merge model part to the text stream

Former-commit-id: 838eb87a961894b072bf79a60bbda63516670d6f

src/llmtuner/hparams/model_args.py

@@ -169,6 +169,10 @@ class ModelArguments:
         default=False,
         metadata={"help": "For debugging purposes, print the status of the parameters in the model."},
     )
+    use_mllm: bool = field(
+        default=False,
+        metadata={"help": "Whether use Multimodal LLM."},
+    )
 
     def __post_init__(self):
         self.compute_dtype = None

src/llmtuner/model/__init__.py

@@ -1,10 +1,9 @@
-from .loader import load_config, load_model, load_tokenizer, load_mm_model
+from .loader import load_config, load_model, load_tokenizer, load_processor
 from .utils.misc import find_all_linear_modules, load_valuehead_params
 
 __all__ = [
     "load_config",
     "load_model",
-    "load_mm_model",
     "load_tokenizer",
     "load_processor",
     "load_valuehead_params",

src/llmtuner/model/adapter.py

@@ -1,4 +1,4 @@
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Union
 
 import torch
 from peft import LoraConfig, LoraModel, PeftModel, TaskType, get_peft_model
@@ -21,11 +21,11 @@ logger = get_logger(__name__)
 
 def init_adapter(
     config: "PretrainedConfig",
-    model: "PreTrainedModel",
+    model: Union["PreTrainedModel","AutoModelForVision2Seq"],
     model_args: "ModelArguments",
     finetuning_args: "FinetuningArguments",
     is_trainable: bool,
-) -> "PreTrainedModel":
+) -> Union["PreTrainedModel","AutoModelForVision2Seq"]:
     r"""
     Initializes the adapters.
 
@@ -196,102 +196,3 @@ def init_adapter(
             logger.info("Loaded adapter(s): {}".format(",".join(model_args.adapter_name_or_path)))
 
     return model
-
-
-def init_mm_adapter(
-        model: "AutoModelForVision2Seq", model_args: "ModelArguments",
-        finetuning_args: "FinetuningArguments",
-        is_trainable: bool,
-        use_clm=True,
-) -> "AutoModelForVision2Seq":
-    if finetuning_args.finetuning_type == "lora":
-        logger.info("Fine-tuning method: {}".format("DoRA" if finetuning_args.use_dora else "LoRA"))
-        adapter_to_resume = None
-
-        if model_args.adapter_name_or_path is not None:
-            is_mergeable = True
-            if getattr(model, "quantization_method", None):  # merge lora in quantized model is unstable
-                assert len(model_args.adapter_name_or_path) == 1, "Quantized model only accepts a single adapter."
-                is_mergeable = False
-
-            if is_deepspeed_zero3_enabled():
-                assert len(model_args.adapter_name_or_path) == 1, "Cannot use multiple adapters in DeepSpeed ZeRO-3."
-                is_mergeable = False
-
-            if (is_trainable and not finetuning_args.create_new_adapter) or (not is_mergeable):
-                adapter_to_merge = model_args.adapter_name_or_path[:-1]
-                adapter_to_resume = model_args.adapter_name_or_path[-1]
-            else:
-                adapter_to_merge = model_args.adapter_name_or_path
-
-            for adapter in adapter_to_merge:
-                model: "LoraModel" = PeftModel.from_pretrained(
-                    model, adapter, offload_folder=model_args.offload_folder
-                )
-                model = model.merge_and_unload()
-
-            if len(adapter_to_merge) > 0:
-                logger.info("Merged {} adapter(s).".format(len(adapter_to_merge)))
-
-            if adapter_to_resume is not None:  # resume lora training
-                model = PeftModel.from_pretrained(
-                    model, adapter_to_resume, is_trainable=is_trainable, offload_folder=model_args.offload_folder
-                )
-
-        if is_trainable and adapter_to_resume is None:  # create new lora weights while training
-            if len(finetuning_args.lora_target) == 1 and finetuning_args.lora_target[0] == "all":
-                target_modules = find_all_linear_modules(model)
-            else:
-                target_modules = finetuning_args.lora_target
-
-            if finetuning_args.use_llama_pro:
-                target_modules = find_expanded_modules(model, target_modules, finetuning_args.num_layer_trainable)
-
-            if (
-                    finetuning_args.use_dora
-                    and getattr(model, "quantization_method", None) is not None
-                    and getattr(model, "quantization_method", None) != QuantizationMethod.BITS_AND_BYTES
-            ):
-                raise ValueError("DoRA is not compatible with PTQ-quantized models.")
-
-            peft_kwargs = {
-                "r": finetuning_args.lora_rank,
-                "target_modules": target_modules,
-                "lora_alpha": finetuning_args.lora_alpha,
-                "lora_dropout": finetuning_args.lora_dropout,
-                "use_rslora": finetuning_args.use_rslora,
-                "modules_to_save": finetuning_args.additional_target,
-            }
-
-            if model_args.use_unsloth:
-                from unsloth import FastLanguageModel  # type: ignore
-
-                unsloth_peft_kwargs = {
-                    "model": model,
-                    "max_seq_length": model_args.model_max_length,
-                    "use_gradient_checkpointing": "unsloth",
-                }
-                model = FastLanguageModel.get_peft_model(**peft_kwargs, **unsloth_peft_kwargs)
-            else:
-                if use_clm:
-                    lora_config = LoraConfig(
-                        task_type=TaskType.CAUSAL_LM,
-                        inference_mode=False,
-                        use_dora=finetuning_args.use_dora,
-                        **peft_kwargs,
-                    )
-                else:
-                    lora_config = LoraConfig(
-                        inference_mode=False,
-                        use_dora=finetuning_args.use_dora,
-                        **peft_kwargs,
-                    )
-                model = get_peft_model(model, lora_config)
-
-        if (not finetuning_args.pure_bf16) and (not finetuning_args.use_badam):
-            for param in filter(lambda p: p.requires_grad, model.parameters()):
-                param.data = param.data.to(torch.float32)
-
-        if model_args.adapter_name_or_path is not None:
-            logger.info("Loaded adapter(s): {}".format(",".join(model_args.adapter_name_or_path)))
-    return model

