[deps] update to transformers 4.52 (#8125)

2025-12-22 23:00:36 +08:00 · 2025-05-21 05:16:18 +08:00
parent 56926d76f9
commit 9ae17cd173
28 changed files with 365 additions and 109 deletions
--- a/src/llamafactory/webui/common.py
+++ b/src/llamafactory/webui/common.py
@@ -205,6 +205,14 @@ def load_eval_results(path: os.PathLike) -> str:
    return f"```json\n{result}\n```\n"


+def calculate_pixels(pixels: str) -> int:
+    r"""Calculate the number of pixels from the expression."""
+    if "*" in pixels:
+        return int(pixels.split("*")[0]) * int(pixels.split("*")[1])
+    else:
+        return int(pixels)
+
+
 def create_ds_config() -> None:
    r"""Create deepspeed config in the current directory."""
    os.makedirs(DEFAULT_CACHE_DIR, exist_ok=True)
--- a/src/llamafactory/webui/components/train.py
+++ b/src/llamafactory/webui/components/train.py
@@ -106,11 +106,11 @@ def create_train_tab(engine: "Engine") -> dict[str, "Component"]:
                use_llama_pro = gr.Checkbox()

            with gr.Column():
+                enable_thinking = gr.Checkbox(value=True)
                report_to = gr.Dropdown(
-                    choices=["none", "all", "wandb", "mlflow", "neptune", "tensorboard"],
-                    value=["none"],
+                    choices=["none", "wandb", "mlflow", "neptune", "tensorboard", "all"],
+                    value="none",
                    allow_custom_value=True,
-                    multiselect=True,
                )

    input_elems.update(
@@ -126,6 +126,7 @@ def create_train_tab(engine: "Engine") -> dict[str, "Component"]:
            mask_history,
            resize_vocab,
            use_llama_pro,
+            enable_thinking,
            report_to,
        }
    )
@@ -143,6 +144,7 @@ def create_train_tab(engine: "Engine") -> dict[str, "Component"]:
            mask_history=mask_history,
            resize_vocab=resize_vocab,
            use_llama_pro=use_llama_pro,
+            enable_thinking=enable_thinking,
            report_to=report_to,
        )
    )
@@ -231,6 +233,42 @@ def create_train_tab(engine: "Engine") -> dict[str, "Component"]:
        )
    )

