.10.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.10.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.10.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.10.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.10.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.10.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.10.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.10.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.10.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.10.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.10.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.10.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.11.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.11.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.11.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.11.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.11.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.11.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.11.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.11.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.11.attn_output.weight,         torch.bfloat16 --> BF16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.11.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.11.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.11.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.12.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.12.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.12.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.12.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.12.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.12.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.12.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.12.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.12.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.12.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.12.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.12.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.13.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.13.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.13.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.13.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.13.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.13.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.13.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.13.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.13.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.13.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.13.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.13.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.14.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.14.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.14.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.14.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.14.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.14.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.14.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.14.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.14.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.14.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.14.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.14.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.15.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.15.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.15.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.15.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.15.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.15.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.15.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.15.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.15.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.15.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.15.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.15.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.16.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.16.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.16.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.16.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.16.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.16.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.16.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.16.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.16.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.16.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.16.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.16.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.17.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.17.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.17.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.17.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.17.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.17.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.17.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.17.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.17.attn_output.weight,         torch.bfloat16 --> BF16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.17.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.17.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.17.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.18.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.18.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.18.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.18.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.18.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.18.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.18.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.18.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.18.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.18.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.18.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.18.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.19.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.19.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.19.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.19.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.19.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.19.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.19.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.19.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.19.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.19.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.19.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.19.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.2.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.2.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.2.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.2.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.2.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.2.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.2.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.2.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.2.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.2.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.2.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.2.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.20.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.20.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.20.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.20.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.20.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.20.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.20.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.20.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.20.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.20.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.20.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.20.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.21.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.21.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.21.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.21.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.21.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.21.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.21.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.21.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.21.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.21.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.21.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.21.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.22.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.22.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.22.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.22.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.22.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.22.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.22.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.22.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.22.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.22.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.22.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.22.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.23.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.23.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.23.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.23.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.23.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.23.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.23.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.23.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.23.attn_output.weight,         torch.bfloat16 --> BF16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.23.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.23.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.23.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.24.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.24.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.24.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.24.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.24.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.24.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.24.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.24.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.24.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.24.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.24.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.24.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.25.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.25.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.25.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.25.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.25.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.25.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.25.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.25.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.25.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.25.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.25.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.25.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.26.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.26.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.26.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.26.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.26.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.26.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.26.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.26.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.26.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.26.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.26.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.26.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.27.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.27.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.27.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.27.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.27.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.27.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.27.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.27.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.27.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.27.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.27.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.27.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.28.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.28.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.28.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.28.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.28.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.28.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.28.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.28.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.28.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.28.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.28.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.28.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.29.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.29.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.29.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.29.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.29.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.29.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.29.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.29.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.29.attn_output.weight,         torch.bfloat16 --> BF16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.29.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.29.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.29.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.3.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.3.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.3.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.3.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.3.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.3.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.3.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.3.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.3.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.3.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.3.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.3.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.30.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.30.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.30.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.30.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.30.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.30.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.30.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.30.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.30.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.30.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.30.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.30.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.31.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.31.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.31.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.31.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.31.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.31.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.31.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.31.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.31.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.31.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.31.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.31.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.32.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.32.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.32.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.32.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.32.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.32.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.32.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.32.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.32.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.32.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.32.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.32.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.33.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.33.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.33.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.33.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.33.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.33.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.33.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.33.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.33.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.33.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.33.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.33.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.34.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.34.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.34.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.34.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.34.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.34.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.34.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.34.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.34.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.34.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.34.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.34.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.35.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.35.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.35.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.35.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.35.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.35.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.35.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.35.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.35.attn_output.weight,         torch.bfloat16 --> BF16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.35.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.35.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.35.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.36.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.36.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.36.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.36.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.36.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.36.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.36.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.36.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.36.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.36.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.36.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.36.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.37.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.37.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.37.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.37.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.37.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.37.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.37.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.37.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.37.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.37.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.37.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.37.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.38.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.38.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.38.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.38.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.38.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.38.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.38.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.38.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.38.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.38.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.38.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.38.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.39.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.39.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.39.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.39.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.39.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.39.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.39.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.39.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.39.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.39.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.39.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.39.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.4.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.4.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.4.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.4.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.4.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.4.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.4.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.4.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.4.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.4.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.4.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.4.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.40.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.40.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.40.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.40.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.40.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.40.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.40.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.40.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.40.attn_output.weight,         torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.40.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.40.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.40.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.41.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.41.ffn_down.weight,            torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.41.ffn_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.41.ffn_up.weight,              torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.41.inp_gate.weight,            torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.41.proj.weight,                torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.41.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.41.attn_k.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.41.attn_output.weight,         torch.bfloat16 --> BF16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.41.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.41.attn_q.weight,              torch.bfloat16 --> BF16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.41.attn_v.weight,              torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.5.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.5.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.5.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.5.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.5.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.5.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.5.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.5.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.5.attn_output.weight,          torch.bfloat16 --> BF16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.5.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.5.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.5.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.6.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.6.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.6.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.6.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.6.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.6.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.6.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.6.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.6.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.6.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.6.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.6.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.7.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.7.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.7.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.7.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.7.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.7.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.7.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.7.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.7.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.7.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.7.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.7.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.8.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.8.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.8.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.8.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.8.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.8.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.8.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.8.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.8.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.8.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.8.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.8.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.9.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.9.ffn_down.weight,             torch.bfloat16 --> BF16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.9.ffn_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.9.ffn_up.weight,               torch.bfloat16 --> BF16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.9.inp_gate.weight,             torch.bfloat16 --> BF16, shape = {2560, 256}
INFO:hf-to-gguf:blk.9.proj.weight,                 torch.bfloat16 --> BF16, shape = {256, 2560}
INFO:hf-to-gguf:blk.9.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.9.attn_k.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:blk.9.attn_output.weight,          torch.bfloat16 --> BF16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.9.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.9.attn_q.weight,               torch.bfloat16 --> BF16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.9.attn_v.weight,               torch.bfloat16 --> BF16, shape = {2560, 512}
INFO:hf-to-gguf:output_norm.weight,                torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:per_layer_model_proj.weight,       torch.bfloat16 --> BF16, shape = {2560, 10752}
INFO:hf-to-gguf:per_layer_proj_norm.weight,        torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:Set meta model
INFO:hf-to-gguf:Set model parameters
INFO:hf-to-gguf:gguf: context length = 131072
INFO:hf-to-gguf:gguf: embedding length = 2560
INFO:hf-to-gguf:gguf: feed forward length = 10240
INFO:hf-to-gguf:gguf: head count = 8
INFO:hf-to-gguf:gguf: key-value head count = 2
WARNING:hf-to-gguf:Unknown RoPE type: proportional
INFO:hf-to-gguf:gguf: rope scaling type = NONE
INFO:hf-to-gguf:gguf: rope theta = 1000000.0
INFO:hf-to-gguf:gguf: rope theta swa = 10000.0
INFO:hf-to-gguf:gguf: rms norm epsilon = 1e-06
INFO:hf-to-gguf:gguf: file type = 32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.context_length', overwriting it with new value 131072 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.head_count', overwriting it with new value 8 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.layer_norm_rms_epsilon', overwriting it with new value 1e-06 of type FLOAT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.key_length', overwriting it with new value 256 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.value_length', overwriting it with new value 256 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.rope.freq_base', overwriting it with new value 1000000.0 of type FLOAT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.head_count_kv', overwriting it with new value 2 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.key_length', overwriting it with new value 512 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.value_length', overwriting it with new value 512 of type UINT32
INFO:hf-to-gguf:Set model quantization version
INFO:hf-to-gguf:Set model tokenizer
INFO:hf-to-gguf:Token '<|tool_call>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<tool_call|>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<|tool_response>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<tool_response|>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<|"|>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<|channel>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<channel|>' is set to USER_DEFINED
WARNING:gguf.vocab:Unknown separator token '<bos>' in TemplateProcessing<pair>
INFO:gguf.vocab:Adding 514906 merge(s).
INFO:gguf.vocab:Setting special token type bos to 2
INFO:gguf.vocab:Setting special token type eos to 1
INFO:gguf.vocab:Setting special token type pad to 0
INFO:gguf.vocab:Setting add_bos_token to True
INFO:gguf.vocab:Setting add_sep_token to False
WARNING:gguf.gguf_writer:Duplicated key name 'tokenizer.ggml.add_bos_token', overwriting it with new value True of type BOOL
INFO:gguf.gguf_writer:Writing the following files:
INFO:gguf.gguf_writer:gguf.gguf: n_tensors = 720, total_size = 15.0G

 

둘 다 머.. 거의 동일한 사이즈로 나왔다. 1기가도 안줄었네

drwxrwxr-x 2 minimonk minimonk        4096  6월  1 13:52 ./
drwxrwxr-x 3 minimonk minimonk        4096  6월  1 13:52 ../
-rw-rw-r-- 1 minimonk minimonk        5105  6월  1 13:51 config.json
-rw-rw-r-- 1 minimonk minimonk 15992595884  6월  1 13:03 model.safetensors
-rw-rw-r-- 1 minimonk minimonk    32170070  6월  1 13:51 tokenizer.json
-rw-rw-r-- 1 minimonk minimonk 15053078208  6월  1 14:01 gguf.gguf

 

그래서 F16 으로 하게 했더니  BF16 이나 F16 이나 용량은 바이트 단위 까지 동일!

-rw-rw-r-- 1 minimonk minimonk 15053078208  6월  1 14:05 gguf.gguf

 

일단 옵션은 outtype 으로 주면된다.

$ /home/minimonk/src/llama.cpp/convert_hf_to_gguf.py ./google/ --outfile gguf.gguf --outtype f16
INFO:hf-to-gguf:Loading model: google
INFO:hf-to-gguf:Model architecture: Gemma4ForConditionalGeneration
INFO:hf-to-gguf:gguf: indexing model part 'model.safetensors'
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:hf-to-gguf:Exporting model...

Writing:   9%|██████▎                                                                | 1.34G/15.0G [00:02<00:27, 503Mbyte/s]

Writing: 100%|███████████████████████████████████████████████████████████████████████| 15.0G/15.0G [00:49<00:00, 305Mbyte/s]
INFO:hf-to-gguf:Model successfully exported to gguf.gguf

 

