[Feature] If i want to start Fine-tuning, must the flashattention be installed?

[snowattitude]

Motivation

GPUS=2 PER_DEVICE_BATCH_SIZE=1 sh ./****2_5_2b_lora.sh

mistake

FlashAttention is not installed.
FlashAttention is not installed.
flash-attention package not found, consider installing for better performance: No module named 'flash_attn'.
Current flash-attenton does not support window_size. Either upgrade or use attn_implementation='eager'.
flash-attention package not found, consider installing for better performance: No module named 'flash_attn'.
Current flash-attenton does not support window_size. Either upgrade or use attn_implementation='eager'.

[Weiyun1025]

It is not required; however, installing it will significantly improve the training speed.

[snowattitude]

how can i turn the flash-attention's switch off? because i can't find how to turn off it

[snowattitude]

It is not required; however, installing it will significantly improve the training speed.

does this code must use Ampere GPU?

[Weiyun1025]

you can try to set _attn_implementation to eager in the config.

[snowattitude]

you can try to set _attn_implementation to eager in the config.

my file is the same as you

[snowattitude]

{
"_commit_hash": null,
"architectures": [
"InternVLChatModel"
],
"auto_map": {
"AutoConfig": "configuration_internvl_chat.InternVLChatConfig",
"AutoModel": "modeling_internvl_chat.InternVLChatModel",
"AutoModelForCausalLM": "modeling_internvl_chat.InternVLChatModel"
},
"downsample_ratio": 0.5,
"dynamic_image_size": true,
"force_image_size": 448,
"llm_config": {
"_name_or_path": "internlm/internlm2_5-1_8b-chat",
"add_cross_attention": false,
"architectures": [
"InternLM2ForCausalLM"
],
"attn_implementation": "flash_attention_2",
"auto_map": {
"AutoConfig": "configuration_internlm2.InternLM2Config",
"AutoModel": "modeling_internlm2.InternLM2ForCausalLM",
"AutoModelForCausalLM": "modeling_internlm2.InternLM2ForCausalLM",
"AutoModelForSequenceClassification": "modeling_internlm2.InternLM2ForSequenceClassification"
},
"bad_words_ids": null,
"begin_suppress_tokens": null,
"bias": false,
"bos_token_id": 1,
"chunk_size_feed_forward": 0,
"cross_attention_hidden_size": null,
"decoder_start_token_id": null,
"diversity_penalty": 0.0,
"do_sample": false,
"early_stopping": false,
"encoder_no_repeat_ngram_size": 0,
"eos_token_id": 2,
"exponential_decay_length_penalty": null,
"finetuning_task": null,
"forced_bos_token_id": null,
"forced_eos_token_id": null,
"hidden_act": "silu",
"hidden_size": 2048,
"id2label": {
"0": "LABEL_0",
"1": "LABEL_1"
},
"initializer_range": 0.02,
"intermediate_size": 8192,
"is_decoder": false,
"is_encoder_decoder": false,
"label2id": {
"LABEL_0": 0,
"LABEL_1": 1
},
"length_penalty": 1.0,
"max_length": 20,
"max_position_embeddings": 32768,
"min_length": 0,
"model_type": "internlm2",
"no_repeat_ngram_size": 0,
"num_attention_heads": 16,
"num_beam_groups": 1,
"num_beams": 1,
"num_hidden_layers": 24,
"num_key_value_heads": 8,
"num_return_sequences": 1,
"output_attentions": false,
"output_hidden_states": false,
"output_scores": false,
"pad_token_id": 2,
"prefix": null,
"pretraining_tp": 1,
"problem_type": null,
"pruned_heads": {},
"remove_invalid_values": false,
"repetition_penalty": 1.0,
"return_dict": true,
"return_dict_in_generate": false,
"rms_norm_eps": 1e-05,
"rope_scaling": {
"factor": 2.0,
"type": "dynamic"
},
"rope_theta": 1000000,
"sep_token_id": null,
"suppress_tokens": null,
"task_specific_params": null,
"temperature": 1.0,
"tf_legacy_loss": false,
"tie_encoder_decoder": false,
"tie_word_embeddings": false,
"tokenizer_class": null,
"top_k": 50,
"top_p": 1.0,
"torch_dtype": "bfloat16",
"torchscript": false,
"transformers_version": "4.37.2",
"typical_p": 1.0,
"use_bfloat16": true,
"use_cache": true,
"vocab_size": 92553
},
"max_dynamic_patch": 12,
"min_dynamic_patch": 1,
"model_type": "internvl_chat",
"ps_version": "v2",
"select_layer": -1,
"template": "internvl2_5",
"torch_dtype": "bfloat16",
"use_backbone_lora": 0,
"use_llm_lora": 0,
"use_thumbnail": true,
"vision_config": {
"architectures": [
"InternVisionModel"
],
"attention_dropout": 0.0,
"drop_path_rate": 0.0,
"dropout": 0.0,
"hidden_act": "gelu",
"hidden_size": 1024,
"image_size": 448,
"initializer_factor": 1.0,
"initializer_range": 0.02,
"intermediate_size": 4096,
"layer_norm_eps": 1e-06,
"model_type": "intern_vit_6b",
"norm_type": "layer_norm",
"num_attention_heads": 16,
"num_channels": 3,
"num_hidden_layers": 24,
"output_attentions": false,
"output_hidden_states": false,
"patch_size": 14,
"qk_normalization": false,
"qkv_bias": true,
"return_dict": true,
"torch_dtype": "bfloat16",
"transformers_version": "4.37.2",
"use_bfloat16": false,
"use_flash_attn": false
}
}