+    with gr.Accordion(open=False) as mm_tab:
+        with gr.Row():
+            freeze_vision_tower = gr.Checkbox(value=True)
+            freeze_multi_modal_projector = gr.Checkbox(value=True)
+            freeze_language_model = gr.Checkbox(value=False)
+
+        with gr.Row():
+            image_max_pixels = gr.Textbox(value="768*768")
+            image_min_pixels = gr.Textbox(value="32*32")
+            video_max_pixels = gr.Textbox(value="256*256")
+            video_min_pixels = gr.Textbox(value="16*16")
+
+    input_elems.update(
+        {
+            freeze_vision_tower,
+            freeze_multi_modal_projector,
+            freeze_language_model,
+            image_max_pixels,
+            image_min_pixels,
+            video_max_pixels,
+            video_min_pixels,
+        }
+    )
+    elem_dict.update(
+        dict(
+            mm_tab=mm_tab,
+            freeze_vision_tower=freeze_vision_tower,
+            freeze_multi_modal_projector=freeze_multi_modal_projector,
+            freeze_language_model=freeze_language_model,
+            image_max_pixels=image_max_pixels,
+            image_min_pixels=image_min_pixels,
+            video_max_pixels=video_max_pixels,
+            video_min_pixels=video_min_pixels,
+        )
+    )
+
    with gr.Accordion(open=False) as galore_tab:
        with gr.Row():
            use_galore = gr.Checkbox()
--- a/src/llamafactory/webui/locales.py
+++ b/src/llamafactory/webui/locales.py
@@ -871,6 +871,28 @@ LOCALES = {
            "info": "拡張ブロックのパラメータのみをトレーニングします。",
        },
    },
+    "enable_thinking": {
+        "en": {
+            "label": "Enable thinking",
+            "info": "Whether or not to enable thinking mode for reasoning models.",
+        },
+        "ru": {
+            "label": "Включить мысли",
+            "info": "Включить режим мысли для моделей решающего характера.",
+        },
+        "zh": {
+            "label": "启用思考模式",
+            "info": "是否启用推理模型的思考模式。",
+        },
+        "ko": {
+            "label": "생각 모드 활성화",
+            "info": "추론 모델의 생각 모드를 활성화할지 여부.",
+        },
+        "ja": {
+            "label": "思考モードを有効化",
+            "info": "推論モデルの思考モードを有効にするかどうか。",
+        },
+    },
    "report_to": {
        "en": {
            "label": "Enable external logger",
@@ -1374,6 +1396,177 @@ LOCALES = {
            "info": "PPO トレーニングにおいて報酬スコアをホワイトニング処理します。",
        },
    },
+    "mm_tab": {
+        "en": {
+            "label": "Multimodal configurations",
+        },
+        "ru": {
+            "label": "Конфигурации мультимедиа",
+        },
+        "zh": {
+            "label": "多模态参数设置",
+        },
+        "ko": {
+            "label": "멀티모달 구성",
+        },
+        "ja": {
+            "label": "多モーダル設定",
+        },
+    },
+    "freeze_vision_tower": {
+        "en": {
+            "label": "Freeze vision tower",
+            "info": "Freeze the vision tower in the model.",
+        },
+        "ru": {
+            "label": "Заморозить башню визиона",
+            "info": "Заморозить башню визиона в модели.",
+        },
+        "zh": {
+            "label": "冻结视觉编码器",
+            "info": "冻结模型中的视觉编码器。",
+        },
+        "ko": {
+            "label": "비전 타워 고정",
+            "info": "모델의 비전 타워를 고정합니다.",
+        },
+        "ja": {
+            "label": "ビジョンタワーの固定",
+            "info": "モデルのビジョンタワーを固定します。",
+        },
+    },
+    "freeze_multi_modal_projector": {
+        "en": {
+            "label": "Freeze multi-modal projector",
+            "info": "Freeze the multi-modal projector in the model.",
+        },
+        "ru": {
+            "label": "Заморозить мультимодальный проектор",
+            "info": "Заморозить мультимодальный проектор в модели.",
+        },
+        "zh": {
+            "label": "冻结多模态投影器",
+            "info": "冻结模型中的多模态投影器。",
+        },
+        "ko": {
+            "label": "멀티모달 프로젝터 고정",
+            "info": "모델의 멀티모달 프로젝터를 고정합니다.",
+        },
+        "ja": {
+            "label": "多モーダルプロジェクターの固定",
+            "info": "モデルの多モーダルプロジェクターを固定します。",
+        },
+    },
+    "freeze_language_model": {
+        "en": {
+            "label": "Freeze language model",
+            "info": "Freeze the language model in the model.",
+        },
+        "ru": {
+            "label": "Заморозить язык модели",
+            "info": "Заморозить язык модели в модели.",
+        },
+        "zh": {
+            "label": "冻结语言模型",
+            "info": "冻结模型中的语言模型。",
+        },
+        "ko": {
+            "label": "언어 모델 고정",
+            "info": "모델의 언어 모델을 고정합니다.",
+        },
+        "ja": {
+            "label": "言語モデルの固定",
+            "info": "モデルの言語モデルを固定します。",
+        },
+    },
+    "image_max_pixels": {
+        "en": {
+            "label": "Image max pixels",