exporting model.. 과 wrting 사이 로그

더보기

INFO:hf-to-gguf:rope_freqs.weight,                 torch.float32 --> F32, shape = {256}
INFO:hf-to-gguf:token_embd.weight,                 torch.bfloat16 --> F16, shape = {2560, 262144}
INFO:hf-to-gguf:per_layer_token_embd.weight,       torch.bfloat16 --> F16, shape = {10752, 262144}
INFO:hf-to-gguf:blk.0.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.0.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.0.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.0.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.0.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.0.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.0.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.0.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.0.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.0.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.0.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.0.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.0.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.1.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.1.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.1.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.1.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.1.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.1.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.1.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.1.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.1.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.1.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.1.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.1.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.1.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.10.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.10.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.10.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.10.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.10.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.10.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.10.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.10.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.10.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.10.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.10.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.10.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.10.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.11.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.11.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.11.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.11.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.11.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.11.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.11.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.11.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.11.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.11.attn_output.weight,         torch.bfloat16 --> F16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.11.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.11.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.11.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.12.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.12.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.12.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.12.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.12.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.12.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.12.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.12.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.12.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.12.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.12.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.12.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.12.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.13.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.13.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.13.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.13.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.13.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.13.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.13.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.13.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.13.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.13.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.13.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.13.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.13.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.14.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.14.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.14.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.14.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.14.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.14.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.14.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.14.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.14.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.14.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.14.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.14.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.14.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.15.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.15.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.15.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.15.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.15.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.15.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.15.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.15.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.15.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.15.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.15.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.15.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.15.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.16.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.16.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.16.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.16.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.16.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.16.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.16.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.16.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.16.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.16.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.16.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.16.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.16.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.17.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.17.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.17.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.17.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.17.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.17.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.17.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.17.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.17.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.17.attn_output.weight,         torch.bfloat16 --> F16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.17.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.17.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.17.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.18.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.18.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.18.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.18.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.18.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.18.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.18.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.18.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.18.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.18.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.18.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.18.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.18.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.19.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.19.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.19.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.19.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.19.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.19.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.19.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.19.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.19.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.19.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.19.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.19.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.19.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.2.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.2.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.2.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.2.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.2.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.2.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.2.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.2.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.2.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.2.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.2.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.2.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.2.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.20.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.20.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.20.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.20.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.20.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.20.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.20.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.20.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.20.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.20.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.20.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.20.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.20.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.21.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.21.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.21.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.21.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.21.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.21.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.21.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.21.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.21.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.21.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.21.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.21.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.21.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.22.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.22.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.22.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.22.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.22.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.22.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.22.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.22.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.22.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.22.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.22.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.22.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.22.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.23.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.23.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.23.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.23.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.23.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.23.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.23.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.23.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.23.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.23.attn_output.weight,         torch.bfloat16 --> F16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.23.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.23.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.23.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.24.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.24.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.24.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.24.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.24.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.24.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.24.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.24.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.24.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.24.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.24.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.24.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.24.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.25.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.25.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.25.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.25.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.25.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.25.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.25.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.25.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.25.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.25.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.25.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.25.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.25.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.26.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.26.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.26.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.26.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.26.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.26.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.26.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.26.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.26.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.26.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.26.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.26.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.26.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.27.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.27.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.27.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.27.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.27.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.27.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.27.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.27.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.27.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.27.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.27.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.27.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.27.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.28.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.28.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.28.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.28.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.28.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.28.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.28.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.28.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.28.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.28.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.28.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.28.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.28.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.29.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.29.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.29.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.29.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.29.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.29.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.29.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.29.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.29.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.29.attn_output.weight,         torch.bfloat16 --> F16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.29.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.29.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.29.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.3.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.3.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.3.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.3.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.3.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.3.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.3.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.3.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.3.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.3.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.3.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.3.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.3.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.30.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.30.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.30.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.30.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.30.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.30.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.30.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.30.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.30.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.30.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.30.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.30.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.30.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.31.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.31.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.31.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.31.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.31.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.31.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.31.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.31.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.31.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.31.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.31.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.31.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.31.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.32.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.32.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.32.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.32.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.32.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.32.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.32.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.32.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.32.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.32.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.32.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.32.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.32.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.33.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.33.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.33.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.33.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.33.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.33.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.33.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.33.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.33.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.33.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.33.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.33.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.33.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.34.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.34.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.34.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.34.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.34.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.34.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.34.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.34.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.34.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.34.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.34.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.34.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.34.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.35.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.35.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.35.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.35.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.35.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.35.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.35.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.35.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.35.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.35.attn_output.weight,         torch.bfloat16 --> F16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.35.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.35.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.35.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.36.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.36.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.36.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.36.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.36.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.36.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.36.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.36.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.36.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.36.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.36.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.36.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.36.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.37.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.37.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.37.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.37.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.37.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.37.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.37.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.37.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.37.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.37.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.37.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.37.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.37.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.38.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.38.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.38.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.38.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.38.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.38.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.38.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.38.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.38.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.38.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.38.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.38.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.38.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.39.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.39.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.39.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.39.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.39.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.39.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.39.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.39.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.39.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.39.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.39.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.39.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.39.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.4.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.4.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.4.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.4.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.4.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.4.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.4.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.4.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.4.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.4.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.4.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.4.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.4.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.40.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.40.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.40.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.40.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.40.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.40.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.40.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.40.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.40.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.40.attn_output.weight,         torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.40.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.40.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.40.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.41.attn_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.layer_output_scale.weight,  torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.41.ffn_down.weight,            torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.41.ffn_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.41.ffn_up.weight,              torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.41.inp_gate.weight,            torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.41.proj.weight,                torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.41.post_attention_norm.weight, torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.post_ffw_norm.weight,       torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.post_norm.weight,           torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.ffn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.41.attn_k_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.41.attn_k.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.41.attn_output.weight,         torch.bfloat16 --> F16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.41.attn_q_norm.weight,         torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.41.attn_q.weight,              torch.bfloat16 --> F16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.41.attn_v.weight,              torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.5.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.5.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.5.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.5.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.5.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.5.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.5.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.5.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.5.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.5.attn_output.weight,          torch.bfloat16 --> F16, shape = {4096, 2560}
INFO:hf-to-gguf:blk.5.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {512}
INFO:hf-to-gguf:blk.5.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 4096}
INFO:hf-to-gguf:blk.5.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 1024}
INFO:hf-to-gguf:blk.6.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.6.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.6.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.6.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.6.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.6.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.6.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.6.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.6.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.6.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.6.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.6.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.6.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.7.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.7.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.7.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.7.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.7.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.7.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.7.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.7.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.7.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.7.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.7.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.7.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.7.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.8.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.8.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.8.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.8.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.8.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.8.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.8.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.8.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.8.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.8.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.8.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.8.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.8.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.9.attn_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.layer_output_scale.weight,   torch.bfloat16 --> F32, shape = {1}
INFO:hf-to-gguf:blk.9.ffn_down.weight,             torch.bfloat16 --> F16, shape = {10240, 2560}
INFO:hf-to-gguf:blk.9.ffn_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.9.ffn_up.weight,               torch.bfloat16 --> F16, shape = {2560, 10240}
INFO:hf-to-gguf:blk.9.inp_gate.weight,             torch.bfloat16 --> F16, shape = {2560, 256}
INFO:hf-to-gguf:blk.9.proj.weight,                 torch.bfloat16 --> F16, shape = {256, 2560}
INFO:hf-to-gguf:blk.9.post_attention_norm.weight,  torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.post_ffw_norm.weight,        torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.post_norm.weight,            torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.ffn_norm.weight,             torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:blk.9.attn_k_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.9.attn_k.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:blk.9.attn_output.weight,          torch.bfloat16 --> F16, shape = {2048, 2560}
INFO:hf-to-gguf:blk.9.attn_q_norm.weight,          torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:blk.9.attn_q.weight,               torch.bfloat16 --> F16, shape = {2560, 2048}
INFO:hf-to-gguf:blk.9.attn_v.weight,               torch.bfloat16 --> F16, shape = {2560, 512}
INFO:hf-to-gguf:output_norm.weight,                torch.bfloat16 --> F32, shape = {2560}
INFO:hf-to-gguf:per_layer_model_proj.weight,       torch.bfloat16 --> F16, shape = {2560, 10752}
INFO:hf-to-gguf:per_layer_proj_norm.weight,        torch.bfloat16 --> F32, shape = {256}
INFO:hf-to-gguf:Set meta model
INFO:hf-to-gguf:Set model parameters
INFO:hf-to-gguf:gguf: context length = 131072
INFO:hf-to-gguf:gguf: embedding length = 2560
INFO:hf-to-gguf:gguf: feed forward length = 10240
INFO:hf-to-gguf:gguf: head count = 8
INFO:hf-to-gguf:gguf: key-value head count = 2
WARNING:hf-to-gguf:Unknown RoPE type: proportional
INFO:hf-to-gguf:gguf: rope scaling type = NONE
INFO:hf-to-gguf:gguf: rope theta = 1000000.0
INFO:hf-to-gguf:gguf: rope theta swa = 10000.0
INFO:hf-to-gguf:gguf: rms norm epsilon = 1e-06
INFO:hf-to-gguf:gguf: file type = 1
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.context_length', overwriting it with new value 131072 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.head_count', overwriting it with new value 8 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.layer_norm_rms_epsilon', overwriting it with new value 1e-06 of type FLOAT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.key_length', overwriting it with new value 256 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.value_length', overwriting it with new value 256 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.rope.freq_base', overwriting it with new value 1000000.0 of type FLOAT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.head_count_kv', overwriting it with new value 2 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.key_length', overwriting it with new value 512 of type UINT32
WARNING:gguf.gguf_writer:Duplicated key name 'gemma4.attention.value_length', overwriting it with new value 512 of type UINT32
INFO:hf-to-gguf:Set model quantization version
INFO:hf-to-gguf:Set model tokenizer
INFO:hf-to-gguf:Token '<|tool_call>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<tool_call|>' is set to USER_DEFINED</tool_call|>
INFO:hf-to-gguf:Token '<|tool_response>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<tool_response|>' is set to USER_DEFINED</tool_response|>
INFO:hf-to-gguf:Token '<|"|>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<|channel>' is set to USER_DEFINED
INFO:hf-to-gguf:Token '<channel|>' is set to USER_DEFINED</channel|>
WARNING:gguf.vocab:Unknown separator token '' in TemplateProcessing
INFO:gguf.vocab:Adding 514906 merge(s).
INFO:gguf.vocab:Setting special token type bos to 2
INFO:gguf.vocab:Setting special token type eos to 1
INFO:gguf.vocab:Setting special token type pad to 0
INFO:gguf.vocab:Setting add_bos_token to True
INFO:gguf.vocab:Setting add_sep_token to False
WARNING:gguf.gguf_writer:Duplicated key name 'tokenizer.ggml.add_bos_token', overwriting it with new value True of type BOOL
INFO:gguf.gguf_writer:Writing the following files:
INFO:gguf.gguf_writer:gguf.gguf: n_tensors = 720, total_size = 15.0G

 

양자화 F16 으로 했더니 14GB 정도 합산해서 먹고 있고

 

llama.cpp로 해보니 애가 오작동한다. 왜 저러니 ㅠㅠ

 

+

$ /mnt/Downloads/llama-b9305/llama-quantize 
usage: /mnt/Downloads/llama-b9305/llama-quantize [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights]
       [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--tensor-type-file]
       [--prune-layers] [--keep-split] [--override-kv] [--dry-run]
       model-f32.gguf [model-quant.gguf] type [nthreads]

  --allow-requantize
                                      allow requantizing tensors that have already been quantized
                                      WARNING: this can severely reduce quality compared to quantizing
                                               from 16bit or 32bit!
  --leave-output-tensor
                                      leave output.weight un(re)quantized
                                      increases model size but may also increase quality, especially when requantizing
  --pure
                                      disable k-quant mixtures and quantize all tensors to the same type
  --imatrix file_name
                                      use data in file_name as importance matrix for quant optimizations
  --include-weights tensor_name
                                      use importance matrix for this/these tensor(s)
  --exclude-weights tensor_name
                                      do not use importance matrix for this/these tensor(s)
  --output-tensor-type ggml_type
                                      use this ggml_type for the output.weight tensor
  --token-embedding-type ggml_type
                                      use this ggml_type for the token embeddings tensor
  --tensor-type tensor_name=ggml_type
                                      quantize this tensor to this ggml_type
                                      this is an advanced option to selectively quantize tensors. may be specified multiple times.
                                      example: --tensor-type attn_q=q8_0
  --tensor-type-file tensor_types.txt
                                      list of tensors to quantize to a specific ggml_type
                                      this is an advanced option to selectively quantize a long list of tensors.
                                      the file should use the same format as above, separated by spaces or newlines.
  --prune-layers L0,L1,L2...
                                      comma-separated list of layer numbers to prune from the model
                                      WARNING: this is an advanced option, use with care.
  --keep-split
                                      generate quantized model in the same shards as input
  --override-kv KEY=TYPE:VALUE
                                      override model metadata by key in the quantized model. may be specified multiple times.
                                      WARNING: this is an advanced option, use with care.
  --dry-run
                                      calculate and show the final quantization size without performing quantization
                                      example: llama-quantize --dry-run model-f32.gguf Q4_K

note: --include-weights and --exclude-weights cannot be used together

-----------------------------------------------------------------------------
 allowed quantization types
-----------------------------------------------------------------------------

  40  or  Q1_0    :  1.125 bpw quantization
   2  or  Q4_0    :  4.34G, +0.4685 ppl @ Llama-3-8B
   3  or  Q4_1    :  4.78G, +0.4511 ppl @ Llama-3-8B
  38  or  MXFP4_MOE :  MXFP4 MoE
   8  or  Q5_0    :  5.21G, +0.1316 ppl @ Llama-3-8B
   9  or  Q5_1    :  5.65G, +0.1062 ppl @ Llama-3-8B
  19  or  IQ2_XXS :  2.06 bpw quantization
  20  or  IQ2_XS  :  2.31 bpw quantization
  28  or  IQ2_S   :  2.5  bpw quantization
  29  or  IQ2_M   :  2.7  bpw quantization
  24  or  IQ1_S   :  1.56 bpw quantization
  31  or  IQ1_M   :  1.75 bpw quantization
  36  or  TQ1_0   :  1.69 bpw ternarization
  37  or  TQ2_0   :  2.06 bpw ternarization
  10  or  Q2_K    :  2.96G, +3.5199 ppl @ Llama-3-8B
  21  or  Q2_K_S  :  2.96G, +3.1836 ppl @ Llama-3-8B
  23  or  IQ3_XXS :  3.06 bpw quantization
  26  or  IQ3_S   :  3.44 bpw quantization
  27  or  IQ3_M   :  3.66 bpw quantization mix
  12  or  Q3_K    : alias for Q3_K_M
  22  or  IQ3_XS  :  3.3 bpw quantization
  11  or  Q3_K_S  :  3.41G, +1.6321 ppl @ Llama-3-8B
  12  or  Q3_K_M  :  3.74G, +0.6569 ppl @ Llama-3-8B
  13  or  Q3_K_L  :  4.03G, +0.5562 ppl @ Llama-3-8B
  25  or  IQ4_NL  :  4.50 bpw non-linear quantization
  30  or  IQ4_XS  :  4.25 bpw non-linear quantization
  15  or  Q4_K    : alias for Q4_K_M
  14  or  Q4_K_S  :  4.37G, +0.2689 ppl @ Llama-3-8B
  15  or  Q4_K_M  :  4.58G, +0.1754 ppl @ Llama-3-8B
  17  or  Q5_K    : alias for Q5_K_M
  16  or  Q5_K_S  :  5.21G, +0.1049 ppl @ Llama-3-8B
  17  or  Q5_K_M  :  5.33G, +0.0569 ppl @ Llama-3-8B
  18  or  Q6_K    :  6.14G, +0.0217 ppl @ Llama-3-8B
   7  or  Q8_0    :  7.96G, +0.0026 ppl @ Llama-3-8B
   1  or  F16     : 14.00G, +0.0020 ppl @ Mistral-7B
  32  or  BF16    : 14.00G, -0.0050 ppl @ Mistral-7B
   0  or  F32     : 26.00G              @ 7B
          COPY    : only copy tensors, no quantizing 

 

F16 에서 Q8로 바꾸니 용량이 절반으로!

