[pt/sft] support use_logits_to_keep & support DeepSeek-R1-0528

docs/source/Instruction/命令行参数.md

@@ -356,10 +356,11 @@ Vera使用`target_modules`, `target_regex`, `modules_to_save`三个参数.
   - 支持的多模态模型参考：https://github.com/modelscope/ms-swift/blob/main/examples/train/packing/qwen2_5_vl.sh
 - packing_cache: 指定 packing 缓存目录。默认值为`None`，表示缓存将存储在环境变量 `$MODELSCOPE_CACHE`所指定的路径下。在跨节点使用 packing 功能时，需确保所有节点的 packing 缓存路径共享且一致。你可以通过设置`MODELSCOPE_CACHE`环境变量，或在命令行中添加 `--packing_cache <shared_path>`参数来实现这一要求。
 - 🔥lazy_tokenize: 是否使用lazy_tokenize。若该参数设置为False，则在训练之前对所有的数据集样本进行tokenize（多模态模型则包括从磁盘中读取图片）。该参数在LLM训练中默认设置为False，而MLLM训练默认为True，节约内存。
+- use_logits_to_keep: 通过在`forward`中根据labels传入logits_to_keep，减少无效logits的计算与存储，从而减少显存占用并加快训练速度。默认为None，进行自动选择。
 - acc_strategy: 训练和验证时计算acc的策略。可选为`seq`和`token`级别的acc，默认为`token`。
 - max_new_tokens: 覆盖生成参数。predict_with_generate=True时的最大生成token数量，默认64。
 - temperature: 覆盖生成参数。predict_with_generate=True时的temperature，默认0。
-- optimizer: plugin的自定义optimizer名称，默认为None。
+- optimizer: plugin的自定义optimizer名称，默认为None。可选optimizer参考[这里](https://github.com/modelscope/ms-swift/blob/main/swift/plugin/optimizer.py)。
 - metric: plugin的自定义metric名称。默认为None，即在predict_with_generate=False的情况下设置为'acc'，在predict_with_generate=True的情况下设置为'nlg'。
 - eval_use_evalscope: 是否使用evalscope进行训练时评测，需要设置该参数来开启评测，具体使用参考[示例](../Instruction/评测.md#训练中评测)。
 - eval_datasets: 评测数据集，可设置多个数据集，用空格分割。

docs/source/Instruction/支持的模型和数据集.md

@@ -398,12 +398,13 @@
 |[deepseek-ai/DeepSeek-Prover-V2-671B](https://modelscope.cn/models/deepseek-ai/DeepSeek-Prover-V2-671B)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|✘|-|[deepseek-ai/DeepSeek-Prover-V2-671B](https://huggingface.co/deepseek-ai/DeepSeek-Prover-V2-671B)|
 |[deepseek-ai/DeepSeek-R1](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1)|deepseek_r1|deepseek_r1|transformers>=4.39.3|✘|-|[deepseek-ai/DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1)|
 |[deepseek-ai/DeepSeek-R1-Zero](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Zero)|deepseek_r1|deepseek_r1|transformers>=4.39.3|✘|-|[deepseek-ai/DeepSeek-R1-Zero](https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero)|
+|[deepseek-ai/DeepSeek-R1-0528](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-0528)|deepseek_r1|deepseek_r1|transformers>=4.39.3|✘|-|[deepseek-ai/DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528)|
 |[cognitivecomputations/DeepSeek-R1-awq](https://modelscope.cn/models/cognitivecomputations/DeepSeek-R1-awq)|deepseek_r1|deepseek_r1|transformers>=4.39.3|✘|-|[cognitivecomputations/DeepSeek-R1-AWQ](https://huggingface.co/cognitivecomputations/DeepSeek-R1-AWQ)|
 |[deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Qwen-7B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Qwen-14B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Qwen-14B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)|
-|[iic/QwenLong-L1-32B](https://modelscope.cn/models/iic/QwenLong-L1-32B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✘|-|[Tongyi-Zhiwen/QwenLong-L1-32B](https://huggingface.co/Tongyi-Zhiwen/QwenLong-L1-32B)|
+|[iic/QwenLong-L1-32B](https://modelscope.cn/models/iic/QwenLong-L1-32B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[Tongyi-Zhiwen/QwenLong-L1-32B](https://huggingface.co/Tongyi-Zhiwen/QwenLong-L1-32B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Llama-8B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)|deepseek_r1_distill|deepseek_r1|-|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Llama-70B)|deepseek_r1_distill|deepseek_r1|-|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B)|
 |[OpenBuddy/openbuddy-llama-65b-v8-bf16](https://modelscope.cn/models/OpenBuddy/openbuddy-llama-65b-v8-bf16)|openbuddy_llama|openbuddy|-|✔|-|[OpenBuddy/openbuddy-llama-65b-v8-bf16](https://huggingface.co/OpenBuddy/openbuddy-llama-65b-v8-bf16)|

