Using Swift to perform inference and fine-tune InternVL-Chat-V1.5

[hjh0119]

Thanks for your awesome work.

swift now supports inference, training of InternVL-Chat-V1.5 model

For more information, please refer to our document

• English
• 中文

For more questions, please raise an issue in the swift repository.

[czczup]

That's great, thank you very much!

[NLP-Learning]

Thanks for your awesome work.

swift now supports inference, training of InternVL-Chat-V1.5 model

For more information, please refer to our document

• English
• 中文

请问按您教程里的来做，最后会报错没有flash_attn，使用的是V100，有解决方案吗？

[INFO:swift] InternVLChatModel: 25514.1861M Params (25514.1861M Trainable [100.0000%]), 402.6563M Buffers. [INFO:swift] system: You are an AI assistant whose name is InternLM (书生·浦语). [INFO:swift] Input exit or quit to exit the conversation. [INFO:swift] Input multi-line to switch to multi-line input mode. [INFO:swift] Input reset-system to reset the system and clear the history. [INFO:swift] Input clear to clear the history. [INFO:swift] Please enter the conversation content first, followed by the path to the multimedia file. <<< 描述一下这张图片的内容 Input a media path or URL <<< https://img2.baidu.com/it/u=2085854734,3872819026&fm=253&fmt=auto&app=138&f=JPEG?w=762&h=500 Exception in thread Thread-2: Traceback (most recent call last): File "/home/sunyuan/.cache/huggingface/modules/transformers_modules/InternVL-Chat-V1-5/modeling_internlm2.py", line 57, in _import_flash_attn from flash_attn import flash_attn_func as _flash_attn_func ModuleNotFoundError: No module named 'flash_attn'

[hjh0119]

请问按您教程里的来做，最后会报错没有flash_attn，使用的是V100，有解决方案吗？

解决方案：将model.config.attn_implementation设为eager，拉下最新代码，用use_flash_attn false

[sunzx8]

请问按您教程里的来做，最后会报错没有flash_attn，使用的是V100，有解决方案吗？

解决方案：将model.config.attn_implementation设为eager，拉下最新代码，用use_flash_attn false

请问全参微调internVL-V1-5用您这个框架需要多大的显存

[sunzx8]

请问按您教程里的来做，最后会报错没有flash_attn，使用的是V100，有解决方案吗？

解决方案：将model.config.attn_implementation设为eager，拉下最新代码，用use_flash_attn false

我看readme里面 写的是只用4*72G就可以全参，请问这是怎么做到的，这个不是26B得参数吗

[hjh0119]

请问全参微调internVL-V1-5用您这个框架需要多大的显存

你好，4*72G是一个参考值，训练显存会根据训练数据的图片大小动态变化，这是我刚才全参训练的一个占用情况

[sunzx8]

请问全参微调internVL-V1-5用您这个框架需要多大的显存

你好，4*72G是一个参考值，训练显存会根据训练数据的图片大小动态变化，这是我刚才全参训练的一个占用情况

很感谢您，请问您的意思是微调的照片分辨率会影响训练的显存吗？请问您这是多大的图片

[sunzx8]

请问全参微调internVL-V1-5用您这个框架需要多大的显存

你好，4*72G是一个参考值，训练显存会根据训练数据的图片大小动态变化，这是我刚才全参训练的一个占用情况

我还是有点不理解，这里的全参数是不是20B的LLM没有参与训练，只有VIT和连接器参与微调了？

[hjh0119]

很感谢您，请问您的意思是微调的照片分辨率会影响训练的显存吗？请问您这是多大的图片

照片大小会影响ViT输出大小进而影响模型的输入序列长度，数据集用的https://www.modelscope.cn/datasets/modelscope/coco_2014_caption/summary

这里的全参数是不是20B的LLM没有参与训练，只有VIT和连接器参与微调了？

全参包括LLM部分，这个是log中训练参数量的信息：[INFO:swift] InternVLChatModel: 25514.1861M Params (25514.1861M Trainable [100.0000%]), 402.6563M Buffers.

