Skip to content

fix: restore torch_dtype when loading layerwise reranker model #1589

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
St4r4x wants to merge 1 commit into
base: master
Choose a base branch
from
Open

fix: restore torch_dtype when loading layerwise reranker model #1589

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 8 additions & 4 deletions FlagEmbedding/finetune/reranker/decoder_only/layerwise/load_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,9 @@
import os
import re
import logging
import torch
from torch import nn
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
from transformers import AutoConfig, AutoTokenizer
from peft import LoraConfig, TaskType, get_peft_model, PeftModel

from FlagEmbedding.finetune.reranker.decoder_only.layerwise.arguments import RerankerModelArguments
Expand Down Expand Up @@ -38,11 +39,14 @@ def find_largest_checkpoint(checkpoint_dir):
return None


def get_model(model_args: RerankerModelArguments, only_for_one_logit):
def get_model(model_args: RerankerModelArguments, only_for_one_logit, training_args):
"""Get the model.

Args:
model_args (RerankerModelArguments): Model arguments instance.
only_for_one_logit: Index of the logit to keep when the classifier head is not multi-head.
training_args: Training arguments instance, used to pick the torch dtype (fp16/bf16) the
model weights are loaded in.

Returns:
transformers.PreTrainedModel or PeftModel: The loaded model.
Expand Down Expand Up @@ -76,7 +80,7 @@ def get_model(model_args: RerankerModelArguments, only_for_one_logit):
model = LayerWiseMiniCPMForCausalLM.from_pretrained(
model_args.model_name_or_path,
trust_remote_code=model_args.trust_remote_code,
# torch_dtype=torch.float16 if training_args.fp16 else torch.bfloat16,
torch_dtype=torch.float16 if training_args.fp16 else torch.bfloat16,
attn_implementation = "flash_attention_2" if model_args.use_flash_attn else None,
token=model_args.token,
cache_dir=model_args.cache_dir,
Expand Down Expand Up @@ -130,7 +134,7 @@ def get_model(model_args: RerankerModelArguments, only_for_one_logit):

model = LayerWiseMiniCPMForCausalLM.from_pretrained(
model_args.model_name_or_path,
# torch_dtype=torch.float16 if training_args.fp16 else torch.bfloat16,
torch_dtype=torch.float16 if training_args.fp16 else torch.bfloat16,
attn_implementation = "flash_attention_2" if model_args.use_flash_attn else None,
token=model_args.token,
cache_dir=model_args.cache_dir,
Expand Down
3 changes: 1 addition & 2 deletions FlagEmbedding/finetune/reranker/decoder_only/layerwise/runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,3 @@
import os
import logging
from typing import Tuple
from pathlib import Path
Expand Down Expand Up @@ -65,7 +64,7 @@ def load_tokenizer_and_model(self) -> Tuple[PreTrainedTokenizer, AbsRerankerMode
# if 'mistral' in self.model_args.model_name_or_path.lower():
tokenizer.padding_side = 'left'

base_model = get_model(self.model_args, tokenizer('Yes', add_special_tokens=False)['input_ids'][-1])
base_model = get_model(self.model_args, tokenizer('Yes', add_special_tokens=False)['input_ids'][-1], self.training_args)

model = CrossDecoderModel(
base_model,
Expand Down