update version

This reverts commit 1800beb13f407ddb881d0af936860643e84ba085.

docs/source/en/optimizers.md

@@ -154,7 +154,7 @@ pip install schedulefree
 
 [Schedule Free optimizer (SFO)](https://hf.co/papers/2405.15682) replaces the base optimizers momentum with a combination of averaging and interpolation. Unlike a traditional scheduler, SFO completely removes the need to anneal the learning rate.
 
-SFO supports the RAdam (`schedule_free_radam`), AdamW (`schedule_free_adamw`) and SGD (`schedule_free_sgd`) optimizers. The RAdam scheduler doesn't require `warmup_steps` or `warmup_ratio`.
+SFO supports the RAdam (`schedule_free_radam`), AdamW (`schedule_free_adamw`) and SGD (`schedule_free_sgd`) optimizers. The RAdam scheduler doesn't require `warmup_steps`.
 
 By default, it is recommended to set `lr_scheduler_type="constant"`. Other `lr_scheduler_type` values may also work, but combining SFO optimizers with other learning rate schedules could affect SFOs intended behavior and performance.
 

docs/source/en/tasks/audio_classification.md

@@ -220,7 +220,7 @@ At this point, only three steps remain:
 ...     gradient_accumulation_steps=4,
 ...     per_device_eval_batch_size=32,
 ...     num_train_epochs=10,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/en/tasks/image_classification.md

@@ -211,7 +211,7 @@ At this point, only three steps remain:
 ...     gradient_accumulation_steps=4,
 ...     per_device_eval_batch_size=16,
 ...     num_train_epochs=3,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/en/tasks/video_classification.md

@@ -378,7 +378,7 @@ Most of the training arguments are self-explanatory, but one that is quite impor
 ...     learning_rate=5e-5,
 ...     per_device_train_batch_size=batch_size,
 ...     per_device_eval_batch_size=batch_size,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/es/tasks/audio_classification.md

@@ -220,7 +220,7 @@ Al llegar a este punto, solo quedan tres pasos:
 ...     gradient_accumulation_steps=4,
 ...     per_device_eval_batch_size=32,
 ...     num_train_epochs=10,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/ja/main_classes/deepspeed.md

@@ -1292,7 +1292,7 @@ DeepSpeed は、`LRRangeTest`、`OneCycle`、`WarmupLR`、および`WarmupDecayL
   したがって、スケジューラを設定しない場合、これがデフォルトで設定されるスケジューラになります。
 
 設定ファイルで `scheduler` エントリを設定しない場合、[`Trainer`] は
-`--lr_scheduler_type`、`--learning_rate`、および `--warmup_steps` または `--warmup_ratio` の値を設定します。
+`--lr_scheduler_type`、`--learning_rate`、および `--warmup_steps` の値を設定します。
 🤗 それのトランスフォーマーバージョン。
 
 以下は、`WarmupLR`の自動構成された`scheduler`エントリの例です。
@@ -1316,8 +1316,7 @@ DeepSpeed は、`LRRangeTest`、`OneCycle`、`WarmupLR`、および`WarmupDecayL
 
 - `warmup_min_lr` の値は `0` です。
 - `warmup_max_lr` と `--learning_rate` の値。
-- `warmup_num_steps` と `--warmup_steps` の値 (指定されている場合)。それ以外の場合は `--warmup_ratio` を使用します
-  トレーニング ステップの数を乗算し、切り上げます。
+- `warmup_num_steps` と `--warmup_steps` の値 (指定されている場合)
 - `total_num_steps` には `--max_steps` の値を指定するか、指定されていない場合は実行時に自動的に導出されます。
   環境、データセットのサイズ、およびその他のコマンド ライン引数 (
   `WarmupDecayLR`)。

docs/source/ja/tasks/audio_classification.md

@@ -219,7 +219,7 @@ MInDS-14 データセットのサンプリング レートは 8khz です (こ
 ...     gradient_accumulation_steps=4,
 ...     per_device_eval_batch_size=32,
 ...     num_train_epochs=10,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/ja/tasks/image_classification.md

@@ -216,7 +216,7 @@ Datasets、🤗 データセット ライブラリから Food-101 データセ
 ...     gradient_accumulation_steps=4,
 ...     per_device_eval_batch_size=16,
 ...     num_train_epochs=3,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/ja/tasks/video_classification.md

@@ -360,7 +360,7 @@ You should probably TRAIN this model on a down-stream task to be able to use it
 ...     learning_rate=5e-5,
 ...     per_device_train_batch_size=batch_size,
 ...     per_device_eval_batch_size=batch_size,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/ko/optimizers.md

@@ -154,7 +154,7 @@ pip install schedulefree
 
 [Schedule Free optimizer (SFO)](https://hf.co/papers/2405.15682)는 기본 옵티마이저의 모멘텀 대신 평균화(averaging)와 보간(interpolation)을 조합하여 사용합니다. 덕분에 기존의 학습률 스케줄러와 달리, SFO는 학습률을 점진적으로 낮추는 절차가 아예 필요 없습니다.
 
-SFO는 RAdam(`schedule_free_radam`), AdamW(`schedule_free_adamw`), SGD(`schedule_free_sgd`) 옵티마이저를 지원합니다. RAdam 스케줄러는 `warmup_steps`나 `warmup_ratio` 설정이 필요하지 않습니다. 
+SFO는 RAdam(`schedule_free_radam`), AdamW(`schedule_free_adamw`), SGD(`schedule_free_sgd`) 옵티마이저를 지원합니다. RAdam 스케줄러는 `warmup_steps`.
 
 기본적으로 `lr_scheduler_type="constant"`로 설정하는 것을 권장합니다. 다른 `lr_scheduler_type` 값도 동작할 순 있으나, SFO 옵티마이저와 다른 학습률 스케줄을 함께 사용하면 SFO의 의도된 동작과 성능에 영향을 줄 수 있습니다. 
 

docs/source/ko/tasks/audio_classification.md

@@ -221,7 +221,7 @@ MinDS-14 데이터 세트의 샘플링 속도는 8khz이므로(이 정보는 [
 ...     gradient_accumulation_steps=4,
 ...     per_device_eval_batch_size=32,
 ...     num_train_epochs=10,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/ko/tasks/image_classification.md

@@ -212,7 +212,7 @@ Hugging Face 계정에 로그인하여 모델을 업로드하고 커뮤니티에
 ...     gradient_accumulation_steps=4,
 ...     per_device_eval_batch_size=16,
 ...     num_train_epochs=3,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/ko/tasks/video_classification.md

@@ -357,7 +357,7 @@ You should probably TRAIN this model on a down-stream task to be able to use it
 ...     learning_rate=5e-5,
 ...     per_device_train_batch_size=batch_size,
 ...     per_device_eval_batch_size=batch_size,
-...     warmup_ratio=0.1,
+...     warmup_steps=0.1,
 ...     logging_steps=10,
 ...     load_best_model_at_end=True,
 ...     metric_for_best_model="accuracy",

docs/source/zh/main_classes/deepspeed.md

@@ -1206,7 +1206,7 @@ DeepSpeed支持`LRRangeTest`、`OneCycle`、`WarmupLR`和`WarmupDecayLR`学习
 - 通过 `--lr_scheduler_type constant_with_warmup` 实现 `WarmupLR`
 - 通过 `--lr_scheduler_type linear` 实现 `WarmupDecayLR`。这也是 `--lr_scheduler_type` 的默认值，因此，如果不配置调度器，这将是默认配置的调度器。
 
-如果在配置文件中不配置 `scheduler` 条目，[`Trainer`] 将使用 `--lr_scheduler_type`、`--learning_rate` 和 `--warmup_steps` 或 `--warmup_ratio` 的值来配置其🤗 Transformers 版本。
+如果在配置文件中不配置 `scheduler` 条目，[`Trainer`] 将使用 `--lr_scheduler_type`、`--learning_rate` 和 `--warmup_steps` 的值来配置其🤗 Transformers 版本。
 
 以下是 `WarmupLR` 的自动配置示例：
 
@@ -1227,7 +1227,7 @@ DeepSpeed支持`LRRangeTest`、`OneCycle`、`WarmupLR`和`WarmupDecayLR`学习
 
 - `warmup_min_lr` 的值为 `0`。
 - `warmup_max_lr` 的值为 `--learning_rate`。
-- `warmup_num_steps` 的值为 `--warmup_steps`（如果提供）。否则，将使用 `--warmup_ratio` 乘以训练步骤的数量，并四舍五入。
+- `warmup_num_steps` 的值为 `--warmup_steps`（如果提供）。
 - `total_num_steps` 的值为 `--max_steps` 或者如果没有提供，将在运行时根据环境、数据集的大小和其他命令行参数（对于 `WarmupDecayLR` 来说需要）自动推导。
 
 当然，您可以接管任何或所有的配置值，并自行设置这些值：

examples/pytorch/audio-classification/README.md

@@ -42,7 +42,7 @@ python run_audio_classification.py \
     --learning_rate 3e-5 \
     --max_length_seconds 1 \
     --attention_mask False \
-    --warmup_ratio 0.1 \
+    --warmup_steps 0.1 \
     --num_train_epochs 5 \
     --per_device_train_batch_size 32 \
     --gradient_accumulation_steps 4 \
@@ -84,7 +84,7 @@ python run_audio_classification.py \
     --learning_rate 3e-4 \
     --max_length_seconds 16 \
     --attention_mask False \
-    --warmup_ratio 0.1 \
+    --warmup_steps 0.1 \
     --num_train_epochs 10 \
     --per_device_train_batch_size 8 \
     --gradient_accumulation_steps 4 \

examples/pytorch/image-pretraining/README.md

@@ -167,7 +167,7 @@ python run_mae.py \
     --lr_scheduler_type cosine \
     --weight_decay 0.05 \
     --num_train_epochs 800 \
-    --warmup_ratio 0.05 \
+    --warmup_steps 0.05 \
     --per_device_train_batch_size 8 \
     --per_device_eval_batch_size 8 \
     --logging_strategy steps \

src/transformers/modelcard.py

@@ -753,8 +753,6 @@ def extract_hyperparameters_from_trainer(trainer):
             hyperparameters["optimizer"] = f"Use {optimizer_name} and the args are:\n{optimizer_args}"
 
     hyperparameters["lr_scheduler_type"] = trainer.args.lr_scheduler_type.value
-    if trainer.args.warmup_ratio != 0.0:
-        hyperparameters["lr_scheduler_warmup_ratio"] = trainer.args.warmup_ratio
     if trainer.args.warmup_steps != 0.0:
         hyperparameters["lr_scheduler_warmup_steps"] = trainer.args.warmup_steps
     if trainer.args.max_steps != -1:

src/transformers/training_args.py

@@ -300,10 +300,9 @@ class TrainingArguments:
             The scheduler type to use. See the documentation of [`SchedulerType`] for all possible values.
         lr_scheduler_kwargs ('dict', *optional*, defaults to {}):
             The extra arguments for the lr_scheduler. See the documentation of each scheduler for possible values.
-        warmup_ratio (`float`, *optional*, defaults to 0.0):
-            Ratio of total training steps used for a linear warmup from 0 to `learning_rate`.
-        warmup_steps (`int`, *optional*, defaults to 0):
-            Number of steps used for a linear warmup from 0 to `learning_rate`. Overrides any effect of `warmup_ratio`.
+        warmup_steps (`int` or `float`, *optional*, defaults to 0):
+            Number of steps used for a linear warmup from 0 to `learning_rate`.  Should be an integer or a float in range `[0,1)`.
+            If smaller than 1, will be interpreted as ratio of steps used for a linear warmup from 0 to `learning_rate`.
         log_level (`str`, *optional*, defaults to `passive`):
             Logger log level to use on the main process. Possible choices are the log levels as strings: 'debug',
             'info', 'warning', 'error' and 'critical', plus a 'passive' level which doesn't set anything and keeps the
@@ -888,10 +887,14 @@ class TrainingArguments:
             )
         },
     )
-    warmup_ratio: float = field(
-        default=0.0, metadata={"help": "Linear warmup over warmup_ratio fraction of total steps."}
+    warmup_ratio: Optional[float] = field(
+        default=None,
+        metadata={
+            "help": "This argument is deprecated and will be removed in v5. Use `warmup_steps` instead as it also works with float values."
+        },
     )
-    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
+
+    warmup_steps: float = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
 
     log_level: str = field(
         default="passive",
@@ -1724,16 +1727,12 @@ class TrainingArguments:
         elif not isinstance(self.report_to, list):
             self.report_to = [self.report_to]
 