-rw-rw-r-- 1 minimonk minimonk  4977169088  4월 29 10:43 gemma-4-E4B-it-Q4_K_M.gguf

-rw-rw-r-- 1 minimonk minimonk 15053078208  6월  1 14:24 gguf.gguf
-rw-rw-r-- 1 minimonk minimonk  5335273152  6월  1 14:32 Q4_K_M.gguf
-rw-rw-r-- 1 minimonk minimonk  8031223488  6월  1 14:27 Q8.gguf

 

$ /mnt/Downloads/llama-b9305/llama-quantize  ./gguf.gguf ./Q8.gguf Q8_0
load_backend: loaded RPC backend from /mnt/Downloads/llama-b9305/libggml-rpc.so
ggml_vulkan: Found 3 Vulkan devices:
ggml_vulkan: 0 = Intel(R) UHD Graphics 630 (CFL GT2) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 65536 | int dot: 0 | matrix cores: none
ggml_vulkan: 1 = NVIDIA GeForce GTX 1080 Ti (NVIDIA) | uma: 0 | fp16: 0 | bf16: 0 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
ggml_vulkan: 2 = NVIDIA GeForce GTX 1080 Ti (NVIDIA) | uma: 0 | fp16: 0 | bf16: 0 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from /mnt/Downloads/llama-b9305/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/Downloads/llama-b9305/libggml-cpu-haswell.so
llama_print_build_info: build = 9305 (63248fc3e)
llama_print_build_info: built with GNU 11.4.0 for Linux x86_64
llama_quantize: quantizing './gguf.gguf' to './Q8.gguf' as Q8_0
llama_model_loader: loaded meta data with 37 key-value pairs and 720 tensors from ./gguf.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = gemma4
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Google
llama_model_loader: - kv   3:                         general.size_label str              = 7.5B
llama_model_loader: - kv   4:                         gemma4.block_count u32              = 42
llama_model_loader: - kv   5:                      gemma4.context_length u32              = 131072
llama_model_loader: - kv   6:                    gemma4.embedding_length u32              = 2560
llama_model_loader: - kv   7:                 gemma4.feed_forward_length u32              = 10240
llama_model_loader: - kv   8:                gemma4.attention.head_count u32              = 8
llama_model_loader: - kv   9:             gemma4.attention.head_count_kv u32              = 2
llama_model_loader: - kv  10:                      gemma4.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  11:                  gemma4.rope.freq_base_swa f32              = 10000.000000
llama_model_loader: - kv  12:    gemma4.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                gemma4.attention.key_length u32              = 512
llama_model_loader: - kv  14:              gemma4.attention.value_length u32              = 512
llama_model_loader: - kv  15:                          general.file_type u32              = 1
llama_model_loader: - kv  16:             gemma4.final_logit_softcapping f32              = 30.000000
llama_model_loader: - kv  17:            gemma4.attention.sliding_window u32              = 512
llama_model_loader: - kv  18:          gemma4.attention.shared_kv_layers u32              = 18
llama_model_loader: - kv  19:    gemma4.embedding_length_per_layer_input u32              = 256
llama_model_loader: - kv  20:    gemma4.attention.sliding_window_pattern arr[bool,42]     = [true, true, true, true, true, false,...
llama_model_loader: - kv  21:            gemma4.attention.key_length_swa u32              = 256
llama_model_loader: - kv  22:          gemma4.attention.value_length_swa u32              = 256
llama_model_loader: - kv  23:                gemma4.rope.dimension_count u32              = 512
llama_model_loader: - kv  24:            gemma4.rope.dimension_count_swa u32              = 256
llama_model_loader: - kv  25:               general.quantization_version u32              = 2
llama_model_loader: - kv  26:                       tokenizer.ggml.model str              = gemma4
llama_model_loader: - kv  27:                      tokenizer.ggml.tokens arr[str,262144]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  28:                      tokenizer.ggml.scores arr[f32,262144]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  29:                  tokenizer.ggml.token_type arr[i32,262144]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  30:                      tokenizer.ggml.merges arr[str,514906]  = ["\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n \n", ...
llama_model_loader: - kv  31:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  32:                tokenizer.ggml.eos_token_id u32              = 1
llama_model_loader: - kv  33:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  34:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  35:               tokenizer.ggml.add_sep_token bool             = false
llama_model_loader: - kv  36:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - type  f32:  339 tensors
llama_model_loader: - type  f16:  381 tensors

/// 여기 생략됨 

llama_model_quantize_impl: model size  = 14340.66 MiB (16.00 BPW)
llama_model_quantize_impl: quant size  =  7644.10 MiB (8.53 BPW)

llama_quantize: quantize time = 99905.90 ms
llama_quantize:    total time = 99905.90 ms

 