docs/source_en/Instruction/Command-line-parameters.md

@@ -365,10 +365,11 @@ Training arguments include the [base arguments](#base-arguments), [Seq2SeqTraine
   - Supported multimodal models reference: https://github.com/modelscope/ms-swift/blob/main/examples/train/packing/qwen2_5_vl.sh
 - packing_cache: Specifies the directory for packing cache. The default value is `None`, which means the cache will be stored in the path defined by the environment variable `$MODELSCOPE_CACHE`. When using the packing feature across multiple nodes, ensure that all nodes share the same packing cache directory. You can achieve this by setting the `MODELSCOPE_CACHE` environment variable or by adding the `--packing_cache <shared_path>` argument in the command line.
 - 🔥lazy_tokenize: Whether to use lazy tokenization. If set to False, all dataset samples are tokenized before training (for multimodal models, this includes reading images from disk). This parameter defaults to False for LLM training, and True for MLLM training, to save memory.
+- use_logits_to_keep: Pass `logits_to_keep` in the `forward` method based on labels to reduce the computation and storage of unnecessary logits, thereby reducing memory usage and accelerating training. The default is `None`, which enables automatic selection.
 - acc_strategy: Strategy for calculating accuracy during training and validation. Options are `seq`-level and `token`-level accuracy, with `token` as the default.
 - max_new_tokens: Generation parameter override. The maximum number of tokens to generate when `predict_with_generate=True`, defaulting to 64.
 - temperature: Generation parameter override. The temperature setting when `predict_with_generate=True`, defaulting to 0.
-- optimizer: Custom optimizer name for the plugin, defaults to None.
+- optimizer: Custom optimizer name for the plugin, defaults to None. Optional optimizer reference: [here](https://github.com/modelscope/ms-swift/blob/main/swift/plugin/optimizer.py).
 - metric: Custom metric name for the plugin. Defaults to None, with the default set to 'acc' when `predict_with_generate=False` and 'nlg' when `predict_with_generate=True`.
 - eval_use_evalscope: Whether to use evalscope for evaluation, this parameter needs to be set to enable evaluation, refer to [example](../Instruction/Evaluation.md#evaluation-during-training). Default is False.
 - eval_datasets: Evaluation datasets, multiple datasets can be set, separated by spaces