[MVP-D77]

@hjh0119 想问下 按照您的教程里，我想对int8版本进行微调，替换掉了--model_type internvl-chat-v1_5-int8 模型名称，加载了从huggingface下载的模型权重，但报了如下的错误，似乎和int8有些关系，对与1.5的版本则不会报错，可以正常微调，想问下这是需要加什么特殊的参数影响么？

/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/bitsandbytes/autograd/_functions.py:316: UserWarning: MatMul8bitLt: inputs will be cast from torch.bfloat16 to float16 during quantization
  warnings.warn(f"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization")
/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/bitsandbytes/autograd/_functions.py:316: UserWarning: MatMul8bitLt: inputs will be cast from torch.float32 to float16 during quantization
  warnings.warn(f"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization")
Traceback (most recent call last):
  File "/data2/renyw/PythonWorkspace/FM-LLM/swift/swift/cli/sft.py", line 5, in <module>
    sft_main()
  File "/data2/renyw/PythonWorkspace/FM-LLM/swift/swift/utils/run_utils.py", line 27, in x_main
    result = llm_x(args, **kwargs)
  File "/data2/renyw/PythonWorkspace/FM-LLM/swift/swift/llm/sft.py", line 256, in llm_sft
    trainer.train(training_args.resume_from_checkpoint)
  File "/data2/renyw/PythonWorkspace/FM-LLM/swift/swift/trainers/trainers.py", line 50, in train
    res = super().train(*args, **kwargs)
  File "/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/transformers/trainer.py", line 1859, in train
    return inner_training_loop(
  File "/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/transformers/trainer.py", line 2203, in _inner_training_loop
    tr_loss_step = self.training_step(model, inputs)
  File "/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/transformers/trainer.py", line 3147, in training_step
    self.accelerator.backward(loss)
  File "/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/accelerate/accelerator.py", line 2125, in backward
    loss.backward(**kwargs)
  File "/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/torch/_tensor.py", line 492, in backward
    torch.autograd.backward(
  File "/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/torch/autograd/__init__.py", line 251, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
  File "/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/torch/autograd/function.py", line 288, in apply
    return user_fn(self, *args)
  File "/data2/renyw/InstallationPackage/anaconda3/envs/swift/lib/python3.10/site-packages/bitsandbytes/autograd/_functions.py", line 479, in backward
    .mul_(state.SCB.unsqueeze(1).mul(1.0 / 127.0))
RuntimeError: The size of tensor a (92576) must match the size of tensor b (92553) at non-singleton dimension 0
Train:   0%|                                                                                                                        | 0/2531 [00:02<?, ?it/s]

[hjh0119]

@MVP-D77 不需要替换model_type， 使用本地模型文件在--model_id_or_path指定路径， 再指定--model_type internvl-chat-v1_5-int8.

[MVP-D77]

@MVP-D77 不需要替换model_type， 使用本地模型文件在--model_id_or_path指定路径， 再指定--model_type internvl-chat-v1_5-int8.

@hjh0119 您好，这是我的微调命令，请您看一下有没有问题，它还是报错，和上面报错信息一样，同样的数据集，internvl-1.5就不会报错，另外想问下，flash_attn如果想开启的话，是加一个设置为true么，我看打印的log文件里面显示use_flash_attn 为null

CUDA_VISIBLE_DEVICES=0,1 swift sft  --model_type  internvl-chat-v1_5-int8  --dataset coco-mini-en-2 --model_id_or_path xxxxxxxxx/InternVL/pretrained/InternVL-Chat-V1-5-Int8

[hjh0119]

@hjh0119 您好，这是我的微调命令，请您看一下有没有问题，它还是报错，和上面报错信息一样，同样的数据集，internvl-1.5就不会报错，另外想问下，flash_attn如果想开启的话，是加一个设置为true么，我看打印的log文件里面显示use_flash_attn 为null

可以复现，修复中 开启flash-attention： --use_flash_attn true 关于swift的问题欢迎到swift提issue

[MVP-D77]

@hjh0119 您好，这是我的微调命令，请您看一下有没有问题，它还是报错，和上面报错信息一样，同样的数据集，internvl-1.5就不会报错，另外想问下，flash_attn如果想开启的话，是加一个设置为true么，我看打印的log文件里面显示use_flash_attn 为null

可以复现，修复中 开启flash-attention： --use_flash_attn true 关于swift的问题欢迎到swift提issue

@hjh0119 谢谢您的回复，也感谢您的工作，我还有两个新的有关internvl微调的问题，已在swift仓库提出issue https://github.com/modelscope/swift/issues/925 非常期待您的回复