더보기

[   1/ 720] output_norm.weight                   - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[   2/ 720] per_layer_model_proj.weight          - [  2560,  10752,      1,      1], type =    f16, size =   52.500 MiB
[   3/ 720] per_layer_proj_norm.weight           - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[   4/ 720] per_layer_token_embd.weight          - [ 10752, 262144,      1,      1], type =    f16, converting to q8_0 .. size =  5376.00 MiB ->  2856.00 MiB
[   5/ 720] rope_freqs.weight                    - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[   6/ 720] token_embd.weight                    - [  2560, 262144,      1,      1], type =    f16, converting to q8_0 .. size =  1280.00 MiB ->   680.00 MiB
[   7/ 720] blk.0.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[   8/ 720] blk.0.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[   9/ 720] blk.0.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  10/ 720] blk.0.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  11/ 720] blk.0.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  12/ 720] blk.0.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  13/ 720] blk.0.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  14/ 720] blk.0.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  15/ 720] blk.0.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  16/ 720] blk.0.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  17/ 720] blk.0.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  18/ 720] blk.0.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  19/ 720] blk.0.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[  20/ 720] blk.0.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  21/ 720] blk.0.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  22/ 720] blk.0.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  23/ 720] blk.0.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  24/ 720] blk.1.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  25/ 720] blk.1.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  26/ 720] blk.1.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  27/ 720] blk.1.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  28/ 720] blk.1.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  29/ 720] blk.1.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  30/ 720] blk.1.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  31/ 720] blk.1.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  32/ 720] blk.1.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  33/ 720] blk.1.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  34/ 720] blk.1.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  35/ 720] blk.1.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  36/ 720] blk.1.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[  37/ 720] blk.1.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  38/ 720] blk.1.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  39/ 720] blk.1.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  40/ 720] blk.1.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  41/ 720] blk.2.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  42/ 720] blk.2.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  43/ 720] blk.2.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  44/ 720] blk.2.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  45/ 720] blk.2.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  46/ 720] blk.2.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  47/ 720] blk.2.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  48/ 720] blk.2.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  49/ 720] blk.2.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  50/ 720] blk.2.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  51/ 720] blk.2.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  52/ 720] blk.2.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  53/ 720] blk.2.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[  54/ 720] blk.2.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  55/ 720] blk.2.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  56/ 720] blk.2.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  57/ 720] blk.2.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  58/ 720] blk.3.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  59/ 720] blk.3.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  60/ 720] blk.3.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  61/ 720] blk.3.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  62/ 720] blk.3.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  63/ 720] blk.3.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  64/ 720] blk.3.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  65/ 720] blk.3.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  66/ 720] blk.3.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  67/ 720] blk.3.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  68/ 720] blk.3.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  69/ 720] blk.3.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  70/ 720] blk.3.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[  71/ 720] blk.3.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  72/ 720] blk.3.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  73/ 720] blk.3.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  74/ 720] blk.3.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  75/ 720] blk.4.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  76/ 720] blk.4.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  77/ 720] blk.4.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  78/ 720] blk.4.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  79/ 720] blk.4.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[  80/ 720] blk.4.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[  81/ 720] blk.4.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[  82/ 720] blk.4.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  83/ 720] blk.4.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  84/ 720] blk.4.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  85/ 720] blk.4.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[  86/ 720] blk.4.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  87/ 720] blk.4.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[  88/ 720] blk.4.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  89/ 720] blk.4.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  90/ 720] blk.4.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  91/ 720] blk.4.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[  92/ 720] blk.5.attn_k.weight                  - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[  93/ 720] blk.5.attn_k_norm.weight             - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[  94/ 720] blk.5.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[  95/ 720] blk.5.attn_output.weight             - [  4096,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[  96/ 720] blk.5.attn_q.weight                  - [  2560,   4096,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[  97/ 720] blk.5.attn_q_norm.weight             - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[  98/ 720] blk.5.attn_v.weight                  - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[  99/ 720] blk.5.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 100/ 720] blk.5.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 101/ 720] blk.5.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 102/ 720] blk.5.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 103/ 720] blk.5.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 104/ 720] blk.5.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 105/ 720] blk.5.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 106/ 720] blk.5.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 107/ 720] blk.5.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 108/ 720] blk.5.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 109/ 720] blk.6.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 110/ 720] blk.6.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 111/ 720] blk.6.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 112/ 720] blk.6.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 113/ 720] blk.6.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 114/ 720] blk.6.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 115/ 720] blk.6.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 116/ 720] blk.6.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 117/ 720] blk.6.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 118/ 720] blk.6.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 119/ 720] blk.6.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 120/ 720] blk.6.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 121/ 720] blk.6.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 122/ 720] blk.6.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 123/ 720] blk.6.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 124/ 720] blk.6.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 125/ 720] blk.6.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 126/ 720] blk.7.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 127/ 720] blk.7.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 128/ 720] blk.7.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 129/ 720] blk.7.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 130/ 720] blk.7.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 131/ 720] blk.7.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 132/ 720] blk.7.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 133/ 720] blk.7.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 134/ 720] blk.7.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 135/ 720] blk.7.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 136/ 720] blk.7.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 137/ 720] blk.7.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 138/ 720] blk.7.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 139/ 720] blk.7.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 140/ 720] blk.7.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 141/ 720] blk.7.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 142/ 720] blk.7.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 143/ 720] blk.8.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 144/ 720] blk.8.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 145/ 720] blk.8.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 146/ 720] blk.8.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 147/ 720] blk.8.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 148/ 720] blk.8.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 149/ 720] blk.8.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 150/ 720] blk.8.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 151/ 720] blk.8.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 152/ 720] blk.8.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 153/ 720] blk.8.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 154/ 720] blk.8.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 155/ 720] blk.8.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 156/ 720] blk.8.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 157/ 720] blk.8.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 158/ 720] blk.8.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 159/ 720] blk.8.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 160/ 720] blk.9.attn_k.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 161/ 720] blk.9.attn_k_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 162/ 720] blk.9.attn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 163/ 720] blk.9.attn_output.weight             - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 164/ 720] blk.9.attn_q.weight                  - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 165/ 720] blk.9.attn_q_norm.weight             - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 166/ 720] blk.9.attn_v.weight                  - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 167/ 720] blk.9.ffn_down.weight                - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 168/ 720] blk.9.ffn_gate.weight                - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 169/ 720] blk.9.ffn_norm.weight                - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 170/ 720] blk.9.ffn_up.weight                  - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 171/ 720] blk.9.inp_gate.weight                - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 172/ 720] blk.9.layer_output_scale.weight      - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 173/ 720] blk.9.post_attention_norm.weight     - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 174/ 720] blk.9.post_ffw_norm.weight           - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 175/ 720] blk.9.post_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 176/ 720] blk.9.proj.weight                    - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 177/ 720] blk.10.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 178/ 720] blk.10.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 179/ 720] blk.10.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 180/ 720] blk.10.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 181/ 720] blk.10.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 182/ 720] blk.10.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 183/ 720] blk.10.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 184/ 720] blk.10.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 185/ 720] blk.10.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 186/ 720] blk.10.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 187/ 720] blk.10.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 188/ 720] blk.10.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 189/ 720] blk.10.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 190/ 720] blk.10.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 191/ 720] blk.10.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 192/ 720] blk.10.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 193/ 720] blk.10.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 194/ 720] blk.11.attn_k.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 195/ 720] blk.11.attn_k_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 196/ 720] blk.11.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 197/ 720] blk.11.attn_output.weight            - [  4096,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 198/ 720] blk.11.attn_q.weight                 - [  2560,   4096,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 199/ 720] blk.11.attn_q_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 200/ 720] blk.11.attn_v.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 201/ 720] blk.11.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 202/ 720] blk.11.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 203/ 720] blk.11.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 204/ 720] blk.11.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 205/ 720] blk.11.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 206/ 720] blk.11.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 207/ 720] blk.11.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 208/ 720] blk.11.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 209/ 720] blk.11.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 210/ 720] blk.11.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 211/ 720] blk.12.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 212/ 720] blk.12.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 213/ 720] blk.12.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 214/ 720] blk.12.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 215/ 720] blk.12.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 216/ 720] blk.12.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 217/ 720] blk.12.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 218/ 720] blk.12.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 219/ 720] blk.12.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 220/ 720] blk.12.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 221/ 720] blk.12.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 222/ 720] blk.12.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 223/ 720] blk.12.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 224/ 720] blk.12.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 225/ 720] blk.12.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 226/ 720] blk.12.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 227/ 720] blk.12.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 228/ 720] blk.13.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 229/ 720] blk.13.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 230/ 720] blk.13.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 231/ 720] blk.13.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 232/ 720] blk.13.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 233/ 720] blk.13.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 234/ 720] blk.13.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 235/ 720] blk.13.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 236/ 720] blk.13.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 237/ 720] blk.13.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 238/ 720] blk.13.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 239/ 720] blk.13.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 240/ 720] blk.13.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 241/ 720] blk.13.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 242/ 720] blk.13.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 243/ 720] blk.13.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 244/ 720] blk.13.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 245/ 720] blk.14.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 246/ 720] blk.14.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 247/ 720] blk.14.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 248/ 720] blk.14.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 249/ 720] blk.14.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 250/ 720] blk.14.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 251/ 720] blk.14.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 252/ 720] blk.14.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 253/ 720] blk.14.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 254/ 720] blk.14.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 255/ 720] blk.14.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 256/ 720] blk.14.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 257/ 720] blk.14.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 258/ 720] blk.14.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 259/ 720] blk.14.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 260/ 720] blk.14.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 261/ 720] blk.14.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 262/ 720] blk.15.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 263/ 720] blk.15.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 264/ 720] blk.15.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 265/ 720] blk.15.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 266/ 720] blk.15.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 267/ 720] blk.15.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 268/ 720] blk.15.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 269/ 720] blk.15.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 270/ 720] blk.15.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 271/ 720] blk.15.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 272/ 720] blk.15.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 273/ 720] blk.15.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 274/ 720] blk.15.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 275/ 720] blk.15.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 276/ 720] blk.15.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 277/ 720] blk.15.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 278/ 720] blk.15.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 279/ 720] blk.16.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 280/ 720] blk.16.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 281/ 720] blk.16.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 282/ 720] blk.16.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 283/ 720] blk.16.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 284/ 720] blk.16.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 285/ 720] blk.16.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 286/ 720] blk.16.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 287/ 720] blk.16.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 288/ 720] blk.16.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 289/ 720] blk.16.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 290/ 720] blk.16.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 291/ 720] blk.16.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 292/ 720] blk.16.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 293/ 720] blk.16.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 294/ 720] blk.16.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 295/ 720] blk.16.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 296/ 720] blk.17.attn_k.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 297/ 720] blk.17.attn_k_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 298/ 720] blk.17.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 299/ 720] blk.17.attn_output.weight            - [  4096,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 300/ 720] blk.17.attn_q.weight                 - [  2560,   4096,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 301/ 720] blk.17.attn_q_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 302/ 720] blk.17.attn_v.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 303/ 720] blk.17.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 304/ 720] blk.17.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 305/ 720] blk.17.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 306/ 720] blk.17.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 307/ 720] blk.17.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 308/ 720] blk.17.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 309/ 720] blk.17.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 310/ 720] blk.17.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 311/ 720] blk.17.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 312/ 720] blk.17.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 313/ 720] blk.18.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 314/ 720] blk.18.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 315/ 720] blk.18.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 316/ 720] blk.18.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 317/ 720] blk.18.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 318/ 720] blk.18.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 319/ 720] blk.18.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 320/ 720] blk.18.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 321/ 720] blk.18.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 322/ 720] blk.18.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 323/ 720] blk.18.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 324/ 720] blk.18.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 325/ 720] blk.18.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 326/ 720] blk.18.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 327/ 720] blk.18.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 328/ 720] blk.18.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 329/ 720] blk.18.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 330/ 720] blk.19.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 331/ 720] blk.19.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 332/ 720] blk.19.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 333/ 720] blk.19.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 334/ 720] blk.19.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 335/ 720] blk.19.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 336/ 720] blk.19.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 337/ 720] blk.19.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 338/ 720] blk.19.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 339/ 720] blk.19.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 340/ 720] blk.19.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 341/ 720] blk.19.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 342/ 720] blk.19.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 343/ 720] blk.19.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 344/ 720] blk.19.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 345/ 720] blk.19.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 346/ 720] blk.19.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 347/ 720] blk.20.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 348/ 720] blk.20.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 349/ 720] blk.20.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 350/ 720] blk.20.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 351/ 720] blk.20.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 352/ 720] blk.20.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 353/ 720] blk.20.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 354/ 720] blk.20.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 355/ 720] blk.20.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 356/ 720] blk.20.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 357/ 720] blk.20.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 358/ 720] blk.20.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 359/ 720] blk.20.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 360/ 720] blk.20.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 361/ 720] blk.20.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 362/ 720] blk.20.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 363/ 720] blk.20.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 364/ 720] blk.21.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 365/ 720] blk.21.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 366/ 720] blk.21.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 367/ 720] blk.21.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 368/ 720] blk.21.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 369/ 720] blk.21.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 370/ 720] blk.21.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 371/ 720] blk.21.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 372/ 720] blk.21.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 373/ 720] blk.21.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 374/ 720] blk.21.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 375/ 720] blk.21.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 376/ 720] blk.21.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 377/ 720] blk.21.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 378/ 720] blk.21.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 379/ 720] blk.21.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 380/ 720] blk.21.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 381/ 720] blk.22.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 382/ 720] blk.22.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 383/ 720] blk.22.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 384/ 720] blk.22.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 385/ 720] blk.22.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 386/ 720] blk.22.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 387/ 720] blk.22.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 388/ 720] blk.22.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 389/ 720] blk.22.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 390/ 720] blk.22.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 391/ 720] blk.22.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 392/ 720] blk.22.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 393/ 720] blk.22.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 394/ 720] blk.22.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 395/ 720] blk.22.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 396/ 720] blk.22.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 397/ 720] blk.22.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 398/ 720] blk.23.attn_k.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 399/ 720] blk.23.attn_k_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 400/ 720] blk.23.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 401/ 720] blk.23.attn_output.weight            - [  4096,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 402/ 720] blk.23.attn_q.weight                 - [  2560,   4096,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 403/ 720] blk.23.attn_q_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 404/ 720] blk.23.attn_v.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 405/ 720] blk.23.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 406/ 720] blk.23.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 407/ 720] blk.23.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 408/ 720] blk.23.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 409/ 720] blk.23.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 410/ 720] blk.23.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 411/ 720] blk.23.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 412/ 720] blk.23.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 413/ 720] blk.23.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 414/ 720] blk.23.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 415/ 720] blk.24.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 416/ 720] blk.24.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 417/ 720] blk.24.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 418/ 720] blk.24.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 419/ 720] blk.24.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 420/ 720] blk.24.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 421/ 720] blk.24.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 422/ 720] blk.24.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 423/ 720] blk.24.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 424/ 720] blk.24.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 425/ 720] blk.24.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 426/ 720] blk.24.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 427/ 720] blk.24.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 428/ 720] blk.24.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 429/ 720] blk.24.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 430/ 720] blk.24.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 431/ 720] blk.24.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 432/ 720] blk.25.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 433/ 720] blk.25.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 434/ 720] blk.25.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 435/ 720] blk.25.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 436/ 720] blk.25.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 437/ 720] blk.25.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 438/ 720] blk.25.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 439/ 720] blk.25.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 440/ 720] blk.25.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 441/ 720] blk.25.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 442/ 720] blk.25.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 443/ 720] blk.25.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 444/ 720] blk.25.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 445/ 720] blk.25.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 446/ 720] blk.25.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 447/ 720] blk.25.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 448/ 720] blk.25.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 449/ 720] blk.26.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 450/ 720] blk.26.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 451/ 720] blk.26.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 452/ 720] blk.26.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 453/ 720] blk.26.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 454/ 720] blk.26.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 455/ 720] blk.26.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 456/ 720] blk.26.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 457/ 720] blk.26.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 458/ 720] blk.26.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 459/ 720] blk.26.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 460/ 720] blk.26.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 461/ 720] blk.26.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 462/ 720] blk.26.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 463/ 720] blk.26.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 464/ 720] blk.26.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 465/ 720] blk.26.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 466/ 720] blk.27.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 467/ 720] blk.27.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 468/ 720] blk.27.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 469/ 720] blk.27.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 470/ 720] blk.27.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 471/ 720] blk.27.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 472/ 720] blk.27.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 473/ 720] blk.27.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 474/ 720] blk.27.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 475/ 720] blk.27.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 476/ 720] blk.27.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 477/ 720] blk.27.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 478/ 720] blk.27.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 479/ 720] blk.27.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 480/ 720] blk.27.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 481/ 720] blk.27.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 482/ 720] blk.27.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 483/ 720] blk.28.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 484/ 720] blk.28.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 485/ 720] blk.28.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 486/ 720] blk.28.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 487/ 720] blk.28.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 488/ 720] blk.28.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 489/ 720] blk.28.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 490/ 720] blk.28.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 491/ 720] blk.28.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 492/ 720] blk.28.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 493/ 720] blk.28.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 494/ 720] blk.28.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 495/ 720] blk.28.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 496/ 720] blk.28.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 497/ 720] blk.28.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 498/ 720] blk.28.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 499/ 720] blk.28.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 500/ 720] blk.29.attn_k.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 501/ 720] blk.29.attn_k_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 502/ 720] blk.29.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 503/ 720] blk.29.attn_output.weight            - [  4096,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 504/ 720] blk.29.attn_q.weight                 - [  2560,   4096,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 505/ 720] blk.29.attn_q_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 506/ 720] blk.29.attn_v.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 507/ 720] blk.29.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 508/ 720] blk.29.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 509/ 720] blk.29.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 510/ 720] blk.29.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 511/ 720] blk.29.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 512/ 720] blk.29.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 513/ 720] blk.29.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 514/ 720] blk.29.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 515/ 720] blk.29.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 516/ 720] blk.29.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 517/ 720] blk.30.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 518/ 720] blk.30.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 519/ 720] blk.30.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 520/ 720] blk.30.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 521/ 720] blk.30.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 522/ 720] blk.30.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 523/ 720] blk.30.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 524/ 720] blk.30.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 525/ 720] blk.30.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 526/ 720] blk.30.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 527/ 720] blk.30.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 528/ 720] blk.30.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 529/ 720] blk.30.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 530/ 720] blk.30.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 531/ 720] blk.30.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 532/ 720] blk.30.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 533/ 720] blk.30.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 534/ 720] blk.31.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 535/ 720] blk.31.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 536/ 720] blk.31.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 537/ 720] blk.31.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 538/ 720] blk.31.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 539/ 720] blk.31.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 540/ 720] blk.31.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 541/ 720] blk.31.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 542/ 720] blk.31.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 543/ 720] blk.31.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 544/ 720] blk.31.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 545/ 720] blk.31.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 546/ 720] blk.31.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 547/ 720] blk.31.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 548/ 720] blk.31.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 549/ 720] blk.31.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 550/ 720] blk.31.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 551/ 720] blk.32.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 552/ 720] blk.32.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 553/ 720] blk.32.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 554/ 720] blk.32.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 555/ 720] blk.32.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 556/ 720] blk.32.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 557/ 720] blk.32.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 558/ 720] blk.32.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 559/ 720] blk.32.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 560/ 720] blk.32.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 561/ 720] blk.32.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 562/ 720] blk.32.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 563/ 720] blk.32.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 564/ 720] blk.32.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 565/ 720] blk.32.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 566/ 720] blk.32.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 567/ 720] blk.32.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 568/ 720] blk.33.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 569/ 720] blk.33.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 570/ 720] blk.33.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 571/ 720] blk.33.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 572/ 720] blk.33.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 573/ 720] blk.33.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 574/ 720] blk.33.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 575/ 720] blk.33.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 576/ 720] blk.33.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 577/ 720] blk.33.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 578/ 720] blk.33.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 579/ 720] blk.33.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 580/ 720] blk.33.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 581/ 720] blk.33.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 582/ 720] blk.33.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 583/ 720] blk.33.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 584/ 720] blk.33.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 585/ 720] blk.34.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 586/ 720] blk.34.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 587/ 720] blk.34.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 588/ 720] blk.34.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 589/ 720] blk.34.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 590/ 720] blk.34.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 591/ 720] blk.34.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 592/ 720] blk.34.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 593/ 720] blk.34.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 594/ 720] blk.34.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 595/ 720] blk.34.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 596/ 720] blk.34.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 597/ 720] blk.34.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 598/ 720] blk.34.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 599/ 720] blk.34.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 600/ 720] blk.34.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 601/ 720] blk.34.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 602/ 720] blk.35.attn_k.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 603/ 720] blk.35.attn_k_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 604/ 720] blk.35.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 605/ 720] blk.35.attn_output.weight            - [  4096,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 606/ 720] blk.35.attn_q.weight                 - [  2560,   4096,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 607/ 720] blk.35.attn_q_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 608/ 720] blk.35.attn_v.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 609/ 720] blk.35.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 610/ 720] blk.35.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 611/ 720] blk.35.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 612/ 720] blk.35.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 613/ 720] blk.35.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 614/ 720] blk.35.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 615/ 720] blk.35.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 616/ 720] blk.35.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 617/ 720] blk.35.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 618/ 720] blk.35.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 619/ 720] blk.36.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 620/ 720] blk.36.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 621/ 720] blk.36.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 622/ 720] blk.36.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 623/ 720] blk.36.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 624/ 720] blk.36.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 625/ 720] blk.36.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 626/ 720] blk.36.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 627/ 720] blk.36.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 628/ 720] blk.36.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 629/ 720] blk.36.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 630/ 720] blk.36.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 631/ 720] blk.36.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 632/ 720] blk.36.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 633/ 720] blk.36.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 634/ 720] blk.36.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 635/ 720] blk.36.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 636/ 720] blk.37.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 637/ 720] blk.37.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 638/ 720] blk.37.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 639/ 720] blk.37.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 640/ 720] blk.37.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 641/ 720] blk.37.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 642/ 720] blk.37.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 643/ 720] blk.37.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 644/ 720] blk.37.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 645/ 720] blk.37.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 646/ 720] blk.37.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 647/ 720] blk.37.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 648/ 720] blk.37.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 649/ 720] blk.37.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 650/ 720] blk.37.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 651/ 720] blk.37.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 652/ 720] blk.37.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 653/ 720] blk.38.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 654/ 720] blk.38.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 655/ 720] blk.38.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 656/ 720] blk.38.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 657/ 720] blk.38.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 658/ 720] blk.38.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 659/ 720] blk.38.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 660/ 720] blk.38.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 661/ 720] blk.38.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 662/ 720] blk.38.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 663/ 720] blk.38.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 664/ 720] blk.38.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 665/ 720] blk.38.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 666/ 720] blk.38.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 667/ 720] blk.38.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 668/ 720] blk.38.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 669/ 720] blk.38.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 670/ 720] blk.39.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 671/ 720] blk.39.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 672/ 720] blk.39.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 673/ 720] blk.39.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 674/ 720] blk.39.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 675/ 720] blk.39.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 676/ 720] blk.39.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 677/ 720] blk.39.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 678/ 720] blk.39.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 679/ 720] blk.39.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 680/ 720] blk.39.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 681/ 720] blk.39.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 682/ 720] blk.39.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 683/ 720] blk.39.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 684/ 720] blk.39.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 685/ 720] blk.39.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 686/ 720] blk.39.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 687/ 720] blk.40.attn_k.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 688/ 720] blk.40.attn_k_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 689/ 720] blk.40.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 690/ 720] blk.40.attn_output.weight            - [  2048,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 691/ 720] blk.40.attn_q.weight                 - [  2560,   2048,      1,      1], type =    f16, converting to q8_0 .. size =    10.00 MiB ->     5.31 MiB
[ 692/ 720] blk.40.attn_q_norm.weight            - [   256,      1,      1,      1], type =    f32, size =    0.001 MiB
[ 693/ 720] blk.40.attn_v.weight                 - [  2560,    512,      1,      1], type =    f16, converting to q8_0 .. size =     2.50 MiB ->     1.33 MiB
[ 694/ 720] blk.40.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 695/ 720] blk.40.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 696/ 720] blk.40.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 697/ 720] blk.40.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 698/ 720] blk.40.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 699/ 720] blk.40.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 700/ 720] blk.40.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 701/ 720] blk.40.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 702/ 720] blk.40.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 703/ 720] blk.40.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 704/ 720] blk.41.attn_k.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 705/ 720] blk.41.attn_k_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 706/ 720] blk.41.attn_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 707/ 720] blk.41.attn_output.weight            - [  4096,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 708/ 720] blk.41.attn_q.weight                 - [  2560,   4096,      1,      1], type =    f16, converting to q8_0 .. size =    20.00 MiB ->    10.62 MiB
[ 709/ 720] blk.41.attn_q_norm.weight            - [   512,      1,      1,      1], type =    f32, size =    0.002 MiB
[ 710/ 720] blk.41.attn_v.weight                 - [  2560,   1024,      1,      1], type =    f16, converting to q8_0 .. size =     5.00 MiB ->     2.66 MiB
[ 711/ 720] blk.41.ffn_down.weight               - [ 10240,   2560,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 712/ 720] blk.41.ffn_gate.weight               - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 713/ 720] blk.41.ffn_norm.weight               - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 714/ 720] blk.41.ffn_up.weight                 - [  2560,  10240,      1,      1], type =    f16, converting to q8_0 .. size =    50.00 MiB ->    26.56 MiB
[ 715/ 720] blk.41.inp_gate.weight               - [  2560,    256,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB
[ 716/ 720] blk.41.layer_output_scale.weight     - [     1,      1,      1,      1], type =    f32, size =    0.000 MiB
[ 717/ 720] blk.41.post_attention_norm.weight    - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 718/ 720] blk.41.post_ffw_norm.weight          - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 719/ 720] blk.41.post_norm.weight              - [  2560,      1,      1,      1], type =    f32, size =    0.010 MiB
[ 720/ 720] blk.41.proj.weight                   - [   256,   2560,      1,      1], type =    f16, converting to q8_0 .. size =     1.25 MiB ->     0.66 MiB