docs/source_en/Instruction/Supported-models-and-datasets.md

@@ -398,12 +398,13 @@ The table below introduces the models integrated with ms-swift:
 |[deepseek-ai/DeepSeek-Prover-V2-671B](https://modelscope.cn/models/deepseek-ai/DeepSeek-Prover-V2-671B)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|✘|-|[deepseek-ai/DeepSeek-Prover-V2-671B](https://huggingface.co/deepseek-ai/DeepSeek-Prover-V2-671B)|
 |[deepseek-ai/DeepSeek-R1](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1)|deepseek_r1|deepseek_r1|transformers>=4.39.3|✘|-|[deepseek-ai/DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1)|
 |[deepseek-ai/DeepSeek-R1-Zero](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Zero)|deepseek_r1|deepseek_r1|transformers>=4.39.3|✘|-|[deepseek-ai/DeepSeek-R1-Zero](https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero)|
+|[deepseek-ai/DeepSeek-R1-0528](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-0528)|deepseek_r1|deepseek_r1|transformers>=4.39.3|✘|-|[deepseek-ai/DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528)|
 |[cognitivecomputations/DeepSeek-R1-awq](https://modelscope.cn/models/cognitivecomputations/DeepSeek-R1-awq)|deepseek_r1|deepseek_r1|transformers>=4.39.3|✘|-|[cognitivecomputations/DeepSeek-R1-AWQ](https://huggingface.co/cognitivecomputations/DeepSeek-R1-AWQ)|
 |[deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Qwen-7B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Qwen-14B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Qwen-14B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)|
-|[iic/QwenLong-L1-32B](https://modelscope.cn/models/iic/QwenLong-L1-32B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✘|-|[Tongyi-Zhiwen/QwenLong-L1-32B](https://huggingface.co/Tongyi-Zhiwen/QwenLong-L1-32B)|
+|[iic/QwenLong-L1-32B](https://modelscope.cn/models/iic/QwenLong-L1-32B)|deepseek_r1_distill|deepseek_r1|transformers>=4.37|✔|-|[Tongyi-Zhiwen/QwenLong-L1-32B](https://huggingface.co/Tongyi-Zhiwen/QwenLong-L1-32B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Llama-8B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)|deepseek_r1_distill|deepseek_r1|-|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)|
 |[deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Llama-70B)|deepseek_r1_distill|deepseek_r1|-|✔|-|[deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B)|
 |[OpenBuddy/openbuddy-llama-65b-v8-bf16](https://modelscope.cn/models/OpenBuddy/openbuddy-llama-65b-v8-bf16)|openbuddy_llama|openbuddy|-|✔|-|[OpenBuddy/openbuddy-llama-65b-v8-bf16](https://huggingface.co/OpenBuddy/openbuddy-llama-65b-v8-bf16)|

swift/llm/argument/train_args.py

@@ -112,7 +112,6 @@ class TrainArguments(SwanlabArguments, TunerArguments, BaseArguments, Seq2SeqTra
 
     # plugin
     loss_type: Optional[str] = field(default=None, metadata={'help': f'loss_func choices: {list(LOSS_MAPPING.keys())}'})
-    optimizer: Optional[str] = None
     metric: Optional[str] = None
 
     # extra

swift/llm/model/model/deepseek.py

@@ -251,6 +251,7 @@ def get_model_tokenizer_deepseek_vl2(model_dir: str, *args, **kwargs):
             ModelGroup([
                 Model('deepseek-ai/DeepSeek-R1', 'deepseek-ai/DeepSeek-R1'),
                 Model('deepseek-ai/DeepSeek-R1-Zero', 'deepseek-ai/DeepSeek-R1-Zero'),
+                Model('deepseek-ai/DeepSeek-R1-0528', 'deepseek-ai/DeepSeek-R1-0528')
             ]),
             ModelGroup([
                 Model('cognitivecomputations/DeepSeek-R1-awq', 'cognitivecomputations/DeepSeek-R1-AWQ'),

swift/llm/model/utils.py

@@ -332,7 +332,7 @@ def get_llm_model(model: torch.nn.Module, model_meta=None):
 
     llm_prefix = getattr(get_model_arch(model_meta.model_arch), 'language_model', None)
     if llm_prefix:
-        llm_model = getattr(model, llm_prefix[0])
+        llm_model = deep_getattr(model, llm_prefix[0])
     else:
         llm_model = model
 

swift/llm/train/sft.py

@@ -75,13 +75,6 @@ def _get_dataset(self):
 
         return train_dataset, val_dataset
 
-    def _get_loss_func(self):
-        args = self.args
-        loss_type = args.loss_type
-        if loss_type is None and args.loss_scale != 'default':
-            loss_type = 'loss_scale'
-        return get_loss_func(loss_type)
-
     def _get_data_collator(self):
         args = self.args
         template = self.template
@@ -141,7 +134,7 @@ def _get_trainer_kwargs(self):
         return {
             'compute_metrics': compute_metrics,
             'preprocess_logits_for_metrics': preprocess_logits_for_metrics,
-            'compute_loss_func': self._get_loss_func()
+            'compute_loss_func': get_loss_func(args.loss_type)
         }
 
     def _save_trainer_state(self, trainer):

swift/trainers/arguments.py

@@ -50,6 +50,7 @@ class TrainArgumentsMixin:
     aligner_lr: Optional[float] = None
     vit_lr: Optional[float] = None
     optimizer: Optional[str] = None
+    use_logits_to_keep: Optional[bool] = None
 