[sunzx8]

很感谢您，请问您的意思是微调的照片分辨率会影响训练的显存吗？请问您这是多大的图片

照片大小会影响ViT输出大小进而影响模型的输入序列长度，数据集用的https://www.modelscope.cn/datasets/modelscope/coco_2014_caption/summary

这里的全参数是不是20B的LLM没有参与训练，只有VIT和连接器参与微调了？

全参包括LLM部分，这个是log中训练参数量的信息：[INFO:swift] InternVLChatModel: 25514.1861M Params (25514.1861M Trainable [100.0000%]), 402.6563M Buffers.

您好，谢谢您的回复，请问如果我想把微调时的max_length扩宽应该怎么做？我现在发现只要我不用默认的2048(比如改成4096)就会报错 " RuntimeError: CUDA error: unspecified launch failure CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with TORCH_USE_CUDA_DSA to enable device-side assertions. "

[1028686314]

你好，请问你们的脚本支持多图嘛？我看输入格式是这样的。如果不支持的话，大概的修改思路能否帮忙指点一下呢

(只支持单轮对话, 每轮对话必须包含一张图片, 支持传入本地路径或URL) {"query": "55555", "response": "66666", "images": ["image_path"]} {"query": "eeeee", "response": "fffff", "images": ["image_path"]} {"query": "EEEEE", "response": "FFFFF", "images": ["image_path"]}

[hjh0119]

@1028686314 多图的路径在images里用逗号隔开

[1028686314]

你好 我想请问下 如果我想输入尽量多的图片进行微调和推理 需要修改哪些设置呀？是修改max_length参数就够了嘛。如果修改这个参数做sft，会不会影响模型效果呀？ 我现在测试下来最多输入8张图左右，现在想利用视频抽帧数据做sft，8张图有点不太能满足需求。

[McCarthyjfq]

请问我用的是官方文件来进行lora微调的，想用swift来merge的时候发现merge后的权重和原权重一摸一样，lora权重并没有被merge进去，是什么原因？是一定也要用swift来微调才行吗？

[McCarthyjfq]

请问我用的是官方文件来进行lora微调的，想用swift来merge的时候发现merge后的权重和原权重一摸一样，lora权重并没有被merge进去，是什么原因？是一定也要用swift来微调才行吗？

有一个提示是sft_args.json not found

[1028686314]

请问我用的是官方文件来进行lora微调的，想用swift来merge的时候发现merge后的权重和原权重一摸一样，lora权重并没有被merge进去，是什么原因？是一定也要用swift来微调才行吗？

有一个提示是sft_args.json not found

这个文件是swift微调之后生成的文件 你没有的话看起来是不行

[McCarthyjfq]

请问我用的是官方文件来进行lora微调的，想用swift来merge的时候发现merge后的权重和原权重一摸一样，lora权重并没有被merge进去，是什么原因？是一定也要用swift来微调才行吗？

有一个提示是sft_args.json not found

这个文件是swift微调之后生成的文件 你没有的话看起来是不行

谢谢！

[1028686314]

请问我用的是官方文件来进行lora微调的，想用swift来merge的时候发现merge后的权重和原权重一摸一样，lora权重并没有被merge进去，是什么原因？是一定也要用swift来微调才行吗？

有一个提示是sft_args.json not found

这个文件是swift微调之后生成的文件 你没有的话看起来是不行

可以麻烦看一下你的sft_args.json文件吗，我试试改一下配置merge试试。因为我看swift是不是不支持图文对话的微调？我如果要改我本来数据集稍微有点麻烦，谢谢！