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

llama.cpp reasoning 옵션  (0) 2026.06.01
torchvision model  (0) 2026.06.01
antigravity 2  (0) 2026.05.28
telegram bot api 로 기능 변경하기  (0) 2026.05.25
python huggingface 저장경로 변경하기  (0) 2026.05.24
Posted by 구차니
모종의 음모/ai 프로그램2026. 5. 28. 10:30

antigravity 2

생각해보니 기존에 쓰던 버전도 몰랐네 -_-?

 

그래서 리눅스용 tar 파일 받고, 실행하는데 머.. 특별한건 없어 보인다.

 

로그인 하고

 

동의하고

 

뜬다.

 

소심하게 좌측상단 메뉴 옆에 Antigravity 누르면 버전 나오고 땡

 

테마 고르고

 

멀 할지 고르고(아무것도 선택안해도 넘어감)

 

짜잔~

이전에 작업했던 것들이 왼쪽에 쭈르륵 뜬다. 이건 좀 좋다고 해야하나 말아야 하나?

 

음.. 저번에 다 써먹었는데 이번에는 좀 쓸만한 양을 주려나?

 

근데, 우분투에서 패키지로 깐게 아니라

설치하는 경로에서 실행해줘야 해서 조금 귀찮다. 어떻게 데스크탑 아이콘 설정 바꾸지?

 

 

Exec 뒤에 경로를 적절히 조정해주면 된다.

$ cd ~/.local/share/applications
$ vi antigravity-ide.desktop
$ cat antigravity-ide.desktop
[Desktop Entry]
Name=Antigravity IDE
Comment=Antigravity IDE v2.0 - Experience liftoff
GenericName=IDE
Exec="/home/minimonk/Antigravity-x64/antigravity" %F
Icon=antigravity
Type=Application
Terminal=false
StartupNotify=true
StartupWMClass=Antigravity
Categories=Development;IDE;TextEditor;
MimeType=application/x-antigravity-workspace;
$ chmod +x antigravity-ide.desktop 
$ update-desktop-database ~/.local/share/applications/

[링크 : https://discuss.ai.google.dev/t/ubuntu-24-04-upgrading-to-antigravity-2-0-and-antigravity-ide/146516/4]

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

torchvision model  (0) 2026.06.01
safetensors to gguf 일단 실패  (0) 2026.06.01
telegram bot api 로 기능 변경하기  (0) 2026.05.25
python huggingface 저장경로 변경하기  (0) 2026.05.24
vllm 설치, 실행 실패 -> 사실상 포기  (0) 2026.05.24
Posted by 구차니
모종의 음모/ai 프로그램2026. 5. 25. 23:12

telegram bot api 로 기능 변경하기

그래서.. openclaw 에서 대화창 인증하고 나서 바로 메뉴가 추가되었던 건가..

그리고 set_my_commands() 로 등록되어도 채팅 내용으로 받아와지진 않는다.

다른 무슨 처리가 별도로 필요한 듯.

 

Bot settings

set_my_commands() Used for setting the list of commands
delete_my_commands() Used for deleting the list of commands
get_my_commands() Used for obtaining the list of commands
get_my_default_administrator_rights() Used for obtaining the default administrator rights for the bot
set_my_default_administrator_rights() Used for setting the default administrator rights for the bot
get_chat_menu_button() Used for obtaining the menu button of a private chat or the default menu button
set_chat_menu_button() Used for setting the menu button of a private chat or the default menu button
set_my_description() Used for setting the description of the bot
get_my_description() Used for obtaining the description of the bot
set_my_short_description() Used for setting the short description of the bot
get_my_short_description() Used for obtaining the short description of the bot
set_my_name() Used for setting the name of the bot
get_my_name() Used for obtaining the name of the bot
set_my_profile_photo() Used for setting the profile photo of the bot
remove_my_profile_photo() Used for removing the profile photo of the bot

[링크 : https://docs.python-telegram-bot.org/en/stable/telegram.bot.html]

 

봇 토큰으로 채팅내 메뉴 목록 보기

import asyncio
from telegram import Bot

async def check_bot_menu():
    # Replace with your actual Bot Token from @BotFather
    bot_token = "YOUR_BOT_TOKEN_HERE"
    bot = Bot(token=bot_token)
    
    print("Fetching active menu commands...")
    # Retrieves a list of BotCommand objects
    commands = await bot.get_my_commands()
    
    if not commands:
        print("No menu commands found. The menu is empty.")
    else:
        print(f"Found {len(commands)} command(s):")
        for cmd in commands:
            print(f"/{cmd.command} - {cmd.description}")

if __name__ == "__main__":
    asyncio.run(check_bot_menu())

 

봇 토큰으로 채팅내 메뉴 비우기

import asyncio
from telegram import Bot

async def remove_bot_menu():
    # Replace with your actual Bot Token from @BotFather
    bot_token = "YOUR_BOT_TOKEN_HERE"
    bot = Bot(token=bot_token)
    
    print("Removing all menu commands...")
    # Passing no arguments clears the commands globally
    await bot.delete_my_commands()
    print("Success! All menu commands have been removed.")

# Run the async function
if __name__ == "__main__":
    asyncio.run(remove_bot_menu())

 

봇 토큰과 채팅id를 이용해서 즉각적으로 메뉴 변경하기.

set_my_commands 로는 대화창을 나가도 바뀌지 않아서, 특정 채팅 아이디 넣어서야 바뀌게 되었음

import asyncio
from telegram import Bot, BotCommand, BotCommandScopeChat

async def force_update_user_menu():
    # 1. 봇 토큰과 대상 유저의 Chat ID를 입력하세요.
    BOT_TOKEN = "YOUR_BOT_TOKEN_HERE"
    USER_CHAT_ID = 123456789  # 숫자로 된 유저의 chat_id 입력
    
    bot = Bot(token=BOT_TOKEN)
    
    # 2. 변경하고 싶은 새로운 메뉴 목록을 정의합니다.
    new_commands = [
        BotCommand(command="home", description="🏠 홈 화면으로"),
        BotCommand(command="mypage", description="👤 내 정보 보기"),
        BotCommand(command="support", description="❓ 고객 센터")
    ]
    
    print(f"User({USER_CHAT_ID})의 메뉴를 즉시 변경합니다...")
    
    # 3. scope를 'BotCommandScopeChat'으로 지정하여 특정 chat_id에 즉시 강제 적용
    await bot.set_my_commands(
        commands=new_commands,
        scope=BotCommandScopeChat(chat_id=USER_CHAT_ID)
    )
    
    print("성공! 해당 유저의 텔레그램 앱 화면에서 메뉴가 즉시 업데이트되었습니다.")

if __name__ == "__main__":
    asyncio.run(force_update_user_menu())

 

요건 아직 테스트 안해봄

from telegram import Update, BotCommand, BotCommandScopeChat
from telegram.ext import Application, CommandHandler, ContextTypes

# 봇 토큰 설정
BOT_TOKEN = "YOUR_BOT_TOKEN_HERE"

# 1. /start 명령어가 들어왔을 때 실행될 함수 (여기서 메뉴를 즉시 변경)
async def start_command(update: Update, context: ContextTypes.DEFAULT_TYPE):
    chat_id = update.effective_chat.id
    
    # 해당 유저의 화면에 보일 새로운 메뉴 정의
    new_menu = [
        BotCommand(command="home", description="🏠 홈 화면으로"),
        BotCommand(command="mypage", description="👤 내 정보 보기")
    ]
    
    # 유저 화면의 메뉴 버튼 즉시 업데이트
    await context.bot.set_my_commands(
        commands=new_menu,
        scope=BotCommandScopeChat(chat_id=chat_id)
    )
    
    await update.message.reply_text(
        "반갑습니다! 메뉴 버튼이 업데이트되었습니다.\n"
        "좌측 하단의 [Menu] 버튼을 누르거나 명령어를 입력해보세요!"
    )

# 2. 메뉴 버튼의 /home 처리를 위한 핸들러 함수
async def home_command(update: Update, context: ContextTypes.DEFAULT_TYPE):
    await update.message.reply_text("🏠 홈 화면으로 이동했습니다.")

# 3. 메뉴 버튼의 /mypage 처리를 위한 핸들러 함수
async def mypage_command(update: Update, context: ContextTypes.DEFAULT_TYPE):
    user = update.effective_user
    await update.message.reply_text(f"👤 [{user.first_name}] 님의 마이페이지입니다.")


def main():
    # 애플리케이션 빌드
    application = Application.builder().token(BOT_TOKEN).build()

    # ⭐ [핵심] 메뉴에 추가한 명령어들과 핸들러 함수를 1:1로 매핑해줍니다.
    application.add_handler(CommandHandler("start", start_command))
    application.add_handler(CommandHandler("home", home_command))       # /home 처리
    application.add_handler(CommandHandler("mypage", mypage_command))   # /mypage 처리

    # 봇 시작 (폴링 방식)
    print("봇이 시작되었습니다. 대화를 시작하세요...")
    application.run_polling()

if __name__ == "__main__":
    main()
저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

safetensors to gguf 일단 실패  (0) 2026.06.01
antigravity 2  (0) 2026.05.28
python huggingface 저장경로 변경하기  (0) 2026.05.24
vllm 설치, 실행 실패 -> 사실상 포기  (0) 2026.05.24
antigravity gemini flash 할당량이 깃털 같구만?  (0) 2026.05.22
Posted by 구차니
모종의 음모/ai 프로그램2026. 5. 24. 22:19

python huggingface 저장경로 변경하기

HF_HOME 환경 변수를 바꿔주면 된다고

리눅스에서는 기본값 ~/.cache/huggingface

[링크 : https://developer0hye.tistory.com/775]

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

antigravity 2  (0) 2026.05.28
telegram bot api 로 기능 변경하기  (0) 2026.05.25
vllm 설치, 실행 실패 -> 사실상 포기  (0) 2026.05.24
antigravity gemini flash 할당량이 깃털 같구만?  (0) 2026.05.22
openclaw agent 구성 관련  (0) 2026.05.22
Posted by 구차니
모종의 음모/ai 프로그램2026. 5. 24. 17:56

vllm 설치, 실행 실패 -> 사실상 포기

실행하라면 이렇게 하라고 하는데

vllm serve google/gemma-4-E4B-it \
  --max-model-len <n_of_tokens> # up to 131072

[링크 : https://docs.vllm.ai/projects/recipes/en/latest/Google/Gemma4.html]

 

아따.. 드럽게 크네. 그나저나 허깅페이스에서 바로 받으려나?

그리고 gguf가 양자화 되서 작은거였나. 기존에 내가 쓰던데 Q4_K_M 이라 4.7기가 정도 되었는데

model.safetensors는 16기가나 된다. 와우

[링크 : https://huggingface.co/google/gemma-4-E4B]

 

vllm : 1080 ti 라니 불량식품이잖아! 퉤!

