微调阶段，fp16训练loss值为0，fp32训练loss正常 #1370
Description
模型
jina-embeddings-v2-base-zh
模型微调参考
FlagEmbedding/examples/finetune/embedder/encoder_only/ 路径下的脚本修改
微调命令：
export WANDB_MODE=disabled
train_data="/xxx/xxx/data/finetune_data_score_v2.jsonl"
num_train_epochs=4
per_device_train_batch_size=256
num_gpus=2
if [ -z "$HF_HUB_CACHE" ]; then
export HF_HUB_CACHE="$HOME/.cache/huggingface/hub"
fi
model_args="
--model_name_or_path /SharedNFS/LLM_model/jina-embeddings-v2-base-zh
--cache_dir $HF_HUB_CACHE
--trust_remote_code True
"
data_args="
--train_data $train_data
--cache_path ~/.cache
--train_group_size 15
--query_max_len 32
--passage_max_len 32
--pad_to_multiple_of 8
--query_instruction_for_retrieval 'Represent this sentence for searching relevant passages: '
--query_instruction_format '{}{}'
--knowledge_distillation True
"
training_args="
--output_dir ./knowledge_distillation_agent_minedHN_score_test_encoder_only_base_jina-embeddings-v2-base-zh
--overwrite_output_dir
--learning_rate 1e-5
--fp16 \ # 唯一区别就是这里要不要指定fp16进行训练
--num_train_epochs $num_train_epochs
--per_device_train_batch_size $per_device_train_batch_size
--dataloader_drop_last True
--warmup_ratio 0.1
--gradient_checkpointing
--deepspeed ../../ds_stage0.json
--logging_steps 1
--save_steps 1000
--negatives_cross_device
--temperature 0.02
--sentence_pooling_method mean
--normalize_embeddings True
--kd_loss_type kl_div
"
cmd="torchrun --nproc_per_node $num_gpus
-m FlagEmbedding.finetune.embedder.encoder_only.base
$model_args
$data_args
$training_args
"
echo $cmd
eval $cmd
log输出（fp16）：
/data/anaconda3/envs/c-mteb/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py:632: UserWarning: torch.utils.checkpoint: the use_reentrant parameter should be passed explicitly. In version 2.5 we will raise an exception if use_reentrant is not passed. use_reentrant=False is recommended, but if you need to preserve the current default behavior, you can pass use_reentrant=True. Refer to docs for more details on the differences between the two variants.
return fn(*args, **kwargs)
0%| | 1/452 [00:04<32:47, 4.36s/it]
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 0, 'epoch': 0.01}
0%| | 1/452 [00:04<32:47, 4.36s/it]
0%| | 2/452 [00:07<26:18, 3.51s/it]
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 0, 'epoch': 0.02}
0%| | 2/452 [00:07<26:18, 3.51s/it]
1%| | 3/452 [00:10<24:00, 3.21s/it]
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 0, 'epoch': 0.03}
1%| | 3/452 [00:10<24:00, 3.21s/it]
1%| | 4/452 [00:13<25:15, 3.38s/it]
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 0, 'epoch': 0.04}
1%| | 4/452 [00:13<25:15, 3.38s/it]
1%| | 5/452 [00:16<24:40, 3.31s/it]
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 0, 'epoch': 0.04}
1%| | 5/452 [00:16<24:40, 3.31s/it]
1%|▏ | 6/452 [00:19<23:28, 3.16s/it]
log输出（fp32）：
/data/anaconda3/envs/c-mteb/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py:632: UserWarning: torch.utils.checkpoint: the use_reentrant parameter should be passed explicitly. In version 2.5 we will raise an exception if use_reentrant is not passed. use_reentrant=False is recommended, but if you need to preserve the current default behavior, you can pass use_reentrant=True. Refer to docs for more details on the differences between the two variants.
return fn(*args, **kwargs)
0%| | 1/452 [00:07<59:15, 7.88s/it]
{'loss': 5.9738, 'learning_rate': 0.0, 'epoch': 0.01}
0%| | 1/452 [00:07<59:15, 7.88s/it]
0%| | 2/452 [00:14<52:21, 6.98s/it]
{'loss': 5.2496, 'learning_rate': 1.8104259678004022e-06, 'epoch': 0.02}
0%| | 2/452 [00:14<52:21, 6.98s/it]
1%| | 3/452 [00:20<50:08, 6.70s/it]
{'loss': 5.9845, 'learning_rate': 2.8694572692954448e-06, 'epoch': 0.03}
1%| | 3/452 [00:20<50:08, 6.70s/it]
1%| | 4/452 [00:28<54:47, 7.34s/it]
{'loss': 5.684, 'learning_rate': 3.6208519356008044e-06, 'epoch': 0.04}
1%| | 4/452 [00:28<54:47, 7.34s/it]
1%| | 5/452 [00:36<54:48, 7.36s/it]
{'loss': 5.5646, 'learning_rate': 4.203678918349396e-06, 'epoch': 0.04}
1%| | 5/452 [00:36<54:48, 7.36s/it]
1%|▏ | 6/452 [00:42<52:31, 7.07s/it]
另外想问一下，loss大概在多少模型差不多收敛