add yi-vl

Former-commit-id: 891b25cb3d709ea82182ca90496034360e1cd5d8

src/llmtuner/data/template.py

@@ -856,6 +856,20 @@ _register_template(
 )
 
 
+_register_template(
+    name="yi-vl",
+    format_user=StringFormatter(slots=["### Human:\n{{content}}\n### Assistant: "]),
+    stop_words=["###"],
+    default_system=(
+        "This is a chat between an inquisitive human and an AI assistant. "
+        "Assume the role of the AI assistant. "
+        "Read all the images carefully, and respond to the human's questions with informative, helpful, detailed and polite answers."
+        "这是一个好奇的人类和一个人工智能助手之间的对话。"
+        "假设你扮演这个AI助手的角色。仔细阅读所有的图像，并对人类的问题做出信息丰富、有帮助、详细的和礼貌的回答。"
+    ),
+)
+
+
 _register_template(
     name="yuan",
     format_user=StringFormatter(slots=["{{content}}", {"token": "<sep>"}]),

src/llmtuner/model/patcher.py

@@ -16,7 +16,7 @@ from .utils.moe import add_z3_leaf_module, configure_moe
 from .utils.quantization import configure_quantization
 from .utils.rope import configure_rope
 from .utils.valuehead import prepare_valuehead_model
-from .utils.visual import autocast_projector_dtype, configure_hidden_size
+from .utils.visual import autocast_projector_dtype, configure_hidden_size, configure_visual
 
 
 if TYPE_CHECKING:
@@ -50,6 +50,7 @@ def patch_config(
     configure_quantization(config, tokenizer, model_args, init_kwargs)
     configure_moe(config, model_args, is_trainable)
     configure_hidden_size(config)
+    configure_visual(config, model_args)
 
     if model_args.use_cache and not is_trainable:
         setattr(config, "use_cache", True)

src/llmtuner/model/utils/visual.py

@@ -1,12 +1,14 @@
 from typing import TYPE_CHECKING, Tuple
 
 import torch
+import transformers
+from torch import nn
 
 from ...extras.logging import get_logger
 
 
 if TYPE_CHECKING:
-    from transformers import PretrainedConfig, PreTrainedModel
+    from transformers import PretrainedConfig, PreTrainedModel, LlavaConfig
 
     from ...hparams import ModelArguments
 
@@ -31,3 +33,25 @@ def autocast_projector_dtype(
         logger.info("Casting multimodal projector outputs in {}.".format(model_args.compute_dtype))
         mm_projector: "torch.nn.Module" = getattr(model, mm_projector_name)
         mm_projector.register_forward_hook(_mm_projector_forward_post_hook)
+
+
+class LlavaMultiModalProjectorYiVL(nn.Module):
+    def __init__(self, config: "LlavaConfig"):
+        super().__init__()
+        self.linear_1 = nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True)
+        self.linear_2 = nn.LayerNorm(config.text_config.hidden_size, bias=True)
+        self.linear_3 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)
+        self.linear_4 = nn.LayerNorm(config.text_config.hidden_size, bias=True)
+        self.act = nn.GELU()
+        self.proj = nn.Sequential(*[self.linear_1, self.linear_2, self.act, self.linear_3, self.linear_4])
+
+    def forward(self, image_features):
+        hidden_states = self.proj(image_features)
+        return hidden_states
+
+
+def configure_visual(config: "PretrainedConfig", model_args: "ModelArguments") -> None:
+    logger = get_logger(__name__)
+    if model_args.visual_inputs and "Yi" in getattr(config.text_config, "_name_or_path", None):
+        transformers.models.llava.modeling_llava.LlavaMultiModalProjector = LlavaMultiModalProjectorYiVL
+        logger.info("Patched Multimodal Projector for Yi-VL.")

src/llmtuner/train/sft/trainer.py

@@ -5,7 +5,7 @@ from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
-from transformers import Seq2SeqTrainer
+from transformers import Seq2SeqTrainer, ProcessorMixin
 
 from ...extras.constants import IGNORE_INDEX
 from ...extras.logging import get_logger
@@ -26,9 +26,10 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
     Inherits Seq2SeqTrainer to compute generative metrics such as BLEU and ROUGE.
     """
 
-    def __init__(self, finetuning_args: "FinetuningArguments", **kwargs) -> None:
+    def __init__(self, finetuning_args: "FinetuningArguments", processor: "ProcessorMixin", **kwargs) -> None:
         super().__init__(**kwargs)
         self.finetuning_args = finetuning_args
+        self.processor = processor
         if finetuning_args.use_badam:
             from badam import clip_grad_norm_for_sparse_tensor
 
@@ -120,3 +121,10 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
             for label, pred in zip(decoded_labels, decoded_preds):
                 res.append(json.dumps({"label": label, "predict": pred}, ensure_ascii=False))
             writer.write("\n".join(res))
+
+    def save_model(self, output_dir: Optional[str] = None, _internal_call: bool = False):
+        super().save_model(output_dir, _internal_call)
+        if self.processor is not None:
+            if output_dir is None:
+                output_dir = self.args.output_dir
+            getattr(self.processor, "image_processor").save_pretrained(output_dir)

src/llmtuner/train/sft/workflow.py

@@ -30,6 +30,7 @@ def run_sft(
 ):
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
+    processor = tokenizer_module["processor"]
     dataset = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
 
@@ -55,6 +56,7 @@ def run_sft(
         model=model,
         args=training_args,
         finetuning_args=finetuning_args,
+        processor=processor,
         tokenizer=tokenizer,
         data_collator=data_collator,
         callbacks=callbacks,