[popoyaya]

+1，how can I finetune the model without the flash-attn?

[popoyaya]

how can i turn the flash-attention's switch off? because i can't find how to turn off it

+1，how can I finetune the model without the flash-attn?

[genkerizer]

You need to install flash_attn after removing code at InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py

[popoyaya]

You need to install flash_attn after removing code at InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py

Thanks！However, I encounter an error when installing flash-attn, and I am unable to resolve it. Is there any fine-tuning method that does not require flash-attn?

[snowattitude]

You need to install flash_attn after removing code at InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py

Thanks！However, I encounter an error when installing flash-attn, and I am unable to resolve it. Is there any fine-tuning method that does not require flash-attn?

i can not see the question image, but u should install by the file named "~.whl"

[genkerizer]

You need to install flash_attn after removing code at InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py

Thanks！However, I encounter an error when installing flash-attn, and I am unable to resolve it. Is there any fine-tuning method that does not require flash-attn?

i can not see the question image, but u should install by the file named "~.whl"

[genkerizer]

You need to install flash_attn after removing code at InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py

Thanks！However, I encounter an error when installing flash-attn, and I am unable to resolve it. Is there any fine-tuning method that does not require flash-attn?

to install Flash_attn, please pull Nvidia devel image, refer in: https://catalog.ngc.nvidia.com/orgs/nvidia/containers/cuda/tags

Example: nvcr.io/nvidia/cuda:12.6.3-cudnn-devel-ubuntu20.04

[popoyaya]

You need to install flash_attn after removing code at InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py

Thanks！However, I encounter an error when installing flash-attn, and I am unable to resolve it. Is there any fine-tuning method that does not require flash-attn?

to install Flash_attn, please pull Nvidia devel image, refer in: https://catalog.ngc.nvidia.com/orgs/nvidia/containers/cuda/tags

Example: nvcr.io/nvidia/cuda:12.6.3-cudnn-devel-ubuntu20.04

Thank you so much! Help a lot!

[popoyaya]

You need to install flash_attn after removing code at InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py

Thanks！However, I encounter an error when installing flash-attn, and I am unable to resolve it. Is there any fine-tuning method that does not require flash-attn?

to install Flash_attn, please pull Nvidia devel image, refer in: https://catalog.ngc.nvidia.com/orgs/nvidia/containers/cuda/tags

Example: nvcr.io/nvidia/cuda:12.6.3-cudnn-devel-ubuntu20.04

I have another question regarding the use of LoRA. Do I only need to set --freeze_backbone False? However, after training, I found that in the config.json, "use_backbone_lora": 0. What else should I do if I want to finetune the visual encoder as well?

Thank you in advance for your help!