{ "model_type": "internvl-chat-v1_5", "model_id_or_path": "/mnt/workspace/workgroup/shiyi/weights/InternVL-Chat-V1-5", "model_revision": "master", "sft_type": "lora", "freeze_parameters": 0.0, "additional_trainable_parameters": [], "tuner_backend": "peft", "template_type": "internvl", "output_dir": "/mnt/workspace/workgroup/swift/output/internvl-chat-v1_5/v0-20240605-173708", "add_output_dir_suffix": true, "ddp_backend": null, "ddp_find_unused_parameters": null, "ddp_broadcast_buffers": null, "seed": 42, "resume_from_checkpoint": null, "ignore_data_skip": false, "dtype": "bf16", "packing": false, "dataset": [ "/mnt/workspace/workgroup/swift/data/test.jsonl" ], "val_dataset": [], "dataset_seed": 42, "dataset_test_ratio": 0.01, "use_loss_scale": false, "system": "You are an AI assistant whose name is InternLM (书生·浦语).", "max_length": 2048, "truncation_strategy": "delete", "check_dataset_strategy": "none", "model_name": [ null, null ], "model_author": [ null, null ], "quant_method": null, "quantization_bit": 0, "hqq_axis": 0, "hqq_dynamic_config_path": null, "bnb_4bit_comp_dtype": "bf16", "bnb_4bit_quant_type": "nf4", "bnb_4bit_use_double_quant": true, "bnb_4bit_quant_storage": null, "lora_target_modules": [ "wqkv" ], "lora_rank": 8, "lora_alpha": 32, "lora_dropout_p": 0.05, "lora_bias_trainable": "none", "lora_modules_to_save": [], "lora_dtype": null, "lora_lr_ratio": null, "use_rslora": false, "use_dora": false, "init_lora_weights": true, "boft_block_size": 4, "boft_block_num": 0, "boft_n_butterfly_factor": 1, "boft_target_modules": [ "DEFAULT" ], "boft_dropout": 0.0, "boft_modules_to_save": [], "vera_rank": 256, "vera_target_modules": [ "DEFAULT" ], "vera_projection_prng_key": 0, "vera_dropout": 0.0, "vera_d_initial": 0.1, "vera_modules_to_save": [], "adapter_act": "gelu", "adapter_length": 128, "use_galore": false, "galore_rank": 128, "galore_target_modules": null, "galore_update_proj_gap": 50, "galore_scale": 1.0, "galore_proj_type": "std", "galore_optim_per_parameter": false, "galore_with_embedding": false, "adalora_target_r": 8, "adalora_init_r": 12, "adalora_tinit": 0, "adalora_tfinal": 0, "adalora_deltaT": 1, "adalora_beta1": 0.85, "adalora_beta2": 0.85, "adalora_orth_reg_weight": 0.5, "ia3_target_modules": [ "DEFAULT" ], "ia3_feedforward_modules": [], "ia3_modules_to_save": [], "llamapro_num_new_blocks": 4, "llamapro_num_groups": null, "neftune_noise_alpha": null, "neftune_backend": "transformers", "lisa_activated_layers": 0, "lisa_step_interval": 20, "gradient_checkpointing": true, "deepspeed": null, "batch_size": 1, "eval_batch_size": 1, "num_train_epochs": 1, "max_steps": -1, "optim": "adamw_torch", "adam_beta1": 0.9, "adam_beta2": 0.999, "learning_rate": 0.0001, "weight_decay": 0.1, "gradient_accumulation_steps": 16, "max_grad_norm": 0.5, "predict_with_generate": false, "lr_scheduler_type": "linear", "warmup_ratio": 0.05, "eval_steps": 50, "save_steps": 50, "save_only_model": false, "save_total_limit": 2, "logging_steps": 5, "dataloader_num_workers": 0, "dataloader_pin_memory": false, "dataloader_drop_last": false, "push_to_hub": false, "hub_model_id": null, "hub_token": null, "hub_private_repo": false, "push_hub_strategy": "push_best", "test_oom_error": false, "disable_tqdm": false, "lazy_tokenize": true, "preprocess_num_proc": 1, "use_flash_attn": null, "ignore_args_error": false, "check_model_is_latest": true, "logging_dir": "/mnt/workspace/workgroup/swift/output/internvl-chat-v1_5/v0-20240605-173708/runs", "report_to": [ "tensorboard" ], "acc_strategy": "token", "save_on_each_node": true, "evaluation_strategy": "steps", "save_strategy": "steps", "save_safetensors": true, "gpu_memory_fraction": null, "include_num_input_tokens_seen": false, "local_repo_path": null, "custom_register_path": null, "custom_dataset_info": null, "device_map_config_path": null, "max_new_tokens": 2048, "do_sample": true, "temperature": 0.3, "top_k": 20, "top_p": 0.7, "repetition_penalty": 1.0, "num_beams": 1, "fsdp": "", "fsdp_config": null, "sequence_parallel_size": 1, "model_layer_cls_name": null, "metric_warmup_step": 0, "fsdp_num": 1, "per_device_train_batch_size": null, "per_device_eval_batch_size": null, "self_cognition_sample": 0, "train_dataset_mix_ratio": 0.0, "train_dataset_mix_ds": [ "ms-bench" ], "train_dataset_sample": -1, "val_dataset_sample": null, "safe_serialization": null, "only_save_model": null, "neftune_alpha": null, "deepspeed_config_path": null, "model_cache_dir": null, "custom_train_dataset_path": [ "/mnt/workspace/workgroup/swift/data/test.jsonl" ], "custom_val_dataset_path": [], "use_self_cognition": false, "lora_use_embedding": false, "lora_use_all": false, "lora_m2s_use_embedding": false, "lora_m2s_use_ln": false, "torch_dtype": "torch.bfloat16", "fp16": false, "bf16": true, "bnb_4bit_compute_dtype": "torch.bfloat16", "load_in_4bit": false, "load_in_8bit": false, "train_sampler_random": true, "training_args": "Seq2SeqTrainingArguments(output_dir='/mnt/workspace/workgroup/swift/output/internvl-chat-v1_5/v0-20240605-173708', overwrite_output_dir=False, do_train=False, do_eval=True, do_predict=False, eval_strategy=<IntervalStrategy.STEPS: 'steps'>, prediction_loss_only=False, per_device_train_batch_size=1, per_device_eval_batch_size=1, per_gpu_train_batch_size=None, per_gpu_eval_batch_size=None, gradient_accumulation_steps=16, eval_accumulation_steps=None, eval_delay=0, learning_rate=0.0001, weight_decay=0.1, adam_beta1=0.9, adam_beta2=0.999, adam_epsilon=1e-08, max_grad_norm=0.5, num_train_epochs=1, max_steps=-1, lr_scheduler_type=<SchedulerType.LINEAR: 'linear'>, lr_scheduler_kwargs={}, warmup_ratio=0.05, warmup_steps=0, log_level='passive', log_level_replica='warning', log_on_each_node=True, logging_dir='/mnt/workspace/workgroup/swift/output/internvl-chat-v1_5/v0-20240605-173708/runs', logging_strategy=<IntervalStrategy.STEPS: 'steps'>, logging_first_step=True, logging_steps=5, logging_nan_inf_filter=True, save_strategy=<IntervalStrategy.STEPS: 'steps'>, save_steps=50, save_total_limit=2, save_safetensors=True, save_on_each_node=True, save_only_model=False, restore_callback_states_from_checkpoint=False, no_cuda=False, use_cpu=False, use_mps_device=False, seed=42, data_seed=None, jit_mode_eval=False, use_ipex=False, bf16=True, fp16=False, fp16_opt_level='O1', half_precision_backend='auto', bf16_full_eval=False, fp16_full_eval=False, tf32=None, local_rank=0, ddp_backend=None, tpu_num_cores=None, tpu_metrics_debug=False, debug=[], dataloader_drop_last=False, eval_steps=50, dataloader_num_workers=0, dataloader_prefetch_factor=None, past_index=-1, run_name='/mnt/workspace/workgroup/swift/output/internvl-chat-v1_5/v0-20240605-173708', disable_tqdm=False, remove_unused_columns=False, label_names=None, load_best_model_at_end=False, metric_for_best_model='loss', greater_is_better=False, ignore_data_skip=False, fsdp=[], fsdp_min_num_params=0, fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}, fsdp_transformer_layer_cls_to_wrap=None, accelerator_config=AcceleratorConfig(split_batches=False, dispatch_batches=None, even_batches=True, use_seedable_sampler=True, non_blocking=False, gradient_accumulation_kwargs=None), deepspeed=None, label_smoothing_factor=0.0, optim=<OptimizerNames.ADAMW_TORCH: 'adamw_torch'>, optim_args=None, adafactor=False, group_by_length=False, length_column_name='length', report_to=['tensorboard'], ddp_find_unused_parameters=None, ddp_bucket_cap_mb=None, ddp_broadcast_buffers=None, dataloader_pin_memory=False, dataloader_persistent_workers=False, skip_memory_metrics=True, use_legacy_prediction_loop=False, push_to_hub=False, resume_from_checkpoint=None, hub_model_id=None, hub_strategy=<HubStrategy.EVERY_SAVE: 'every_save'>, hub_token=None, hub_private_repo=False, hub_always_push=False, gradient_checkpointing=True, gradient_checkpointing_kwargs=None, include_inputs_for_metrics=False, eval_do_concat_batches=True, fp16_backend='auto', evaluation_strategy='steps', push_to_hub_model_id=None, push_to_hub_organization=None, push_to_hub_token=None, mp_parameters='', auto_find_batch_size=False, full_determinism=False, torchdynamo=None, ray_scope='last', ddp_timeout=1800, torch_compile=False, torch_compile_backend=None, torch_compile_mode=None, dispatch_batches=None, split_batches=None, include_tokens_per_second=False, include_num_input_tokens_seen=False, neftune_noise_alpha=None, optim_target_modules=None, batch_eval_metrics=False, sortish_sampler=True, predict_with_generate=False, generation_max_length=None, generation_num_beams=None, generation_config=GenerationConfig {\n "do_sample": true,\n "eos_token_id": 2,\n "max_new_tokens": 2048,\n "pad_token_id": 2,\n "temperature": 0.3,\n "top_k": 20,\n "top_p": 0.7\n}\n, train_sampler_random=True, push_hub_strategy='push_best', acc_strategy='token', additional_saved_files=[], metric_warmup_step=0, train_dataset_sample=4)" }

[6Roy]

请问下大家，Mini-InternVL-1.5在swift下微调支持这种图文交织的格式么，类似这种支持多回合对话，每回合可包含多图或无图，支持传入本地路径或URL [ {"conversations": [ {"from": "user", "value": "Picture 1:img_path\n11111"}, {"from": "assistant", "value": "22222"} ]}, {"conversations": [ {"from": "user", "value": "Picture 1:img_path\nPicture 2:img_path2\nPicture 3:img_path3\naaaaa"}, {"from": "assistant", "value": "bbbbb"}, {"from": "user", "value": "Picture 1:img_path\nccccc"}, {"from": "assistant", "value": "ddddd"} ]}, {"conversations": [ {"from": "user", "value": "AAAAA"}, {"from": "assistant", "value": "BBBBB"}, {"from": "user", "value": "CCCCC"}, {"from": "assistant", "value": "DDDDD"} ]} ]

[6Roy]

你好，请问你们的脚本支持多图嘛？我看输入格式是这样的。如果不支持的话，大概的修改思路能否帮忙指点一下呢

(只支持单轮对话, 每轮对话必须包含一张图片, 支持传入本地路径或URL) {"query": "55555", "response": "66666", "images": ["image_path"]} {"query": "eeeee", "response": "fffff", "images": ["image_path"]} {"query": "EEEEE", "response": "FFFFF", "images": ["image_path"]}

@1028686314 多图的路径在images里用逗号隔开

请问下大佬，Mini-InternVL-1.5在swift下微调支持这种图文交织的格式么，类似这种支持多回合对话，每回合可包含多图或无图，支持传入本地路径或URL [ {"conversations": [ {"from": "user", "value": "Picture 1:img_path\n11111"}, {"from": "assistant", "value": "22222"} ]}, {"conversations": [ {"from": "user", "value": "Picture 1:img_path\nPicture 2:img_path2\nPicture 3:img_path3\naaaaa"}, {"from": "assistant", "value": "bbbbb"}, {"from": "user", "value": "Picture 1:img_path\nccccc"}, {"from": "assistant", "value": "ddddd"} ]}, {"conversations": [ {"from": "user", "value": "AAAAA"}, {"from": "assistant", "value": "BBBBB"}, {"from": "user", "value": "CCCCC"}, {"from": "assistant", "value": "DDDDD"} ]} ]

[fourierer]

@1028686314 多图的路径在images里用逗号隔开

@1028686314 多图的路径在images里用逗号隔开

您好，使用swift按照逗号隔开来微调多图任务，推理的时候我可以按照intern-vl的多图推理方式进行推理么 ，就是图中的torch.cat方式

[G-z-w]

感谢关注，具体请参考swift实现。

[radna0]

Can you provide example code to work with internVL2 Model on TPUs XLA @hjh0119