Version 0.42.0 New Features EasyDelState,`fully automated loading…

… and converting model`

lib/python/EasyDel/__init__.py

@@ -114,39 +114,27 @@
  get_mem as get_mem
 )
 
-from .transform.llama import (
- llama_from_pretrained as llama_from_pretrained,
- llama_convert_flax_to_pt as llama_convert_flax_to_pt,
- llama_convert_hf_to_flax_load as llama_convert_hf_to_flax_load,
- llama_convert_hf_to_flax as llama_convert_hf_to_flax,
- llama_easydel_to_hf as llama_easydel_to_hf
-)
-from .transform.mpt import (
- mpt_convert_flax_to_pt_1b as mpt_convert_flax_to_pt_1b,
- mpt_convert_pt_to_flax_1b as mpt_convert_pt_to_flax_1b,
- mpt_convert_pt_to_flax_7b as mpt_convert_pt_to_flax_7b,
- mpt_convert_flax_to_pt_7b as mpt_convert_flax_to_pt_7b,
- mpt_from_pretrained as mpt_from_pretrained
-)
-
-from .transform.falcon import (
- falcon_convert_pt_to_flax_7b as falcon_convert_pt_to_flax_7b,
+from .transform import (
+ huggingface_to_easydel as huggingface_to_easydel,
+ easystate_to_huggingface_model as easystate_to_huggingface_model,
+ easystate_to_torch as easystate_to_torch,
  falcon_convert_flax_to_pt_7b as falcon_convert_flax_to_pt_7b,
  falcon_from_pretrained as falcon_from_pretrained,
  falcon_convert_hf_to_flax as falcon_convert_hf_to_flax,
- falcon_easydel_to_hf as falcon_easydel_to_hf
-)
-from .transform.mistral import (
- mistral_convert_hf_to_flax as mistral_convert_hf_to_flax,
+ mpt_convert_pt_to_flax_1b as mpt_convert_pt_to_flax_1b,
+ mpt_convert_pt_to_flax_7b as mpt_convert_pt_to_flax_7b,
+ mpt_convert_flax_to_pt_7b as mpt_convert_flax_to_pt_7b,
+ mpt_from_pretrained as mpt_from_pretrained,
  mistral_convert_hf_to_flax_load as mistral_convert_hf_to_flax_load,
  mistral_convert_flax_to_pt as mistral_convert_flax_to_pt,
  mistral_from_pretrained as mistral_from_pretrained,
- mistral_convert_pt_to_flax as mistral_convert_pt_to_flax,
- mistral_easydel_to_hf as mistral_easydel_to_hf
-)
-
-from .transform.easydel_transform import (
- huggingface_to_easydel as huggingface_to_easydel
+ falcon_convert_pt_to_flax_7b as falcon_convert_pt_to_flax_7b,
+ mistral_convert_hf_to_flax as mistral_convert_hf_to_flax,
+ mpt_convert_flax_to_pt_1b as mpt_convert_flax_to_pt_1b,
+ llama_convert_flax_to_pt as llama_convert_flax_to_pt,
+ llama_convert_hf_to_flax_load as llama_convert_hf_to_flax_load,
+ llama_convert_hf_to_flax as llama_convert_hf_to_flax,
+ llama_from_pretrained as llama_from_pretrained
 )
 from .etils import (
  EasyDelOptimizers as EasyDelOptimizers,

lib/python/EasyDel/transform/__init__.py

@@ -1,10 +1,31 @@
-from .llama import llama_from_pretrained, llama_convert_flax_to_pt, llama_convert_hf_to_flax_load, \
- llama_convert_hf_to_flax, llama_easydel_to_hf
-from .mpt import mpt_convert_flax_to_pt_1b, mpt_convert_pt_to_flax_1b, mpt_convert_pt_to_flax_7b, \
- mpt_convert_flax_to_pt_7b, mpt_from_pretrained
-from .falcon import falcon_convert_pt_to_flax_7b, falcon_convert_flax_to_pt_7b, falcon_from_pretrained, \
- falcon_convert_hf_to_flax, falcon_easydel_to_hf
-from .mistral import mistral_convert_hf_to_flax, mistral_convert_hf_to_flax_load, \
- mistral_convert_flax_to_pt, \
- mistral_from_pretrained, mistral_convert_pt_to_flax, mistral_easydel_to_hf
-from .easydel_transform import easystate_to_huggingface_model, easystate_to_torch
+from .llama import (
+ llama_from_pretrained,
+ llama_convert_flax_to_pt,
+ llama_convert_hf_to_flax_load,
+ llama_convert_hf_to_flax,
+)
+from .mpt import (
+ mpt_convert_flax_to_pt_1b,
+ mpt_convert_pt_to_flax_1b,
+ mpt_convert_pt_to_flax_7b,
+ mpt_convert_flax_to_pt_7b,
+ mpt_from_pretrained
+)
+from .falcon import (
+ falcon_convert_pt_to_flax_7b,
+ falcon_convert_flax_to_pt_7b,
+ falcon_from_pretrained,
+ falcon_convert_hf_to_flax,
+)
+from .mistral import (
+ mistral_convert_hf_to_flax,
+ mistral_convert_hf_to_flax_load,
+ mistral_convert_flax_to_pt,
+ mistral_from_pretrained,
+)
+
+from .easydel_transform import (
+ huggingface_to_easydel,
+ easystate_to_huggingface_model,
+ easystate_to_torch
+)

lib/python/EasyDel/transform/falcon.py

@@ -147,17 +147,6 @@ def falcon_convert_flax_to_pt_7b(state_dict_flax, num_hidden_layers: int, device
  return state_dict
 
 
-def falcon_easydel_to_hf(path, config: FalconConfig):
- """
- Takes path to easydel saved ckpt and return the model in pytorch (Transformers Huggingface)
- """
- torch_params = load_and_convert_checkpoint_to_torch(path)
- edited_params = {}
- for k, v in torch_params.items():
- edited_params[k.replace('.kernel', '.weight').replace('.embedding', '.weight')] = v
- model = FalconForCausalLM(config=config)
- model.load_state_dict(edited_params)
- return model
 
 
 def falcon_convert_hf_to_flax(state_dict: Dict[str, torch.Tensor], config: FalconConfig, device):