update ui

src/llamafactory/webui/components/train.py

@@ -95,12 +95,11 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
 
         with gr.Row():
             with gr.Column():
-                resize_vocab = gr.Checkbox()
                 packing = gr.Checkbox()
-                efficient_packing = gr.Checkbox()
+                neat_packing = gr.Checkbox()
 
             with gr.Column():
-                upcast_layernorm = gr.Checkbox()
+                resize_vocab = gr.Checkbox()
                 use_llama_pro = gr.Checkbox()
 
             with gr.Column():
@@ -114,10 +113,9 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
             warmup_steps,
             neftune_alpha,
             optim,
-            resize_vocab,
             packing,
-            efficient_packing,
+            neat_packing,
-            upcast_layernorm,
+            resize_vocab,
             use_llama_pro,
             shift_attn,
             report_to,
@@ -131,10 +129,9 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
             warmup_steps=warmup_steps,
             neftune_alpha=neftune_alpha,
             optim=optim,
-            resize_vocab=resize_vocab,
             packing=packing,
-            efficient_packing=efficient_packing,
+            neat_packing=neat_packing,
-            upcast_layernorm=upcast_layernorm,
+            resize_vocab=resize_vocab,
             use_llama_pro=use_llama_pro,
             shift_attn=shift_attn,
             report_to=report_to,
@@ -331,7 +328,7 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
     )
 
     dataset.focus(list_datasets, [dataset_dir, training_stage], [dataset], queue=False)
-    training_stage.change(change_stage, [training_stage], [dataset, packing, efficient_packing], queue=False)
+    training_stage.change(change_stage, [training_stage], [dataset, packing], queue=False)
     reward_model.focus(list_checkpoints, [model_name, finetuning_type], [reward_model], queue=False)
     model_name.change(list_output_dirs, [model_name, finetuning_type, current_time], [output_dir], queue=False)
     finetuning_type.change(list_output_dirs, [model_name, finetuning_type, current_time], [output_dir], queue=False)

src/llamafactory/webui/locales.py

@@ -494,20 +494,6 @@ LOCALES = {
             "info": "使用的优化器：adamw_torch、adamw_8bit 或 adafactor。",
         },
     },
-    "resize_vocab": {
-        "en": {
-            "label": "Resize token embeddings",
-            "info": "Resize the tokenizer vocab and the embedding layers.",
-        },
-        "ru": {
-            "label": "Изменение размера токенных эмбеддингов",
-            "info": "Изменить размер словаря токенизатора и слоев эмбеддинга.",
-        },
-        "zh": {
-            "label": "更改词表大小",
-            "info": "更改分词器词表和嵌入层的大小。",
-        },
-    },
     "packing": {
         "en": {
             "label": "Pack sequences",
@@ -522,32 +508,32 @@ LOCALES = {
             "info": "将序列打包为等长样本。",
         },
     },
-    "efficient_packing": {
+    "neat_packing": {
         "en": {
-            "label": "Pack sequences for efficient training",
+            "label": "Use neat packing",
-            "info": "Pack sequences into samples of fixed length without cross-contamination attention for efficient training.",
+            "info": "Avoid cross-attention between packed sequences.",
         },
         "ru": {
-            "label": "Пакетные последовательности для эффективного обучения",
+            "label": "Используйте аккуратную упаковку",
-            "info": "Упакуйте последовательности в образцы фиксированной длины без учета перекрестного загрязнения для эффективного обучения.",
+            "info": "избегайте перекрестного внимания между упакованными последовательностями.",
         },
         "zh": {
-            "label": "打包序列以实现高效训练",
+            "label": "使用无污染打包",
-            "info": "为了提高训练效率，将序列打包成固定长度的样本，无需注意交叉污染。",
+            "info": "避免打包后的序列产生交叉注意力。",
         },
     },
-    "upcast_layernorm": {
+    "resize_vocab": {
         "en": {
-            "label": "Upcast LayerNorm",
+            "label": "Resize token embeddings",
-            "info": "Upcast weights of layernorm in float32.",
+            "info": "Resize the tokenizer vocab and the embedding layers.",
         },
         "ru": {
-            "label": "Приведение весов LayerNorm",
+            "label": "Изменение размера токенных эмбеддингов",
-            "info": "Приведение весов LayerNorm к float32.",
+            "info": "Изменить размер словаря токенизатора и слоев эмбеддинга.",
         },
         "zh": {
-            "label": "缩放归一化层",
+            "label": "更改词表大小",
-            "info": "将归一化层权重缩放至 32 位精度。",
+            "info": "更改分词器词表和嵌入层的大小。",
         },
     },
     "use_llama_pro": {

src/llamafactory/webui/runner.py

@@ -138,10 +138,9 @@ class Runner:
             warmup_steps=get("train.warmup_steps"),
             neftune_noise_alpha=get("train.neftune_alpha") or None,
             optim=get("train.optim"),
+            packing=get("train.packing") or get("train.neat_packing"),
+            neat_packing=get("train.neat_packing"),
             resize_vocab=get("train.resize_vocab"),
-            packing=get("train.packing"),
-            efficient_packing=get("train.efficient_packing"),
-            upcast_layernorm=get("train.upcast_layernorm"),
             use_llama_pro=get("train.use_llama_pro"),
             shift_attn=get("train.shift_attn"),
             report_to="all" if get("train.report_to") else "none",