src/llmtuner/model/loader.py

@@ -1,11 +1,11 @@
-from typing import TYPE_CHECKING, Any, Dict
+from typing import TYPE_CHECKING, Any, Dict, Union
 
 from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, AutoProcessor, AutoModelForVision2Seq
 from trl import AutoModelForCausalLMWithValueHead
 
 from ..extras.logging import get_logger
 from ..extras.misc import count_parameters, try_download_model_from_ms
-from .adapter import init_adapter, init_mm_adapter
+from .adapter import init_adapter
 from .patcher import patch_config, patch_model, patch_tokenizer, patch_valuehead_model
 from .utils.misc import load_valuehead_params, register_autoclass
 from .utils.mod import convert_pretrained_model_to_mod, load_mod_pretrained_model
@@ -111,7 +111,7 @@ def load_model(
     finetuning_args: "FinetuningArguments",
     is_trainable: bool = False,
     add_valuehead: bool = False,
-) -> "PreTrainedModel":
+) -> Union["PreTrainedModel", "AutoModelForVision2Seq"]:
     r"""
     Loads pretrained model.
     """
@@ -133,6 +133,9 @@ def load_model(
 
         if model_args.mixture_of_depths == "load":
             model = load_mod_pretrained_model(**init_kwargs)
+        else:
+            if model_args.use_mllm:
+                model = AutoModelForVision2Seq.from_pretrained(**init_kwargs)
             else:
                 model = AutoModelForCausalLM.from_pretrained(**init_kwargs)
 
@@ -183,55 +186,3 @@ def load_model(
             )
 
     return model
-
-
-def load_mm_model(
-        processor: "AutoProcessor",
-        model_args: "ModelArguments",
-        finetuning_args: "FinetuningArguments",
-        is_trainable: bool = False,
-        add_valuehead: bool = False,
-        use_clm=True,
-) -> "AutoModelForVision2Seq":
-    r"""
-    Loads pretrained model. Must after load_tokenizer.
-    """
-    tokenizer = processor.tokenizer
-    init_kwargs = _get_init_kwargs(model_args)
-    config = AutoConfig.from_pretrained(model_args.model_name_or_path, **init_kwargs)
-    patch_config(config, tokenizer, model_args, init_kwargs, is_trainable)
-
-    model = None
-    if model is None:
-        init_kwargs["config"] = config
-        init_kwargs["pretrained_model_name_or_path"] = model_args.model_name_or_path
-        model: "AutoModelForVision2Seq" = AutoModelForVision2Seq.from_pretrained(**init_kwargs)
-    patch_model(model, tokenizer, model_args, is_trainable)
-    register_autoclass(config, model, tokenizer)
-
-    model = init_mm_adapter(model, model_args, finetuning_args, is_trainable, use_clm)
-
-    if not is_trainable:
-        model.requires_grad_(False)
-        model.eval()
-    else:
-        model.train()
-
-    trainable_params, all_param = count_parameters(model)
-    if is_trainable:
-        param_stats = "trainable params: {:d} || all params: {:d} || trainable%: {:.4f}".format(
-            trainable_params, all_param, 100 * trainable_params / all_param
-        )
-    else:
-        param_stats = "all params: {:d}".format(all_param)
-    logger.info(param_stats)
-
-    if model_args.print_param_status:
-        for name, param in model.named_parameters():
-            print(
-                "name: {}, dtype: {}, device: {}, trainable: {}".format(
-                    name, param.dtype, param.device, param.requires_grad
-                )
-            )
-
-    return model

src/llmtuner/train/sftmm/workflow.py

@@ -4,7 +4,7 @@ from typing import TYPE_CHECKING, List, Optional
 from ...data import split_dataset, get_mm_dataset
 from ...extras.misc import get_logits_processor
 from ...extras.ploting import plot_loss
-from ...model import load_tokenizer, load_processor, load_mm_model
+from ...model import load_tokenizer, load_processor, load_model
 from ..utils import create_modelcard_and_push
 from .metric import ComputeMetrics
 from .trainer import CustomSeq2SeqTrainer
@@ -29,10 +29,7 @@ def run_sft_mm(
     CHAT_TEMPLATE = """{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. {% for message in messages %}{% if message['role'] == 'user' %}USER: {% else %}ASSISTANT: {% endif %}{% for item in message['content'] %}{% if item['type'] == 'text' %}{{ item['text'] }}{% elif item['type'] == 'image' %}<image>{% endif %}{% endfor %}{% if message['role'] == 'user' %} {% else %}{{eos_token}}{% endif %}{% endfor %}{% if add_generation_prompt %}ASSISTANT: {% endif %}"""
     tokenizer.chat_template = CHAT_TEMPLATE
     processor.tokenizer = tokenizer
-    use_clm = True
+    model = load_model(processor.tokenizer, model_args, finetuning_args, training_args.do_train)
-    if "blip" in model_args.model_name_or_path:
-        use_clm = False
-    model = load_mm_model(processor, model_args, finetuning_args, training_args.do_train, use_clm=use_clm)
     dataset = get_mm_dataset(processor, model_args, data_args, training_args, stage="sft")
     if getattr(model, "is_quantized", False) and not training_args.do_train:
         setattr(model, "_hf_peft_config_loaded", True)  # hack here: make model compatible with prediction