$ vllm serve google/gemma-4-E4B-it
(APIServer pid=52696) INFO 05-24 22:20:19 [utils.py:306]
(APIServer pid=52696) INFO 05-24 22:20:19 [utils.py:306]        █     █     █▄   ▄█
(APIServer pid=52696) INFO 05-24 22:20:19 [utils.py:306]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.21.0
(APIServer pid=52696) INFO 05-24 22:20:19 [utils.py:306]   █▄█▀ █     █     █     █  model   google/gemma-4-E4B-it
(APIServer pid=52696) INFO 05-24 22:20:19 [utils.py:306]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
(APIServer pid=52696) INFO 05-24 22:20:19 [utils.py:306]
(APIServer pid=52696) INFO 05-24 22:20:19 [utils.py:240] non-default args: {'model_tag': 'google/gemma-4-E4B-it', 'model': 'google/gemma-4-E4B-it'}
(APIServer pid=52696) Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
config.json: 5.14kB [00:00, 17.7MB/s]
processor_config.json: 1.69kB [00:00, 1.70MB/s]
(APIServer pid=52696) INFO 05-24 22:20:29 [model.py:568] Resolved architecture: Gemma4ForConditionalGeneration
(APIServer pid=52696) WARNING 05-24 22:20:29 [model.py:1982] Your device 'NVIDIA GeForce GTX 1080 Ti' (with compute capability 6.1) doesn't support torch.bfloat16. Falling back to torch.float16 for compatibility.
(APIServer pid=52696) WARNING 05-24 22:20:29 [model.py:2035] Casting torch.bfloat16 to torch.float16.
(APIServer pid=52696) INFO 05-24 22:20:29 [model.py:1697] Using max model len 131072
(APIServer pid=52696) INFO 05-24 22:20:29 [config.py:101] Gemma4 model has heterogeneous head dimensions (head_dim=256, global_head_dim=512). Forcing TRITON_ATTN backend to prevent mixed-backend numerical divergence.
(APIServer pid=52696) INFO 05-24 22:20:29 [vllm.py:886] Asynchronous scheduling is enabled.
(APIServer pid=52696) INFO 05-24 22:20:29 [kernel.py:212] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native'])
tokenizer_config.json: 2.10kB [00:00, 2.05MB/s]
tokenizer.json: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 32.2M/32.2M [00:02<00:00, 14.8MB/s]
chat_template.jinja: 17.3kB [00:00, 12.6MB/s]
generation_config.json: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 208/208 [00:00<00:00, 1.68MB/s]
(EngineCore pid=52767) INFO 05-24 22:21:24 [core.py:109] Initializing a V1 LLM engine (v0.21.0) with config: model='google/gemma-4-E4B-it', speculative_config=None, tokenizer='google/gemma-4-E4B-it', skip_t
okenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=131072, download_dir=None, load_format=auto, tensor_parallel_size=
1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=False,
enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=Fal
se, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_tr
aces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_itera
tion_details=False), seed=0, served_model_name=google/gemma-4-E4B-it, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.VLLM_COMPILE:
3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'ir_enable_torch_wrap': True, 'splitting_ops': ['vllm::unified_attention_w
ith_output', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm
::gdn_attention_core_xpu', 'vllm::olmo_hybrid_gdn_full_forward', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::deepseek_v4_attention', 'vllm::unified_kv_
cache_update', 'vllm::unified_mla_kv_cache_update'], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'en
coder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_as
serts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups
': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256, 272, 288, 304, 3
20, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496, 512], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant':
False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 512, 'dynamic_s
hapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []},
 kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native']), enable_flashinfer_autotune=False, moe_backend='auto')
(EngineCore pid=52767) /home/minimonk/.local/lib/python3.10/site-packages/torch/cuda/__init__.py:371: UserWarning: Found GPU0 NVIDIA GeForce GTX 1080 Ti which is of compute capability (CC) 6.1.
(EngineCore pid=52767) The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:
(EngineCore pid=52767) - 7.5 which supports hardware CC >=7.5,<8.0
(EngineCore pid=52767) - 8.0 which supports hardware CC >=8.0,<9.0 except {8.7}
(EngineCore pid=52767) - 8.6 which supports hardware CC >=8.6,<9.0 except {8.7}
(EngineCore pid=52767) - 9.0 which supports hardware CC >=9.0,<10.0
(EngineCore pid=52767) - 8.6 which supports hardware CC >=8.6,<9.0 except {8.7}                                                                                                                      [108/299]
(EngineCore pid=52767) - 9.0 which supports hardware CC >=9.0,<10.0
(EngineCore pid=52767) - 10.0 which supports hardware CC >=10.0,<11.0 except {10.1}
(EngineCore pid=52767) - 12.0 which supports hardware CC >=12.0,<13.0
(EngineCore pid=52767) Please follow the instructions at https://pytorch.org/get-started/locally/ to install a PyTorch release that supports one of these CUDA versions: 12.6
(EngineCore pid=52767)   _warn_unsupported_code(d, device_cc, code_ccs)
(EngineCore pid=52767) /home/minimonk/.local/lib/python3.10/site-packages/torch/cuda/__init__.py:371: UserWarning: Found GPU1 NVIDIA GeForce GTX 1080 Ti which is of compute capability (CC) 6.1.
(EngineCore pid=52767) The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:
(EngineCore pid=52767) - 7.5 which supports hardware CC >=7.5,<8.0                                                                                                                                            (EngineCore pid=52767) - 8.0 which supports hardware CC >=8.0,<9.0 except {8.7}
(EngineCore pid=52767) - 8.6 which supports hardware CC >=8.6,<9.0 except {8.7}
(EngineCore pid=52767) - 9.0 which supports hardware CC >=9.0,<10.0
(EngineCore pid=52767) - 10.0 which supports hardware CC >=10.0,<11.0 except {10.1}
(EngineCore pid=52767) - 12.0 which supports hardware CC >=12.0,<13.0
(EngineCore pid=52767) Please follow the instructions at https://pytorch.org/get-started/locally/ to install a PyTorch release that supports one of these CUDA versions: 12.6
(EngineCore pid=52767)   _warn_unsupported_code(d, device_cc, code_ccs)
(EngineCore pid=52767) /home/minimonk/.local/lib/python3.10/site-packages/torch/cuda/__init__.py:489: UserWarning:
(EngineCore pid=52767) NVIDIA GeForce GTX 1080 Ti with CUDA capability sm_61 is not compatible with the current PyTorch installation.
(EngineCore pid=52767) The current PyTorch install supports CUDA capabilities sm_75 sm_80 sm_86 sm_90 sm_100 sm_120.
(EngineCore pid=52767) If you want to use the NVIDIA GeForce GTX 1080 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/
(EngineCore pid=52767)
(EngineCore pid=52767)   queued_call()
(EngineCore pid=52767) INFO 05-24 22:21:30 [parallel_state.py:1410] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.40.238:47913 backend=nccl
(EngineCore pid=52767) INFO 05-24 22:21:30 [parallel_state.py:1723] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A
(EngineCore pid=52767) WARNING 05-24 22:21:31 [topk_topp_sampler.py:61] FlashInfer top-p/top-k sampling not supported on compute capability 6.1; falling back to PyTorch-native sampler. Set VLLM_USE_FLASHINF
ER_SAMPLER=0 to silence.
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140] EngineCore failed to start.
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140] Traceback (most recent call last):
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 1114, in run_engine_core
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     engine_core = EngineCoreProc(*args, engine_index=dp_rank, **kwargs)
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     return func(*args, **kwargs)
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 880, in __init__
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     super().__init__(
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 118, in __init__
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.model_executor = executor_class(vllm_config)
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     return func(*args, **kwargs)
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/executor/abstract.py", line 109, in __init__
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self._init_executor()
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/executor/uniproc_executor.py", line 60, in _init_executor
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.driver_worker.init_device()
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/worker_base.py", line 317, in init_device
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.worker.init_device()  # type: ignore
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     return func(*args, **kwargs)
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_worker.py", line 330, in init_device
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.model_runner = GPUModelRunnerV1(self.vllm_config, self.device)
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_model_runner.py", line 629, in __init__
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.input_batch = InputBatch(
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_input_batch.py", line 171, in __init__
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.block_table = MultiGroupBlockTable(
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 267, in __init__
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.block_tables = [
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 268, in <listcomp>
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     BlockTable(
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 268, in <listcomp>                          [54/299]
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     BlockTable(
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 70, in __init__
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.block_table = self._make_buffer(
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 218, in _make_buffer
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     return CpuGpuBuffer(
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/utils.py", line 120, in __init__
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]     self.gpu = torch.zeros_like(self.cpu, device=device)
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140] torch.AcceleratorError: CUDA error: no kernel image is available for execution on the device                                                       
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140] Search for `cudaErrorNoKernelImageForDevice' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140] CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140] For debugging consider passing CUDA_LAUNCH_BLOCKING=1
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140] Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
(EngineCore pid=52767) ERROR 05-24 22:21:32 [core.py:1140]
(EngineCore pid=52767) Process EngineCore:
(EngineCore pid=52767) Traceback (most recent call last):
(EngineCore pid=52767)   File "/usr/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
(EngineCore pid=52767)     self.run()
(EngineCore pid=52767)   File "/usr/lib/python3.10/multiprocessing/process.py", line 108, in run
(EngineCore pid=52767)     self._target(*self._args, **self._kwargs)
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 1144, in run_engine_core
(EngineCore pid=52767)     raise e
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 1114, in run_engine_core
(EngineCore pid=52767)     engine_core = EngineCoreProc(*args, engine_index=dp_rank, **kwargs)
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=52767)     return func(*args, **kwargs)
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 880, in __init__
(EngineCore pid=52767)     super().__init__(
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 118, in __init__
(EngineCore pid=52767)     self.model_executor = executor_class(vllm_config)
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=52767)     return func(*args, **kwargs)
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/executor/abstract.py", line 109, in __init__
(EngineCore pid=52767)     self._init_executor()
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/executor/uniproc_executor.py", line 60, in _init_executor
(EngineCore pid=52767)     self.driver_worker.init_device()
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/worker_base.py", line 317, in init_device
(EngineCore pid=52767)     self.worker.init_device()  # type: ignore
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=52767)     return func(*args, **kwargs)
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_worker.py", line 330, in init_device
(EngineCore pid=52767)     self.model_runner = GPUModelRunnerV1(self.vllm_config, self.device)
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_model_runner.py", line 629, in __init__
(EngineCore pid=52767)     self.input_batch = InputBatch(
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_input_batch.py", line 171, in __init__
(EngineCore pid=52767)     self.block_table = MultiGroupBlockTable(
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 267, in __init__
(EngineCore pid=52767)     self.block_tables = [
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 268, in <listcomp>
(EngineCore pid=52767)     BlockTable(
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 70, in __init__
(EngineCore pid=52767)     self.block_table = self._make_buffer(
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 218, in _make_buffer
(EngineCore pid=52767)     return CpuGpuBuffer(
(EngineCore pid=52767)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/utils.py", line 120, in __init__
(EngineCore pid=52767)     self.gpu = torch.zeros_like(self.cpu, device=device)
(EngineCore pid=52767) torch.AcceleratorError: CUDA error: no kernel image is available for execution on the device
(EngineCore pid=52767) Search for `cudaErrorNoKernelImageForDevice' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.
(EngineCore pid=52767) CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
(EngineCore pid=52767) For debugging consider passing CUDA_LAUNCH_BLOCKING=1
(EngineCore pid=52767) Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
(EngineCore pid=52767)
[rank0]:[W524 22:21:32.199264489 ProcessGroupNCCL.cpp:1575] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.
org/docs/stable/distributed.html#shutdown (function operator())
(APIServer pid=52696) Traceback (most recent call last):
(APIServer pid=52696)   File "/home/minimonk/.local/bin/vllm", line 8, in <module>
(APIServer pid=52696)     sys.exit(main())
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/cli/main.py", line 92, in main
(APIServer pid=52696)     args.dispatch_function(args)
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/cli/serve.py", line 122, in cmd
(APIServer pid=52696)     uvloop.run(run_server(args))
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/uvloop/__init__.py", line 69, in run
(APIServer pid=52696)     return loop.run_until_complete(wrapper())
(APIServer pid=52696)   File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/uvloop/__init__.py", line 48, in wrapper
(APIServer pid=52696)     return await main
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 693, in run_server
(APIServer pid=52696)     await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 707, in run_server_worker
(APIServer pid=52696)     async with build_async_engine_client(
(APIServer pid=52696)   File "/usr/lib/python3.10/contextlib.py", line 199, in __aenter__
(APIServer pid=52696)     return await anext(self.gen)
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 100, in build_async_engine_client
(APIServer pid=52696)     async with build_async_engine_client_from_engine_args(
(APIServer pid=52696)   File "/usr/lib/python3.10/contextlib.py", line 199, in __aenter__
(APIServer pid=52696)     return await anext(self.gen)
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 136, in build_async_engine_client_from_engine_args
(APIServer pid=52696)     async_llm = AsyncLLM.from_vllm_config(
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/async_llm.py", line 217, in from_vllm_config
(APIServer pid=52696)     return cls(
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/async_llm.py", line 146, in __init__
(APIServer pid=52696)     self.engine_core = EngineCoreClient.make_async_mp_client(
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(APIServer pid=52696)     return func(*args, **kwargs)
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core_client.py", line 130, in make_async_mp_client
(APIServer pid=52696)     return AsyncMPClient(*client_args)
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(APIServer pid=52696)     return func(*args, **kwargs)
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core_client.py", line 900, in __init__
(APIServer pid=52696)     super().__init__(
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core_client.py", line 535, in __init__
(APIServer pid=52696)     with launch_core_engines(
(APIServer pid=52696)   File "/usr/lib/python3.10/contextlib.py", line 142, in __exit__
(APIServer pid=52696)     next(self.gen)
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/utils.py", line 1128, in launch_core_engines
(APIServer pid=52696)     wait_for_engine_startup(
(APIServer pid=52696)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/utils.py", line 1187, in wait_for_engine_startup
(APIServer pid=52696)     raise RuntimeError(
(APIServer pid=52696) RuntimeError: Engine core initialization failed. See root cause above. Failed core proc(s): {}

 

파스칼은 아예 지원 하드웨어에서 빼버린건가?

[링크 : https://docs.vllm.ai/en/latest/features/quantization/]

 

fork로 이런것도 존재하는데, 아래 pascal-pkgs-ci로 대체 된다고

[링크 : https://github.com/cduk/vllm-pascal]

 

도커로 시도해야하나..

