wassname/alpaca_convert
1
0
Fork 0
mirror of https://github.com/wassname/alpaca_convert.git synced 2026-06-27 17:47:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

misc

Browse Source
This commit is contained in:
wassname
2023-04-22 00:24:57 +00:00
parent 4155cbb641
commit 2e7bf22d4a
2 changed files with 9 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
mjc_notes.md
+5 -4
View File
@@ -55,10 +55,10 @@ python scripts/download-model.py tloen/alpaca-lora-7b
python scripts/download-model.py decapoda-research/llama-7b-hf python scripts/download-model.py decapoda-research/llama-7b-hf
# convert # convert
python scripts/export_hf_checkpoint.py ./models/decapoda-research_llama-7b-hf -l loras/tloen_alpaca-lora-7b python scripts/01_export_hf_checkpoint.py ./models/decapoda-research_llama-7b-hf -l loras/tloen_alpaca-lora-7b
python scripts/export_hf_checkpoint.py ./models/decapoda-research_llama-13b-hf -l loras/chansung_alpaca-lora-13b # crash! 50GB+ needed python scripts/01_export_hf_checkpoint.py ./models/decapoda-research_llama-13b-hf -l loras/chansung_alpaca-lora-13b # crash! 50GB+ needed
python scripts/export_hf_checkpoint.py ./models/decapoda-research_llama-30b-hf -l loras/chansung_alpaca-lora-30b python scripts/01_export_hf_checkpoint.py ./models/decapoda-research_llama-30b-hf -l loras/chansung_alpaca-lora-30b
python scripts/export_hf_checkpoint.py ./models/decapoda-research_llama-60b-hf -l loras/chansung_alpaca-lora-60b python scripts/01_export_hf_checkpoint.py ./models/decapoda-research_llama-60b-hf -l loras/chansung_alpaca-lora-60b
# test # test
python scripts/test_01_delora.py models/tloen_alpaca-lora-7b-delorified python scripts/test_01_delora.py models/tloen_alpaca-lora-7b-delorified
@@ -66,6 +66,7 @@ python scripts/test_01_delora.py models/chansung_alpaca-lora-13b-delorified
# now compare what was generated during conversion `test_prompts.txt`, to the loaded version # now compare what was generated during conversion `test_prompts.txt`, to the loaded version
# 4bit... # 4bit...
CUDA_VISIBLE_DEVICES=0 python llama.py ./models/tloen_alpaca-lora-7b-delorified c4 --wbits 4 --true-sequential --act-order --groupsize 128 --save_safetensors ./models/tloen_alpaca-lora-7b-delorified/llama7b-4bit-128g.safetensors
# ggml conversion... # ggml conversion...
``` ```
scripts/export_hf_checkpoint.py → scripts/01_export_hf_checkpoint.py
+4 -18
View File
@@ -14,17 +14,6 @@ def main(BASE_MODEL, LORA_MODEL, output_path=None):
if output_path is None: if output_path is None:
output_path = 'models/' + LORA_MODEL.split('/')[-1] + '-delorified' output_path = 'models/' + LORA_MODEL.split('/')[-1] + '-delorified'
# BASE_MODEL = os.environ.get("BASE_MODEL", None)
# assert (
# BASE_MODEL
# ), "Please specify a value for BASE_MODEL environment variable, e.g. `export BASE_MODEL=huggyllama/llama-7b`" # noqa: E501
# LORA_MODEL = os.environ.get("BASE_MODEL", None)
# assert (
# LORA_MODEL
# ), "Please specify a value for LORA_MODEL environment variable, e.g. `export BASE_MODEL=tloen/alpaca-lora-7b`" # noqa: E501
tokenizer = LlamaTokenizer.from_pretrained(BASE_MODEL) tokenizer = LlamaTokenizer.from_pretrained(BASE_MODEL)
base_model = LlamaForCausalLM.from_pretrained( base_model = LlamaForCausalLM.from_pretrained(
@@ -33,8 +22,6 @@ def main(BASE_MODEL, LORA_MODEL, output_path=None):
torch_dtype=torch.float16, torch_dtype=torch.float16,
device_map={"": "cpu"}, device_map={"": "cpu"},
) )
# TODO or load 4 bit?
first_weight = base_model.model.layers[0].self_attn.q_proj.weight first_weight = base_model.model.layers[0].self_attn.q_proj.weight
first_weight_old = first_weight.clone() first_weight_old = first_weight.clone()
@@ -70,9 +57,8 @@ def main(BASE_MODEL, LORA_MODEL, output_path=None):
LlamaForCausalLM.save_pretrained( LlamaForCausalLM.save_pretrained(
base_model, output_path, state_dict=deloreanized_sd, max_shard_size="400MB" base_model, output_path, state_dict=deloreanized_sd, max_shard_size="400MB"
) )
print(f'output {output_path}')
LlamaTokenizer.save_pretrained(tokenizer, output_path) LlamaTokenizer.save_pretrained(tokenizer, output_path)
# FIXME also save tokenizer print(f'output {output_path}')
from alpaca_convert.test import test_conversation from alpaca_convert.test import test_conversation
o = test_conversation(lora_model.float(), tokenizer) o = test_conversation(lora_model.float(), tokenizer)
@@ -83,10 +69,10 @@ def main(BASE_MODEL, LORA_MODEL, output_path=None):
if __name__=="__main__": if __name__=="__main__":
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('model', type=str) parser.add_argument('model', type=str, help='Huggingface model e.g. `./models/decapoda-research_llama-7b-hf`')
parser.add_argument('-l', '--lora', type=str, default='main', help='Lora repo or path e.g. `tloen/alpaca-lora-7b`') parser.add_argument('-l', '--lora', type=str, default='main', help='Lora repo or path e.g. `tloen/alpaca-lora-7b` or `./loras/tloen_alpaca-lora-7b`')
parser.add_argument('-o', '--output', type=Path, default=None) parser.add_argument('-o', '--output', type=Path, default=None)
"e.g. ./hf_ckpt. default will be lora name" "if None default will be `./models/tloen_alpaca-lora-7b_delorafied` or similar"
args = parser.parse_args() args = parser.parse_args()
main(args.model, args.lora, args.output) main(args.model, args.lora, args.output)