Misc. bug: ValueError: Duplicated key name 'deepseek2.attention.key_length' #14093
Description
Name and Version
We need a fp16 DeepSeeek-R1 model, thus I have converted the fp8 Deepseek-R1 model to the bf16 model using the fp8_cast_bf16.py script(from the "inference" folder of deepseek-ai/DeepSeek-V3).
When using the convert_hf_to_gguf.py script to convert the bf16 model to fp16 model, I have met the following error:
Traceback (most recent call last):
File "/mnt/host_root/home/jovyan/lich/llama.cpp/convert_hf_to_gguf.py", line 6102, in <module>
main()
File "/mnt/host_root/home/jovyan/lich/llama.cpp/convert_hf_to_gguf.py", line 6096, in main
model_instance.write()
File "/mnt/host_root/home/jovyan/lich/llama.cpp/convert_hf_to_gguf.py", line 402, in write
self.prepare_metadata(vocab_only=False)
File "/mnt/host_root/home/jovyan/lich/llama.cpp/convert_hf_to_gguf.py", line 486, in prepare_metadata
super().prepare_metadata(vocab_only=vocab_only)
File "/mnt/host_root/home/jovyan/lich/llama.cpp/convert_hf_to_gguf.py", line 392, in prepare_metadata
self.set_gguf_parameters()
File "/mnt/host_root/home/jovyan/lich/llama.cpp/convert_hf_to_gguf.py", line 4834, in set_gguf_parameters
self.gguf_writer.add_key_length(hparams["kv_lora_rank"] + hparams["qk_rope_head_dim"])
File "/mnt/host_root/home/jovyan/lich/llama.cpp/gguf-py/gguf/gguf_writer.py", line 687, in add_key_length
self.add_uint32(Keys.Attention.KEY_LENGTH.format(arch=self.arch), length)
File "/mnt/host_root/home/jovyan/lich/llama.cpp/gguf-py/gguf/gguf_writer.py", line 290, in add_uint32
self.add_key_value(key, val, GGUFValueType.UINT32)
File "/mnt/host_root/home/jovyan/lich/llama.cpp/gguf-py/gguf/gguf_writer.py", line 273, in add_key_value
raise ValueError(f'Duplicated key name {key!r}')
ValueError: Duplicated key name 'deepseek2.attention.key_length'
Any idea about this? Can I just skip this check?
Thanks!
Operating systems
No response
Which llama.cpp modules do you know to be affected?
No response
Command line
python convert_hf_to_gguf.py /mnt/host_root/home/lich/gguf/DeepSeek-R1-bf16 --outfile /mnt/host_root/home/models/DeepSeek-R1-fp16.gguf --outtype f16Problem description & steps to reproduce
python fp8_cast_bf16.py --input-fp8-hf-path /home/admin/models/deepseek-ai/DeepSeek-R1/ --output-bf16-hf-path /mnt/host_root/home/lich/gguf/DeepSeek-R1-bf16
python convert_hf_to_gguf.py /mnt/host_root/home/lich/gguf/DeepSeek-R1-bf16 --outfile /mnt/host_root/home/models/DeepSeek-R1-fp16.gguf --outtype f16
First Bad Commit
No response