[링크 : https://github.com/sasha0552/pascal-pkgs-ci]

    [링크 : https://github.com/vllm-project/vllm/issues/19542]

 

위의 도커를 불러오게 하면되려나? 볼륨은 로컬 캐싱에서 HF_HOME 으로 변경해주면 좋을듯

docker run -itd --name gemma4 \
    --ipc=host \
    --network host \
    --shm-size 16G \
    --gpus all \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    vllm/vllm-openai:latest \
        --model google/gemma-4-31B-it \
        --tensor-parallel-size 2 \
        --max-model-len 32768 \
        --gpu-memory-utilization 0.90 \
        --host 0.0.0.0 \
        --port 8000

[링크 : https://docs.vllm.ai/projects/recipes/en/latest/Google/Gemma4.html#docker-deployment]

 

아래와 같이 도커에서 그래픽 카드를 인식하지 못하면

docker: Error response from daemon: could not select device driver "" with capabilities: [[gpu]]

 

nvidia-container-toolkit을 설치하고 도커를 재기동하면 된단다

$ distribution=$(. /etc/os-release;echo $ID$VERSION_ID)    && curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -    && curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list
$ sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit
$ sudo systemctl restart docker

[링크 : https://bluecolorsky.tistory.com/110]

 

docker run -itd --name gemma4 \
    --ipc=host \
    --network host \
    --gpus all \
    -v /mnt/huggingface:/root/.cache/huggingface \
    ghcr.io/sasha0552/vllm\
        --model google/gemma-4-e4b-it \
        --tensor-parallel-size 2 \
        --max-model-len 131072\
        --gpu-memory-utilization 0.90 \
        --host 0.0.0.0 \
        --port 8000

[링크 : https://github.com/sasha0552/pascal-pkgs-ci/pkgs/container/vllm]

 

에라모르겠다 ㅋㅋ

$ docker ps -a
CONTAINER ID   IMAGE                    COMMAND                  CREATED          STATUS                      PORTS     NAMES
fdd1ba582924   ghcr.io/sasha0552/vllm   "python3 -m vllm.ent…"   42 seconds ago   Exited (1) 26 seconds ago             gemma4

[링크 : https://data-newbie.tistory.com/m/1012]

[링크 : https://coding-review.tistory.com/m/608]

 

+

2026.05.25

docker run -itd \
  --name gemma4 \
  --ipc=host \
  --network host \
  --gpus all \
  -v /mnt/huggingface:/root/.cache/huggingface \
  ghcr.io/sasha0552/vllm

 

문제없이 되는거 같으면서도

왜 qwen3-0.6B가 언급이 되지?

INFO 05-25 05:00:09 [__init__.py:241] Automatically detected platform cuda.
(APIServer pid=1) INFO 05-25 05:00:11 [api_server.py:1873] vLLM API server version 999.999.999
(APIServer pid=1) INFO 05-25 05:00:11 [utils.py:326] non-default args: {}
(APIServer pid=1) INFO 05-25 05:00:19 [__init__.py:742] Resolved architecture: Qwen3ForCausalLM
(APIServer pid=1) WARNING 05-25 05:00:19 [__init__.py:2828] Your device 'NVIDIA GeForce GTX 1080 Ti' (with compute capability 6.1) doesn't support torch.bfloat16. Falling back to torch.float16 for compatibility.
(APIServer pid=1) WARNING 05-25 05:00:19 [__init__.py:2879] Casting torch.bfloat16 to torch.float16.
(APIServer pid=1) INFO 05-25 05:00:19 [__init__.py:1774] Using max model len 40960
(APIServer pid=1) WARNING 05-25 05:00:19 [arg_utils.py:1806] Compute Capability < 8.0 is not supported by the V1 Engine. Falling back to V0.
(APIServer pid=1) WARNING 05-25 05:00:19 [arg_utils.py:1580] Chunked prefill is enabled by default for models with max_model_len > 32K. Chunked prefill might not work with some features or models. If you encounter any issues, please disable by launching with --enable-chunked-prefill=False.
(APIServer pid=1) INFO 05-25 05:00:20 [scheduler.py:222] Chunked prefill is enabled with max_num_batched_tokens=2048.
(APIServer pid=1) INFO 05-25 05:00:20 [api_server.py:295] Started engine process with PID 36
INFO 05-25 05:00:24 [__init__.py:241] Automatically detected platform cuda.
INFO 05-25 05:00:25 [llm_engine.py:222] Initializing a V0 LLM engine (v999.999.999) with config: model='Qwen/Qwen3-0.6B', speculative_config=None, tokenizer='Qwen/Qwen3-0.6B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config={}, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=40960, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, decoding_config=DecodingConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_backend=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Qwen/Qwen3-0.6B, enable_prefix_caching=None, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":0,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":null,"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"cudagraph_mode":0,"use_cudagraph":true,"cudagraph_num_of_warmups":0,"cudagraph_capture_sizes":[256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"pass_config":{"enable_fusion":false,"enable_noop":false},"max_capture_size":256,"local_cache_dir":null}, use_cached_outputs=True,
INFO 05-25 05:00:28 [cuda.py:374] Cannot use FlashAttention-2 backend for Volta and Turing GPUs.
INFO 05-25 05:00:28 [cuda.py:419] Using XFormers backend.
INFO 05-25 05:00:28 [parallel_state.py:1134] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
INFO 05-25 05:00:28 [model_runner.py:1080] Starting to load model Qwen/Qwen3-0.6B...
INFO 05-25 05:00:29 [weight_utils.py:296] Using model weights format ['*.safetensors']
INFO 05-25 05:00:29 [weight_utils.py:349] No model.safetensors.index.json found in remote.
Loading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  3.47it/s]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  3.47it/s]

INFO 05-25 05:00:29 [default_loader.py:267] Loading weights took 0.32 seconds
INFO 05-25 05:00:30 [model_runner.py:1112] Model loading took 1.1201 GiB and 1.275699 seconds
INFO 05-25 05:00:31 [worker.py:296] Memory profiling takes 1.07 seconds
INFO 05-25 05:00:31 [worker.py:296] the current vLLM instance can use total_gpu_memory (10.90GiB) x gpu_memory_utilization (0.90) = 9.81GiB
INFO 05-25 05:00:31 [worker.py:296] model weights take 1.12GiB; non_torch_memory takes 0.04GiB; PyTorch activation peak memory takes 1.39GiB; the rest of the memory reserved for KV Cache is 7.26GiB.
INFO 05-25 05:00:31 [executor_base.py:114] # cuda blocks: 4247, # CPU blocks: 2340
INFO 05-25 05:00:31 [executor_base.py:119] Maximum concurrency for 40960 tokens per request: 1.66x
INFO 05-25 05:00:34 [model_runner.py:1383] Capturing cudagraphs for decoding. This may lead to unexpected consequences if the model is not static. To run the model in eager mode, set 'enforce_eager=True' or use '--enforce-eager' in the CLI. If out-of-memory error occurs during cudagraph capture, consider decreasing `gpu_memory_utilization` or switching to eager mode. You can also reduce the `max_num_seqs` as needed to decrease memory usage.
Capturing CUDA graph shapes: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 35/35 [00:12<00:00,  2.85it/s]
INFO 05-25 05:00:46 [model_runner.py:1535] Graph capturing finished in 12 secs, took 0.19 GiB
INFO 05-25 05:00:46 [llm_engine.py:417] init engine (profile, create kv cache, warmup model) took 16.41 seconds
(APIServer pid=1) INFO 05-25 05:00:46 [api_server.py:1679] Supported_tasks: ['generate']
(APIServer pid=1) WARNING 05-25 05:00:46 [__init__.py:1658] Default sampling parameters have been overridden by the model's Hugging Face generation config recommended from the model creator. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
(APIServer pid=1) INFO 05-25 05:00:46 [serving_responses.py:124] Using default chat sampling params from model: {'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}
(APIServer pid=1) INFO 05-25 05:00:47 [serving_chat.py:135] Using default chat sampling params from model: {'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}
(APIServer pid=1) INFO 05-25 05:00:47 [serving_completion.py:77] Using default completion sampling params from model: {'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}
(APIServer pid=1) INFO 05-25 05:00:47 [api_server.py:1948] Starting vLLM API server 0 on http://0.0.0.0:8000
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:36] Available routes are:
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /openapi.json, Methods: HEAD, GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /docs, Methods: HEAD, GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /docs/oauth2-redirect, Methods: HEAD, GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /redoc, Methods: HEAD, GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /health, Methods: GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /load, Methods: GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /ping, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /ping, Methods: GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /tokenize, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /detokenize, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/models, Methods: GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /version, Methods: GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/responses, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/responses/{response_id}, Methods: GET
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/responses/{response_id}/cancel, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/chat/completions, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/completions, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/embeddings, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /pooling, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /classify, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /score, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/score, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/audio/transcriptions, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/audio/translations, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /rerank, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v1/rerank, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /v2/rerank, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /scale_elastic_ep, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /is_scaling_elastic_ep, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /invocations, Methods: POST
(APIServer pid=1) INFO 05-25 05:00:47 [launcher.py:44] Route: /metrics, Methods: GET
(APIServer pid=1) INFO:     Started server process [1]
(APIServer pid=1) INFO:     Waiting for application startup.
(APIServer pid=1) INFO:     Application startup complete.

 

파스칼 P40 용으로 시도하는데 여전히 안된다. 아놔.. 포기!

[링크 : https://github.com/uaysk/vllm-pascal]

 