-        if self.warmup_ratio < 0 or self.warmup_ratio > 1:
-            raise ValueError("warmup_ratio must lie in range [0,1]")
-        elif self.warmup_ratio > 0 and self.warmup_steps > 0:
-            logger.info(
-                "Both warmup_ratio and warmup_steps given, warmup_steps will override any effect of warmup_ratio"
-                " during training"
-            )
+        if self.warmup_ratio is not None:
+            logger.warning("warmup_ratio is deprecated and will be removed in v5.2. Use `warmup_steps` instead.")
+            self.warmup_steps = self.warmup_ratio
 
-        if not isinstance(self.warmup_steps, int) or self.warmup_steps < 0:
-            raise ValueError("warmup_steps must be of type int and must be 0 or a positive integer.")
+        if self.warmup_steps < 0:
+            raise ValueError("warmup_steps must be an integer or a float")
 
         if isinstance(self.fsdp, bool):
             self.fsdp = [FSDPOption.FULL_SHARD] if self.fsdp else ""
@@ -2275,7 +2274,7 @@ class TrainingArguments:
         Get number of steps used for a linear warmup.
         """
         warmup_steps = (
-            self.warmup_steps if self.warmup_steps > 0 else math.ceil(num_training_steps * self.warmup_ratio)
+            int(self.warmup_steps) if self.warmup_steps >= 1 else math.ceil(num_training_steps * self.warmup_steps)
         )
         return warmup_steps
 
@@ -2771,8 +2770,8 @@ class TrainingArguments:
         name: Union[str, SchedulerType] = "linear",
         num_epochs: float = 3.0,
         max_steps: int = -1,
-        warmup_ratio: float = 0,
-        warmup_steps: int = 0,
+        warmup_steps: float = 0,
+        warmup_ratio: Optional[float] = None,
     ):
         """
         A method that regroups all arguments linked to the learning rate scheduler and its hyperparameters.
@@ -2787,11 +2786,9 @@ class TrainingArguments:
                 If set to a positive number, the total number of training steps to perform. Overrides `num_train_epochs`.
                 For a finite dataset, training is reiterated through the dataset (if all data is exhausted) until
                 `max_steps` is reached.
-            warmup_ratio (`float`, *optional*, defaults to 0.0):
-                Ratio of total training steps used for a linear warmup from 0 to `learning_rate`.
-            warmup_steps (`int`, *optional*, defaults to 0):
-                Number of steps used for a linear warmup from 0 to `learning_rate`. Overrides any effect of
-                `warmup_ratio`.
+            warmup_steps (`float`, *optional*, defaults to 0):
+                Number of steps used for a linear warmup from 0 to `learning_rate`.  Should be an integer or a float in range `[0,1)`.
+                If smaller than 1, will be interpreted as ratio of steps used for a linear warmup from 0 to `learning_rate`.
 
         Example:
 
@@ -2799,15 +2796,18 @@ class TrainingArguments:
         >>> from transformers import TrainingArguments
 
         >>> args = TrainingArguments("working_dir")
-        >>> args = args.set_lr_scheduler(name="cosine", warmup_ratio=0.05)
-        >>> args.warmup_ratio
+        >>> args = args.set_lr_scheduler(name="cosine", warmup_steps=0.05)
+        >>> args.warmup_steps
         0.05
         ```
         """
+        if warmup_ratio is not None:
+            logger.warning("warmup_ratio is deprecated and will be removed in v5. Use `warmup_steps` instead.")
+            warmup_steps = warmup_ratio
+
         self.lr_scheduler_type = SchedulerType(name)
         self.num_train_epochs = num_epochs
         self.max_steps = max_steps
-        self.warmup_ratio = warmup_ratio
         self.warmup_steps = warmup_steps
         return self