Using convert.py with a fine tuned phi-2

[FiveTechSoft]

We are loading phi-2 from HF using load_in_8bit=True and torch_dtype=torch.float16, then we fine tune it and finally we save it locally.

When running convert.py ./phi-2 we get this error:
File "/content/convert.py", line 764, in convert
data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
KeyError: 'I8'

If we try the same using load_in_8bit=False then we get:
File "/content/convert.py", line 257, in loadHFTransformerJson
f_norm_eps = config["rms_norm_eps"],
KeyError: 'rms_norm_eps'

how to generate a GGUF from a fine tuned phi-2 ? many thanks

[Rishbah-76]

Having the same kind of issue with falcon finetuned 7b bf16 model
""
ggllm.cpp/convert.py", line 761, in convert
data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
KeyError: 'BF16'
""

[MotorCityCobra]

Do you have reason to believe it is supported?

[apepkuss]

Same issue with phi-2.

[compilade]

how to generate a GGUF from a fine tuned phi-2 ?

I suggest using convert-hf-to-gguf.py (which is where conversion from phi-2 models was implemented) instead of convert.py.

When new models come out on HuggingFace, their conversion is usually added in convert-hf-to-gguf.py (the "hf" in there stands for HuggingFace, I think), probably because it allows sharing a lot of metadata-loading code (especially regarding tokenizers, but also tensor names). It's also more obvious (to me, at least) where to add support for a new model in convert-hf-to-gguf.py than in convert.py. So that's possibly why convert.py doesn't also support converting phi-2.

[apepkuss]

how to generate a GGUF from a fine tuned phi-2 ?

I suggest using convert-hf-to-gguf.py (which is where conversion from phi-2 models was implemented) instead of convert.py.

When new models come out on HuggingFace, their conversion is usually added in convert-hf-to-gguf.py (the "hf" in there stands for HuggingFace, I think), probably because it allows sharing a lot of metadata-loading code (especially regarding tokenizers, but also tensor names). It's also more obvious (to me, at least) where to add support for a new model in convert-hf-to-gguf.py than in convert.py. So that's possibly why convert.py doesn't also support converting phi-2.

Works for me. Thanks a lot!

[tgalery]

I have a similar issue, after fine-tuning, I get KeyError: 'n_layer' (line 1064 of convert-hf-to-gguf.py), otherwise, I get KeyError: 'rms_norm_eps' (line 257 of convert.py).

Doing some more debugging in the convert-hf-to-gguf.py file, it seems that the config I get for the Phi2 model is very different from the expected one. The config from my fine tuned model is:

{
	'_name_or_path': 'microsoft/phi-2',
	'architectures': ['PhiForCausalLM'],
	'attention_dropout': 0.0,
	'auto_map': {
		'AutoConfig': 'microsoft/phi-2--configuration_phi.PhiConfig',
		'AutoModelForCausalLM': 'microsoft/phi-2--modeling_phi.PhiForCausalLM'
	},
	'bos_token_id': 50256,
	'embd_pdrop': 0.0,
	'eos_token_id': 50256,
	'hidden_act': 'gelu_new',
	'hidden_size': 2560,
	'initializer_range': 0.02,
	'intermediate_size': 10240,
	'layer_norm_eps': 1e-05,
	'max_position_embeddings': 2048,
	'model_type': 'phi',
	'num_attention_heads': 32,
	'num_hidden_layers': 32,
	'num_key_value_heads': 32,
	'partial_rotary_factor': 0.4,
	'qk_layernorm': False,
	'resid_pdrop': 0.1,
	'rope_scaling': None,
	'rope_theta': 10000.0,
	'tie_word_embeddings': False,
	'torch_dtype': 'bfloat16',
	'transformers_version': '4.37.2',
	'use_cache': True,
	'vocab_size': 51200
}

Whereas Phi2Model.set_gguf_parameters, expects keys like n_layer, n_head, n_positions, n_embed, layer_norm_epsilon, and rotary_dim.

If it's mapping old keys to new ones, I'm happy to work on a MR, but there seems to be some info lost there.

[compilade]

I have a similar issue, after fine-tuning, I get KeyError: 'n_layer' (line 1064 of convert-hf-to-gguf.py), otherwise, I get KeyError: 'rms_norm_eps' (line 257 of convert.py).

@tgalery The line numbers in the errors you got seem different than those in the latest commit.

Whereas Phi2Model.set_gguf_parameters, expects keys like n_layer, n_head, n_positions, n_embed, layer_norm_epsilon, and rotary_dim.

If it's mapping old keys to new ones, I'm happy to work on a MR, but there seems to be some info lost there.

The fix you're describing is already implemented, but maybe your local checkout of llama.cpp hasn't been updated in a while. Converting Phi-2 with convert-hf-to-gguf.py with such a config.json should work as of at least #4903 (which was merged on January 13), since it checks for num_hidden_layers (which seems to exist in your config.json) before falling back to n_layer.

This is what you should see in convert-hf-to-gguf.py if your source tree of llama.cpp is recent enough:

llama.cpp/convert-hf-to-gguf.py

Lines 1287 to 1296 in 5cdb371

	@Model.register("PhiForCausalLM")
	class Phi2Model(Model):
	model_arch = gguf.MODEL_ARCH.PHI2
	def set_gguf_parameters(self):
	block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
	rot_pct = self.find_hparam(["partial_rotary_factor"])
	n_embd = self.find_hparam(["hidden_size", "n_embd"])
	n_head = self.find_hparam(["num_attention_heads", "n_head"])

If it still doesn't work with the latest version, please do tell. Hope this helps :)

[tgalery]

Oh I see, llama.ccp was pulled inside a python repo via git submodules, will update that and try again. Many thanks.

[tgalery]

@compilade thanks for the explanation. It works like a charm. Just a question, I'm working on a pipeline and for some model types, say Mistral-7B, one needs to use the convert.py script. For some others, say phi-2, we need to use convert-hf-to-gguf.py . Is there a plan to unify these ?

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.