$ vllm serve google/gemma-4-E4B-it
(APIServer pid=65016) INFO 05-25 21:22:17 [utils.py:306]
(APIServer pid=65016) INFO 05-25 21:22:17 [utils.py:306]        █     █     █▄   ▄█
(APIServer pid=65016) INFO 05-25 21:22:17 [utils.py:306]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.21.0
(APIServer pid=65016) INFO 05-25 21:22:17 [utils.py:306]   █▄█▀ █     █     █     █  model   google/gemma-4-E4B-it
(APIServer pid=65016) INFO 05-25 21:22:17 [utils.py:306]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
(APIServer pid=65016) INFO 05-25 21:22:17 [utils.py:306]
(APIServer pid=65016) INFO 05-25 21:22:17 [utils.py:240] non-default args: {'model_tag': 'google/gemma-4-E4B-it', 'model': 'google/gemma-4-E4B-it'}
(APIServer pid=65016) Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
config.json: 5.14kB [00:00, 4.19MB/s]
processor_config.json: 1.69kB [00:00, 6.85MB/s]
(APIServer pid=65016) INFO 05-25 21:22:19 [model.py:568] Resolved architecture: Gemma4ForConditionalGeneration
(APIServer pid=65016) WARNING 05-25 21:22:19 [model.py:1982] Your device 'NVIDIA GeForce GTX 1080 Ti' (with compute capability 6.1) doesn't support torch.bfloat16. Falling back to torch.float16 for compatibility.
(APIServer pid=65016) WARNING 05-25 21:22:19 [model.py:2035] Casting torch.bfloat16 to torch.float16.
(APIServer pid=65016) INFO 05-25 21:22:19 [model.py:1697] Using max model len 131072
(APIServer pid=65016) INFO 05-25 21:22:19 [config.py:101] Gemma4 model has heterogeneous head dimensions (head_dim=256, global_head_dim=512). Forcing TRITON_ATTN backend to prevent mixed-backend numerical divergence.
(APIServer pid=65016) INFO 05-25 21:22:19 [vllm.py:886] Asynchronous scheduling is enabled.
(APIServer pid=65016) INFO 05-25 21:22:19 [kernel.py:212] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native'])
tokenizer_config.json: 2.10kB [00:00, 2.10MB/s]
tokenizer.json: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 32.2M/32.2M [00:02<00:00, 11.8MB/s]
chat_template.jinja: 17.3kB [00:00, 13.0MB/s]
generation_config.json: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 208/208 [00:00<00:00, 1.66MB/s]
(EngineCore pid=65068) INFO 05-25 21:23:14 [core.py:109] Initializing a V1 LLM engine (v0.21.0) with config: model='google/gemma-4-E4B-it', speculative_config=None, tokenizer='google/gemma-4-E4B-it', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=131072, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=False, quantization=None, quantization_config=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=0, served_model_name=google/gemma-4-E4B-it, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'ir_enable_torch_wrap': True, 'splitting_ops': ['vllm::unified_attention_with_output', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::gdn_attention_core_xpu', 'vllm::olmo_hybrid_gdn_full_forward', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::deepseek_v4_attention', 'vllm::unified_kv_cache_update', 'vllm::unified_mla_kv_cache_update'], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_vision_items_per_batch': 0, 'encoder_cudagraph_max_frames_per_batch': None, 'compile_sizes': [], 'compile_ranges_endpoints': [2048], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496, 512], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 512, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': False, 'static_all_moe_layers': []}, kernel_config=KernelConfig(ir_op_priority=IrOpPriorityConfig(rms_norm=['native'], fused_add_rms_norm=['native']), enable_flashinfer_autotune=False, moe_backend='auto')
(EngineCore pid=65068) Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
(EngineCore pid=65068) /home/minimonk/.local/lib/python3.10/site-packages/torch/cuda/__init__.py:371: UserWarning: Found GPU0 NVIDIA GeForce GTX 1080 Ti which is of compute capability (CC) 6.1.
(EngineCore pid=65068) The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:
(EngineCore pid=65068) - 7.5 which supports hardware CC >=7.5,<8.0
(EngineCore pid=65068) - 8.0 which supports hardware CC >=8.0,<9.0 except {8.7}
(EngineCore pid=65068) - 8.6 which supports hardware CC >=8.6,<9.0 except {8.7}
(EngineCore pid=65068) - 9.0 which supports hardware CC >=9.0,<10.0
(EngineCore pid=65068) - 10.0 which supports hardware CC >=10.0,<11.0 except {10.1}
(EngineCore pid=65068) - 12.0 which supports hardware CC >=12.0,<13.0
(EngineCore pid=65068) Please follow the instructions at https://pytorch.org/get-started/locally/ to install a PyTorch release that supports one of these CUDA versions: 12.6
(EngineCore pid=65068)   _warn_unsupported_code(d, device_cc, code_ccs)
(EngineCore pid=65068) /home/minimonk/.local/lib/python3.10/site-packages/torch/cuda/__init__.py:371: UserWarning: Found GPU1 NVIDIA GeForce GTX 1080 Ti which is of compute capability (CC) 6.1.
(EngineCore pid=65068) The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:
(EngineCore pid=65068) - 7.5 which supports hardware CC >=7.5,<8.0
(EngineCore pid=65068) - 8.0 which supports hardware CC >=8.0,<9.0 except {8.7}
(EngineCore pid=65068) - 8.6 which supports hardware CC >=8.6,<9.0 except {8.7}
(EngineCore pid=65068) - 9.0 which supports hardware CC >=9.0,<10.0
(EngineCore pid=65068) - 10.0 which supports hardware CC >=10.0,<11.0 except {10.1}
(EngineCore pid=65068) - 12.0 which supports hardware CC >=12.0,<13.0
(EngineCore pid=65068) Please follow the instructions at https://pytorch.org/get-started/locally/ to install a PyTorch release that supports one of these CUDA versions: 12.6
(EngineCore pid=65068)   _warn_unsupported_code(d, device_cc, code_ccs)
(EngineCore pid=65068) /home/minimonk/.local/lib/python3.10/site-packages/torch/cuda/__init__.py:489: UserWarning:
(EngineCore pid=65068) NVIDIA GeForce GTX 1080 Ti with CUDA capability sm_61 is not compatible with the current PyTorch installation.
(EngineCore pid=65068) The current PyTorch install supports CUDA capabilities sm_75 sm_80 sm_86 sm_90 sm_100 sm_120.
(EngineCore pid=65068) If you want to use the NVIDIA GeForce GTX 1080 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/
(EngineCore pid=65068)
(EngineCore pid=65068)   queued_call()
(EngineCore pid=65068) INFO 05-25 21:23:20 [parallel_state.py:1410] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.40.238:39589 backend=nccl
(EngineCore pid=65068) INFO 05-25 21:23:20 [parallel_state.py:1723] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A
(EngineCore pid=65068) WARNING 05-25 21:23:21 [topk_topp_sampler.py:61] FlashInfer top-p/top-k sampling not supported on compute capability 6.1; falling back to PyTorch-native sampler. Set VLLM_USE_FLASHINFER_SAMPLER=0 to silence.
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140] EngineCore failed to start.
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140] Traceback (most recent call last):
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 1114, in run_engine_core
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     engine_core = EngineCoreProc(*args, engine_index=dp_rank, **kwargs)
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     return func(*args, **kwargs)
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 880, in __init__
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     super().__init__(
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 118, in __init__
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.model_executor = executor_class(vllm_config)
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     return func(*args, **kwargs)
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/executor/abstract.py", line 109, in __init__
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self._init_executor()
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/executor/uniproc_executor.py", line 60, in _init_executor
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.driver_worker.init_device()
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/worker_base.py", line 317, in init_device
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.worker.init_device()  # type: ignore
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     return func(*args, **kwargs)
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_worker.py", line 330, in init_device
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.model_runner = GPUModelRunnerV1(self.vllm_config, self.device)
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_model_runner.py", line 629, in __init__
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.input_batch = InputBatch(
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_input_batch.py", line 171, in __init__
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.block_table = MultiGroupBlockTable(
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 267, in __init__
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.block_tables = [
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 268, in <listcomp>
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     BlockTable(
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 70, in __init__
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.block_table = self._make_buffer(
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 218, in _make_buffer
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     return CpuGpuBuffer(
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/utils.py", line 120, in __init__
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]     self.gpu = torch.zeros_like(self.cpu, device=device)
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140] torch.AcceleratorError: CUDA error: no kernel image is available for execution on the device
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140] Search for `cudaErrorNoKernelImageForDevice' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140] CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140] For debugging consider passing CUDA_LAUNCH_BLOCKING=1
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140] Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
(EngineCore pid=65068) ERROR 05-25 21:23:21 [core.py:1140]
(EngineCore pid=65068) Process EngineCore:
(EngineCore pid=65068) Traceback (most recent call last):
(EngineCore pid=65068)   File "/usr/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
(EngineCore pid=65068)     self.run()
(EngineCore pid=65068)   File "/usr/lib/python3.10/multiprocessing/process.py", line 108, in run
(EngineCore pid=65068)     self._target(*self._args, **self._kwargs)
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 1144, in run_engine_core
(EngineCore pid=65068)     raise e
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 1114, in run_engine_core
(EngineCore pid=65068)     engine_core = EngineCoreProc(*args, engine_index=dp_rank, **kwargs)
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=65068)     return func(*args, **kwargs)
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 880, in __init__
(EngineCore pid=65068)     super().__init__(
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 118, in __init__
(EngineCore pid=65068)     self.model_executor = executor_class(vllm_config)
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=65068)     return func(*args, **kwargs)
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/executor/abstract.py", line 109, in __init__
(EngineCore pid=65068)     self._init_executor()
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/executor/uniproc_executor.py", line 60, in _init_executor
(EngineCore pid=65068)     self.driver_worker.init_device()
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/worker_base.py", line 317, in init_device
(EngineCore pid=65068)     self.worker.init_device()  # type: ignore
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(EngineCore pid=65068)     return func(*args, **kwargs)
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_worker.py", line 330, in init_device
(EngineCore pid=65068)     self.model_runner = GPUModelRunnerV1(self.vllm_config, self.device)
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_model_runner.py", line 629, in __init__
(EngineCore pid=65068)     self.input_batch = InputBatch(
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/gpu_input_batch.py", line 171, in __init__
(EngineCore pid=65068)     self.block_table = MultiGroupBlockTable(
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 267, in __init__
(EngineCore pid=65068)     self.block_tables = [
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 268, in <listcomp>
(EngineCore pid=65068)     BlockTable(
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 70, in __init__
(EngineCore pid=65068)     self.block_table = self._make_buffer(
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/worker/block_table.py", line 218, in _make_buffer
(EngineCore pid=65068)     return CpuGpuBuffer(
(EngineCore pid=65068)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/utils.py", line 120, in __init__
(EngineCore pid=65068)     self.gpu = torch.zeros_like(self.cpu, device=device)
(EngineCore pid=65068) torch.AcceleratorError: CUDA error: no kernel image is available for execution on the device
(EngineCore pid=65068) Search for `cudaErrorNoKernelImageForDevice' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.
(EngineCore pid=65068) CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
(EngineCore pid=65068) For debugging consider passing CUDA_LAUNCH_BLOCKING=1
(EngineCore pid=65068) Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
(EngineCore pid=65068)
[rank0]:[W525 21:23:22.743241268 ProcessGroupNCCL.cpp:1575] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
(APIServer pid=65016) Traceback (most recent call last):
(APIServer pid=65016)   File "/home/minimonk/.local/bin/vllm", line 8, in <module>
(APIServer pid=65016)     sys.exit(main())
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/cli/main.py", line 92, in main
(APIServer pid=65016)     args.dispatch_function(args)
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/cli/serve.py", line 122, in cmd
(APIServer pid=65016)     uvloop.run(run_server(args))
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/uvloop/__init__.py", line 69, in run
(APIServer pid=65016)     return loop.run_until_complete(wrapper())
(APIServer pid=65016)   File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/uvloop/__init__.py", line 48, in wrapper
(APIServer pid=65016)     return await main
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 693, in run_server
(APIServer pid=65016)     await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 707, in run_server_worker
(APIServer pid=65016)     async with build_async_engine_client(
(APIServer pid=65016)   File "/usr/lib/python3.10/contextlib.py", line 199, in __aenter__
(APIServer pid=65016)     return await anext(self.gen)
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 100, in build_async_engine_client
(APIServer pid=65016)     async with build_async_engine_client_from_engine_args(
(APIServer pid=65016)   File "/usr/lib/python3.10/contextlib.py", line 199, in __aenter__
(APIServer pid=65016)     return await anext(self.gen)
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/entrypoints/openai/api_server.py", line 136, in build_async_engine_client_from_engine_args
(APIServer pid=65016)     async_llm = AsyncLLM.from_vllm_config(
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/async_llm.py", line 217, in from_vllm_config
(APIServer pid=65016)     return cls(
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/async_llm.py", line 146, in __init__
(APIServer pid=65016)     self.engine_core = EngineCoreClient.make_async_mp_client(
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(APIServer pid=65016)     return func(*args, **kwargs)
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core_client.py", line 130, in make_async_mp_client
(APIServer pid=65016)     return AsyncMPClient(*client_args)
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper
(APIServer pid=65016)     return func(*args, **kwargs)
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core_client.py", line 900, in __init__
(APIServer pid=65016)     super().__init__(
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/core_client.py", line 535, in __init__
(APIServer pid=65016)     with launch_core_engines(
(APIServer pid=65016)   File "/usr/lib/python3.10/contextlib.py", line 142, in __exit__
(APIServer pid=65016)     next(self.gen)
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/utils.py", line 1128, in launch_core_engines
(APIServer pid=65016)     wait_for_engine_startup(
(APIServer pid=65016)   File "/home/minimonk/.local/lib/python3.10/site-packages/vllm/v1/engine/utils.py", line 1187, in wait_for_engine_startup
(APIServer pid=65016)     raise RuntimeError(
(APIServer pid=65016) RuntimeError: Engine core initialization failed. See root cause above. Failed core proc(s): {}
저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

telegram bot api 로 기능 변경하기  (0) 2026.05.25
python huggingface 저장경로 변경하기  (0) 2026.05.24
antigravity gemini flash 할당량이 깃털 같구만?  (0) 2026.05.22
openclaw agent 구성 관련  (0) 2026.05.22
mxfp4  (0) 2026.05.22
Posted by 구차니
모종의 음모/ai 프로그램2026. 5. 22. 22:27

antigravity gemini flash 할당량이 깃털 같구만?

한번 찾아보니 가격 3배 이상, 토큰 3배 증가

사실상 9배 증가라는 소문이 있네

 

[링크 : https://wikidocs.net/blog/@jaehong/13849/]

 

근데 순식간에(?) geminu 3.0 flash 를 날리고 gemini 3.5 flash + 3.1 pro 를 같은 할당량으로 묶는건 너무 개매너 아니냐 -_-

[링크 : https://ai.google.dev/gemini-api/docs/pricing?hl=ko]

 

+

2026.05.29

느낌상 질문 하나당 한칸 줄어드는 느낌..

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

python huggingface 저장경로 변경하기  (0) 2026.05.24
vllm 설치, 실행 실패 -> 사실상 포기  (0) 2026.05.24
openclaw agent 구성 관련  (0) 2026.05.22
mxfp4  (0) 2026.05.22
텔레그램 봇 - 메시지 관련  (0) 2026.05.22
Posted by 구차니
모종의 음모/ai 프로그램2026. 5. 22. 12:44

openclaw agent 구성 관련

음.. local LLM이 하나 뿐이면 느려서 어떻게 구성해야 할까? ㅜㅠ

 

[링크 : https://brunch.co.kr/@bbt/53]

[링크 : https://docs.openclaw.ai/ko/concepts/multi-agent]

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

vllm 설치, 실행 실패 -> 사실상 포기  (0) 2026.05.24
antigravity gemini flash 할당량이 깃털 같구만?  (0) 2026.05.22
mxfp4  (0) 2026.05.22
텔레그램 봇 - 메시지 관련  (0) 2026.05.22
openclaw 설치 telegram 신기하네?  (0) 2026.05.22
Posted by 구차니
모종의 음모/ai 프로그램2026. 5. 22. 12:38

mxfp4

microscale(mx) fp4

[링크 : https://junhan-ai.tistory.com/602]

[링크 : https://www.themoonlight.io/ko/review/mx-pushing-the-limits-of-microscaling-formats-for-efficient-large-language-model-serving]

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

antigravity gemini flash 할당량이 깃털 같구만?  (0) 2026.05.22
openclaw agent 구성 관련  (0) 2026.05.22
텔레그램 봇 - 메시지 관련  (0) 2026.05.22
openclaw 설치 telegram 신기하네?  (0) 2026.05.22
wan2.2 + comfyui 시도  (0) 2026.05.21
Posted by 구차니
모종의 음모/ai 프로그램2026. 5. 22. 11:41

텔레그램 봇 - 메시지 관련

inlinebutton - 메시지 하단에 생기는 버튼

replybutton - 사용자 입력창 하단에 생기는 버튼

[링크 : https://naimjae.tistory.com/11]

 

메시지 서식(markdown 등)

[링크 : https://naimjae.tistory.com/8]

 

[링크 : https://docs.python-telegram-bot.org/en/stable/telegram.menubuttoncommands.html]

    [링크 : https://docs.python-telegram-bot.org/en/stable/telegram.html]

 

+

2026.05.25

Bot settings

set_my_commands() Used for setting the list of commands
delete_my_commands() Used for deleting the list of commands
get_my_commands() Used for obtaining the list of commands
get_my_default_administrator_rights() Used for obtaining the default administrator rights for the bot
set_my_default_administrator_rights() Used for setting the default administrator rights for the bot
get_chat_menu_button() Used for obtaining the menu button of a private chat or the default menu button
set_chat_menu_button() Used for setting the menu button of a private chat or the default menu button
set_my_description() Used for setting the description of the bot
get_my_description() Used for obtaining the description of the bot
set_my_short_description() Used for setting the short description of the bot
get_my_short_description() Used for obtaining the short description of the bot
set_my_name() Used for setting the name of the bot
get_my_name() Used for obtaining the name of the bot
set_my_profile_photo() Used for setting the profile photo of the bot
remove_my_profile_photo() Used for removing the profile photo of the bot

[링크 : https://docs.python-telegram-bot.org/en/stable/telegram.bot.html]

저작자표시 (새창열림)

'모종의 음모 > ai 프로그램' 카테고리의 다른 글

openclaw agent 구성 관련  (0) 2026.05.22
mxfp4  (0) 2026.05.22
openclaw 설치 telegram 신기하네?  (0) 2026.05.22
wan2.2 + comfyui 시도  (0) 2026.05.21
wan2.2 + comfyui  (0) 2026.05.19
Posted by 구차니

티스토리툴바