+            "info": "The maximum number of pixels of image inputs.",
+        },
+        "ru": {
+            "label": "Максимальное количество пикселей изображения",
+            "info": "Максимальное количество пикселей изображения.",
+        },
+        "zh": {
+            "label": "图像最大像素",
+            "info": "输入图像的最大像素数。",
+        },
+        "ko": {
+            "label": "이미지 최대 픽셀",
+            "info": "이미지 입력의 최대 픽셀 수입니다.",
+        },
+        "ja": {
+            "label": "画像最大ピクセル",
+            "info": "画像入力の最大ピクセル数です。",
+        },
+    },
+    "image_min_pixels": {
+        "en": {
+            "label": "Image min pixels",
+            "info": "The minimum number of pixels of image inputs.",
+        },
+        "ru": {
+            "label": "Минимальное количество пикселей изображения",
+            "info": "Минимальное количество пикселей изображения.",
+        },
+        "zh": {
+            "label": "图像最小像素",
+            "info": "输入图像的最小像素数。",
+        },
+        "ko": {
+            "label": "이미지 최소 픽셀",
+            "info": "이미지 입력의 최소 픽셀 수입니다.",
+        },
+        "ja": {
+            "label": "画像最小ピクセル",
+            "info": "画像入力の最小ピクセル数です。",
+        },
+    },
+    "video_max_pixels": {
+        "en": {
+            "label": "Video max pixels",
+            "info": "The maximum number of pixels of video inputs.",
+        },
+        "ru": {
+            "label": "Максимальное количество пикселей видео",
+            "info": "Максимальное количество пикселей видео.",
+        },
+        "zh": {
+            "label": "视频最大像素",
+            "info": "输入视频的最大像素数。",
+        },
+        "ko": {
+            "label": "비디오 최대 픽셀",
+            "info": "비디오 입력의 최대 픽셀 수입니다.",
+        },
+        "ja": {
+            "label": "ビデオ最大ピクセル",
+            "info": "ビデオ入力の最大ピクセル数です。",
+        },
+    },
+    "video_min_pixels": {
+        "en": {
+            "label": "Video min pixels",
+            "info": "The minimum number of pixels of video inputs.",
+        },
+        "ru": {
+            "label": "Минимальное количество пикселей видео",
+            "info": "Минимальное количество пикселей видео.",
+        },
+        "zh": {
+            "label": "视频最小像素",
+            "info": "输入视频的最小像素数。",
+        },
+        "ko": {
+            "label": "비디오 최소 픽셀",
+            "info": "비디오 입력의 최소 픽셀 수입니다.",
+        },
+        "ja": {
+            "label": "ビデオ最小ピクセル",
+            "info": "ビデオ入力の最小ピクセル数です。",
+        },
+    },
    "galore_tab": {
        "en": {
            "label": "GaLore configurations",
@@ -2468,23 +2661,6 @@ LOCALES = {
            "label": "HTML タグをエスケープ",
        },
    },
-    "enable_thinking": {
-        "en": {
-            "label": "Enable thinking",
-        },
-        "ru": {
-            "label": "Включить мышление",
-        },
-        "zh": {
-            "label": "启用思考",
-        },
-        "ko": {
-            "label": "사고를 활성화하다",
-        },
-        "ja": {
-            "label": "思考を可能にする",
-        },
-    },
    "clear_btn": {
        "en": {
            "value": "Clear history",
--- a/src/llamafactory/webui/runner.py
+++ b/src/llamafactory/webui/runner.py
@@ -29,6 +29,7 @@ from .common import (
    DEFAULT_CACHE_DIR,
    DEFAULT_CONFIG_DIR,
    abort_process,
+    calculate_pixels,
    gen_cmd,
    get_save_dir,
    load_args,
@@ -162,7 +163,15 @@ class Runner:
            mask_history=get("train.mask_history"),
            resize_vocab=get("train.resize_vocab"),
            use_llama_pro=get("train.use_llama_pro"),
+            enable_thinking=get("train.enable_thinking"),
            report_to=get("train.report_to"),
+            freeze_vision_tower=get("train.freeze_vision_tower"),
+            freeze_multi_modal_projector=get("train.freeze_multi_modal_projector"),
+            freeze_language_model=get("train.freeze_language_model"),
+            image_max_pixels=calculate_pixels(get("train.image_max_pixels")),
+            image_min_pixels=calculate_pixels(get("train.image_min_pixels")),
+            video_max_pixels=calculate_pixels(get("train.video_max_pixels")),
+            video_min_pixels=calculate_pixels(get("train.video_min_pixels")),
            use_galore=get("train.use_galore"),
            use_apollo=get("train.use_apollo"),
            use_badam=get("train.use_badam"),
@@ -256,12 +265,6 @@ class Runner:
            args["badam_switch_interval"] = get("train.badam_switch_interval")
            args["badam_update_ratio"] = get("train.badam_update_ratio")

-        # report_to
-        if "none" in args["report_to"]:
-            args["report_to"] = "none"
-        elif "all" in args["report_to"]:
-            args["report_to"] = "all"
-
        # swanlab config
        if get("train.use_swanlab"):
            args["swanlab_project"] = get("train.swanlab_project")