     # torchacc
     metric_warmup_step: Optional[float] = 0

swift/trainers/trainers.py

@@ -1,5 +1,6 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 # Part of the implementation is borrowed from huggingface/transformers.
+import inspect
 import os
 from contextlib import contextmanager, nullcontext
 from functools import wraps
@@ -15,10 +16,12 @@
 from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
 from transformers.utils import is_peft_available
 
-from swift.utils import JsonlWriter, Serializer, gc_collect
+from swift.utils import JsonlWriter, Serializer, gc_collect, get_logger, is_mp
 from .arguments import Seq2SeqTrainingArguments, TrainingArguments
 from .mixin import DataLoaderMixin, SwiftMixin
 
+logger = get_logger()
+
 
 class Trainer(SwiftMixin, HfTrainer):
     args: TrainingArguments
@@ -152,15 +155,39 @@ def prediction_step(
         return None, response_list, labels_list
 
     def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
+        from swift.plugin.loss import get_loss_func
         loss_kwargs = {}
         labels = None
-        if (self.label_smoother is not None or self.compute_loss_func is not None) and 'labels' in inputs:
-            labels = inputs.pop('labels')
-
+        compute_loss_func = self.compute_loss_func
         loss_scale = inputs.pop('loss_scale', None)
         if loss_scale is not None:
             loss_kwargs['loss_scale'] = loss_scale
+            if compute_loss_func is None:
+                compute_loss_func = get_loss_func('loss_scale')
+        if (self.label_smoother is not None or compute_loss_func is not None) and 'labels' in inputs:
+            labels = inputs.pop('labels')
 
+        base_model = self.template.get_base_model(self.model)
+        use_logits_to_keep = self.args.use_logits_to_keep
+        if use_logits_to_keep is None:
+            # padding_free or packing
+            use_logits_to_keep = 'labels' in inputs and 'logits_to_keep' in inspect.signature(
+                base_model.forward).parameters
+        logger.info_once(f'use_logits_to_keep: {use_logits_to_keep}')
+
+        if use_logits_to_keep:
+            if inputs['labels'].shape[0] == 1:
+                loss_mask = (inputs['labels'] != -100)[0]
+                inputs['labels'] = inputs['labels'][:, loss_mask]
+                inputs['labels'] = nn.functional.pad(inputs['labels'], (1, 0), value=-100)
+                inputs['logits_to_keep'] = nn.functional.pad(loss_mask[1:], (0, 1), value=True)
+                if is_mp():
+                    inputs['logits_to_keep'] = inputs['logits_to_keep'].cpu()
+            else:
+                inputs['logits_to_keep'] = (inputs['labels'].shape[-1] -
+                                            (torch.ne(inputs['labels'], -100).int().argmax(-1))).max().item() + 1
+                assert inputs['logits_to_keep'] > 0
+                inputs['labels'] = inputs['labels'][:, -inputs['logits_to_keep']:]
         with self.template.compute_loss_context(self.model, inputs):
             outputs = model(**inputs)
         # Save past state if it exists
@@ -188,8 +215,8 @@ def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=N
             else:
                 model_name = unwrapped_model._get_name()
             # User-defined compute_loss function
-            if self.compute_loss_func is not None:
-                loss = self.compute_loss_func(outputs, labels, num_items_in_batch=num_items_in_batch, **loss_kwargs)
+            if compute_loss_func is not None:
+                loss = compute_loss_func(outputs, labels, num_items_in_batch=num_items_in_batch, **loss_kwargs)
             elif model_name in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES.values():
                 loss = self.label_smoother(outputs, labels, shift_labels=True)
             else: