Merge pull request #648 from OptimalScale/lianqing/multi_modal_training

[FIX] Fix multi-modal training

examples/finetune_multi_modal.py

@@ -62,7 +62,8 @@ def main():
     model = AutoModel.get_model(model_args, tune_strategy='none',
                                 ds_config=pipeline_args.deepspeed,
                                 custom_model=True,
-                                with_deepspeed=False)
+                                with_deepspeed=False,
+                                pipeline_args=pipeline_args)
     # FIXME check if need to move this part to hf_encoder_decoder.py
     for param in model.backend_model.parameters():
         param.requires_grad = False

scripts/multimodal/run_finetune_multi_modal_stage1.sh

@@ -9,7 +9,8 @@ model_name_or_path=Salesforce/blip2-flan-t5-xxl
 # https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K
 dataset_path=./data/llava_cc3m_pretrain_595k/chat.json
 image_folder=./data/llava_cc3m_pretrain_595k/images
-output_dir=output_models/finetune
+output_dir=output_models/finetune_llava-336px-vicuna-7b-v1.3_stage1
+
 deepspeed_args="--master_port=12000"
 
 while [[ $# -ge 1 ]]; do
@@ -62,9 +63,8 @@ deepspeed ${deepspeed_args} \
     --llm_model_name_or_path lmsys/vicuna-7b-v1.5 \
     --image_aspect_ratio None \
     --fp16 True \
-    --learning_rate 2e-5 \
-    --gradient_accumulation_steps 1 \
-    --per_device_train_batch_size 2 \
+    --gradient_accumulation_steps 4 \
+    --per_device_train_batch_size 8 \
     --learning_rate 2e-3 \
     --weight_decay 0. \
     --warmup_ratio 0.03 \
@@ -77,5 +77,6 @@ deepspeed ${deepspeed_args} \
     --save_steps 5000 \
     --dataloader_num_workers 1 \
     --num_train_epochs 1 \
+    --save_language_projection True \
     | tee ${log_dir}/train.log \
     2> ${log_dir}/train.err

scripts/multimodal/run_finetune_multi_modal_stage2.sh

@@ -8,9 +8,9 @@
 model_name_or_path=Salesforce/blip2-flan-t5-xxl
 # Please download the coco dataset and the conversation file from
 # https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_instruct_80k.json
-dataset_path=./data/llava_instruct_80k.json
+dataset_path=./data/llava_instruct_80k_truncated.json
 image_folder=./data/coco2017/train2017
-output_dir=output_models/finetune
+output_dir=output_models/finetune_llava-336px-vicuna-7b-v1.3_stage2
 deepspeed_args="--master_port=12000"
 
 while [[ $# -ge 1 ]]; do
@@ -50,31 +50,39 @@ fi
 if [ ! -f data/llava_instruct_80k.json ]; then
   cd data && ./download.sh llava_instruction_finetune_80k && cd -
 fi
+
+if [ ! -f data/llava_instruct_80k_truncated.json ]; then
+  python utils/preprocess_multimodal_data.py \
+      --data_path data/llava_instruct_80k.json \
+      --save_path data/llava_instruct_80k_truncated.json
+fi
+
 # Finetune
 exp_id=finetune
 project_dir=$(cd "$(dirname $0)"/..; pwd)
 log_dir=${project_dir}/log/${exp_id}
 mkdir -p ${output_dir} ${log_dir}
-
+# train batch size is set as 4
+# default in llava is 16
 deepspeed ${deepspeed_args} \
   examples/finetune_multi_modal.py \
-    --deepspeed configs/ds_config_multimodal.json \
+    --deepspeed configs/ds_config_zero2.json \
     --arch_type vision_encoder_decoder \
     --llava_loading True \
     --model_name_or_path ${model_name_or_path} \
-    --image_encoder_name_or_path openai/clip-vit-large-patch14 \
+    --image_encoder_name_or_path openai/clip-vit-large-patch14-336 \
     --pretrained_language_projection_path output_models/llava-336px-pretrain-vicuna-7b-v1.3_language_projection.pth \
     --dataset_path ${dataset_path} \
     --output_dir ${output_dir} --overwrite_output_dir \
     --image_folder ${image_folder} \
     --custom_vision_model True \
-    --llm_model_name_or_path lmsys/vicuna-7b-v1.5 \
+    --llm_model_name_or_path lmsys/vicuna-7b-v1.3 \
     --image_aspect_ratio None \
     --fp16 True \
     --learning_rate 2e-5 \
-    --gradient_accumulation_steps 1 \
-    --per_device_train_batch_size 2 \
-    --learning_rate 2e-3 \
+    --gradient_accumulation_steps 8 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 4 \
     --weight_decay 0. \
     --warmup_ratio 0.03 \
     --lr_scheduler_type "cosine" \

scripts/multimodal/run_vis_chatbot_llava.sh

@@ -1,3 +1,4 @@
+# only work for gpu mem > 25G; fail to do 4 bit and 8 bit inference.
 model_name_or_path=Salesforce/blip2-flan-t5-xxl
 llava_pretrain_model_path="output_models/llava-v1-0719-336px-lora-merge-vicuna-13b-v1.3/"
 deepspeed_args="--master_port=12000"
@@ -23,13 +24,13 @@ deepspeed ${deepspeed_args} \
     --custom_model True \
     --chatbot_type llava \
     --prompt_structure '{input_text} ASSISTANT:' \
-    --low_resource True \
     --llava_loading True \
     --model_name_or_path ${model_name_or_path} \
-    --image_encoder_name_or_path openai/clip-vit-large-patch14 \
+    --image_encoder_name_or_path openai/clip-vit-large-patch14-336 \
     --custom_vision_model True \
     --llm_model_name_or_path lmsys/vicuna-13b-v1.5 \
     --llava_pretrain_model_path ${llava_pretrain_model_path}"*.bin" \
     --with_deepspeed False \
+    --save_pretrain_model_path "output_models/lmflow_llava-v1-0719-336px-lora-merge-vicuna-13b-v1.3" \
     ${@:1}
 

src/lmflow/args.py

@@ -538,14 +538,20 @@ class FinetunerArguments(TrainingArguments):
     remove_unused_columns: Optional[bool] = field(
         default=False,
         metadata={
-            "help": "wheather to remove the unused columns in collate fn"}
+            "help": "whether to remove the unused columns in collate fn"}
     )
     finetune_part: Optional[str] = field(
         default="language_projection",
         metadata={
             "help": "the module to finetune."
         }
     )
+    save_language_projection: Optional[str] = field(
+        default=False,
+        metadata={
+            "help": "whether to save language projection layer in multi-modal models."
+        }
+    )
 
 
 @dataclass

src/lmflow/models/hf_encoder_decoder_model.py

@@ -33,7 +33,7 @@
 )
 
 import torch
-from transformers.deepspeed import HfDeepSpeedConfig
+from transformers.deepspeed import HfDeepSpeedConfig, HfTrainerDeepSpeedConfig
 
 from transformers.testing_utils import CaptureLogger
 
@@ -89,6 +89,7 @@ def __init__(
         use_accelerator=False,
         custom_model=False,
         with_deepspeed=True,
+        pipeline_args=None,
         *args,
         **kwargs
     ):
@@ -119,7 +120,10 @@ def __init__(
                 raise NotImplementedError(
                     f"Currently encoder2decoder model is not supported with accelerator"
                 )
-            dschf = HfDeepSpeedConfig(ds_config)
+            # dschf = HfDeepSpeedConfig(ds_config)
+            dschf = HfTrainerDeepSpeedConfig(ds_config)
+            if pipeline_args is not None:
+                dschf.trainer_config_process(pipeline_args)
             peft_model_id = model_args.lora_model_path
             # NOTE: Currently offload is not supported by llama
             if "llama" in model_args.model_name_or_path and model_args.use_ram_optimized_load:
@@ -198,10 +202,11 @@ def __init__(
                         kwargs = dict(
                             torch_dtype=torch.float16,
                             load_in_8bit=True,
-                            device_map="auto"
+                            device_map="auto",
                         )
                     else:
-                        kwargs = {}
+                        # kwargs = dict(torch_dtype=torch.float16)
+                        kwargs = dict(device_map="auto")
                     if (model_args.image_encoder_name_or_path is None and
                         model_args.qformer_name_or_path is None and
                         model_args.llm_model_name_or_path is None):

src/lmflow/models/vision2seq_model.py

@@ -22,6 +22,7 @@
     PreTrainedModel,
 )
 from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.deepspeed import is_deepspeed_zero3_enabled
 
 from lmflow.models.base_model import BaseModel
 from lmflow.models.vision_encoder import build_vision_tower
@@ -77,15 +78,18 @@ def __init__(self,
                     torch.zeros(1, config.num_query_tokens,
                                 config.qformer_config.hidden_size))
                 self.qformer = Blip2QFormerModel(config.qformer_config)
-        if low_resource:
-            kwargs = dict(
-                torch_dtype=torch.float16,
-                load_in_8bit=True,
-                device_map="auto"
-            )
-        else:
-            kwargs = {}
+        kwargs = dict()
         if language_model_name_or_path is not None:
+            if low_resource:
+                kwargs = dict(
+                    torch_dtype=torch.float16,
+                    load_in_8bit=True,
+                    device_map="auto",
+                    low_cpu_mem_usage=True)
+            else:
+                if not is_deepspeed_zero3_enabled:
+                    kwargs = dict(device_map="auto",
+                            torch_dtype=torch.float16)
             language_model = AutoModelForCausalLM.from_pretrained(
                 language_model_name_or_path, **kwargs)
             config.text_config = language_model.config
@@ -97,7 +101,7 @@ def __init__(self,
                 language_model = AutoModelForSeq2SeqLM.from_config(
                     config.text_config, **kwargs)
         # Update _tied_weights_keys using the base model used.
-        if language_model._tied_weights_keys is not None:
+        if getattr(language_model, "_tied_weights_keys", None) is not None:
             self._tied_weights_keys = [f"language_model.{k}" for k in language_model._tied_weights_keys]
 
         self.language_model = language_model
@@ -237,7 +241,6 @@ def forward(
             batch_size = pixel_values.shape[0]
         else:
             batch_size = 1
-
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -298,6 +301,9 @@ def forward(
         # if inputs_embeds is not None:
         #     print("input_embeds", inputs_embeds.shape)
         # attention_mask.shape, inputs_embeds.shape)
+        # TODO remove this code by fixing the ddp training issue
+        inputs_embeds = inputs_embeds.to(
+            self.language_model.lm_head.weight.dtype)
         outputs = self.language_model(
             input_ids=input_ids,
             attention_mask=attention_mask,
@@ -356,7 +362,7 @@ def processor_image_token_in_minigpt4(self,
 
         # concatenate query embeddings with prompt embeddings
         inputs_embeds = self.get_input_embeddings()(input_ids)
-        inputs_embeds = inputs_embeds.to(language_model_inputs.device)
+        inputs_embeds = inputs_embeds.to(device=language_model_inputs.device)
         # concatenate the text embeddings with image embeddings
         inputs_embeds_with_images = []
         attention_mask_with_images = []
@@ -426,7 +432,6 @@ def generate(
             batch_size = pixel_values.shape[0]
         else:
             batch_size = 1
-
         if not self.custom_vision_model:
             # do the processing as blip2 and mini gpt-4;
             image_embeds = self.vision_model(
@@ -473,6 +478,8 @@ def generate(
                     self.language_model.model)
         # convert the dtype.
         # FIXME check when need to do this
+        inputs_embeds = inputs_embeds.to(
+            device=self.language_model.lm_head.weight.device)
         inputs_embeds = inputs_embeds.to(
             self.language_model.lm_head.weight.dtype)
         outputs = self.language_model.generate(

src/lmflow/models/vision_encoder/clip_encoder.py

@@ -128,7 +128,6 @@ def prepare_inputs_labels_for_multimodal(
             image_features = [x.flatten(0, 1) for x in image_features]
         else:
             image_features = self.encode_images(images, language_projection)
-
         new_input_embeds = []
         new_labels = [] if labels is not None else None
         cur_image_idx = 0
@@ -164,6 +163,7 @@ def prepare_inputs_labels_for_multimodal(
                         cur_new_labels.append(cur_labels[image_token_start:image_token_start+1])
                         cur_labels = cur_labels[image_token_start+2:]
                 else:
+                    cur_input_ids = cur_input_ids.to(device=language_model.device)
                     cur_new_input_embeds.append(language_model.embed_tokens(cur_input_ids[:image_token_start]))
                     cur_new_input_embeds.append(cur_image_features)
                     if labels is not None:

src/lmflow/pipeline/finetuner.py

@@ -296,7 +296,6 @@ def compute_metrics(eval_preds):
             callbacks=trainer_callbacks
         )
         # Training
-        # import pdb; pdb.set_trace()
         if training_args.do_train:
             checkpoint = None
             last_checkpoint = self.last_checkpoint
@@ -312,7 +311,14 @@ def compute_metrics(eval_preds):
                 if model_args.save_aggregated_lora:
                     model.merge_lora_weights()
                 model.save(finetuner_args.output_dir,model_args.save_aggregated_lora)
-
+            # save language_projection for multi-modal model;
+            if self.finetuner_args.save_language_projection:
+                language_projection_state = trainer.model.language_projection.state_dict()
+                torch.save(
+                    osp.join(
+                        self.finetuner_args.output_dir,
+                        "language_projection.pth"),
+                    language_projection_state)
             metrics = train_result.metrics
 
             max_train_samples = (

src/lmflow/pipeline/inferencer.py

@@ -123,7 +123,7 @@ def inference(
         temperature: float=0.0,
         prompt_structure: str='{input}',
         remove_image_flag: bool=False,
-        chatbot_format: str="mini_gpt",
+        chatbot_type: str="mini_gpt",
     ):
         """
         Perform inference for a model
@@ -165,9 +165,9 @@ def inference(
                 input['images'] = np.array(input['images'])
             if remove_image_flag:
                 # remove the image flag <ImageHere> in tokenization;
-                if chatbot_format == "mini_gpt":
+                if chatbot_type == "mini_gpt":
                     image_split_flag = "<ImageHere>"
-                elif chatbot_format:
+                elif chatbot_type:
                     image_split_flag = "<image>"
                 else:
                     raise NotImplementedError
@@ -186,7 +186,7 @@ def inference(
                     ).to(device=self.local_rank)
                     input_ids.append(temp_inputs['input_ids'])
                     attention_mask.append(temp_inputs['attention_mask'])
-                    if chatbot_format == "llava":
+                    if chatbot_type == "llava":
                         # add the flag for inserting the image.
                         # TODO should merge the way of handling image flag in minigpt and llava.
                         index_tensor = torch.tensor(
@@ -200,7 +200,7 @@ def inference(
                         temp_inputs["input_ids"].shape[1])
                 if len(image_token_indexes) > 1:
                     image_token_indexes = image_token_indexes[:-1]
-                    if chatbot_format == "llava":
+                    if chatbot_type == "llava":
                         input_ids = input_ids[:-1]
                         attention_mask = attention_mask[:-1]
                 inputs = temp_inputs
@@ -219,7 +219,6 @@ def inference(
                     raise NotImplementedError(
                         f"device \"{self.inferencer_args.device}\" is not supported"
                     )
-
             if remove_image_flag:
                 inputs["image_token_indexes"] = image_token_indexes
                 inputs["one_sample_multiple_images"] = True
@@ -553,4 +552,4 @@ def speculative_sampling(input_ids: torch.Tensor,
 
 
     def stream_inference(self):
-        raise NotImplementedError("Streaming output for SpeculativeInferencer is not supported yet")
+        raise NotImplementedError("Streaming output for SpeculativeInferencer is not supported yet")

src/lmflow/utils/multimodal.py

@@ -42,10 +42,9 @@ def adapt_llava_model_to_lmflow_type(state_dict):
         key = key.replace("model.embed_tokens",
                           "language_model.model.embed_tokens")
         key = key.replace("model.mm_projector", "language_projection")
-        key = key.replace("lm_head", "model.language_model.lm_head")
-        key = key.replace("model.norm", "language_model.model.layers")
+        key = key.replace("lm_head", "language_model.lm_head")
+        key = key.replace("model.norm", "language_model.model.norm")
         if "vision_tower" in key:
             continue
         new_state_dict[key] = item
-    return new_state_dict
-
+    return new_state_dict