Megatron-LM源码系列(四)：重计算(recompute)

github: https://github.com/NVIDIA/Megatron-LM

1. recompute参数配置

在megatron/arguments.py中有重计算的参数配置如下：

    group.add_argument('--recompute-activations', action='store_true',
                       help='recompute activation to allow for training '
                       'with larger models, sequences, and batch sizes.')
    group.add_argument('--recompute-granularity', type=str, default=None,
                       choices=['full', 'selective'],
                       help='Checkpoint activations to allow for training '
                       'with larger models, sequences, and batch sizes. '
                       'It is supported at two granularities 1) full: '
                       'whole transformer layer is recomputed, '
                       '2) selective: core attention part of the transformer '
                       'layer is recomputed.')
    group.add_argument('--distribute-saved-activations',
                       action='store_true',
                       help='If set, distribute recomputed activations '
                       'across model parallel group.')
    group.add_argument('--recompute-method', type=str, default=None,
                       choices=['uniform', 'block'],
                       help='1) uniform: uniformly divide the total number of '
                       'Transformer layers and recompute the input activation of '
                       'each divided chunk at specified granularity, '
                       '2) recompute the input activations of only a set number of '
                       'individual Transformer layers per pipeline stage and do the '
                       'rest without any recomputing at specified granularity'
                       'default) do not apply activations recompute to any layers')
    group.add_argument('--recompute-num-layers', type=int, default=1,
                       help='1) uniform: the number of Transformer layers in each '
                       'uniformly divided recompute unit, '
                       '2) block: the number of individual Transformer layers '
                       'to recompute within each pipeline stage.')

说明：

• --recompute-activations: 设置recompute_activations等同于recompute_granularity为selective；selective运行效率更高，大部分场景只设置这个就可以。如果显存更紧张时，再通过recompute-granularity来进行full的设置。
• --recompute-granularity: 支持不同颗粒度的重计算，设为full会重计算整个transformer层，设为selective只会重算transformer中的core_attention部分。
• --distribute-saved-activations: 按TP并行度分开存储activation。
• --recompute-method: uniform计算会把所有的transformer layer分为若干组，分别把每组的input activation保存在内存中, GPU显存不足时，可通过设大每个组内的layer数来运行更大的model；block是针对pipeline并行的每个stage，checkpoint部分transformer layer的input activation, 剩余部分不进行checkpoint缓存，对于一个pipeline stage中有8层的来说，当设为5时，前5层中每一层的input activation都会被缓存，后3层在反向的时候正常计算。
• --recompute-num-layers: 对于uniform类型，表示设置在每个重计算的transformer layer group中的层数, 默认为1表示对每一层transformer layer都分别进行checkpoint；对于block类型，设为N表示单个pipeline stage中的前N个layers会缓存input activation。

2. 源码详解

2.1 --recompute-activations

设置recompute_activations等同于recompute_granularity为selective，设置后会覆盖recompute_granularity的值。

    if args.recompute_activations:
        args.recompute_granularity = 'selective'
    del args.recompute_activations

2.2 --recompute-granularity

支持不同颗粒度的重计算，设为full会重计算整个transformer层，设为selective只会重算transformer中的core_attention部分。下图红框中为选择重计算的部分：

入口是在megatron/model/transformer.py文件中的ParallelTransformer函数中, 在forward中如果为full的话会对整个前向进行checkpoint操作。

class ParallelTransformer(MegatronModule):

    def forward(self, hidden_states, attention_mask,
                encoder_output=None, enc_dec_attn_mask=None,
                retriever_input=None,
                retriever_output=None,
                retriever_attn_mask=None,
                inference_params=None,
                rotary_pos_emb=None):
                ...
                # Forward pass.
                if self.recompute_granularity == 'full':
                    hidden_states = self._checkpointed_forward(hidden_states,
                                                               attention_mask,
                                                               encoder_output,
                                                               enc_dec_attn_mask,
                                                               rotary_pos_emb,
                                                               is_first_microbatch)
                ...                                                            

具体执行checkpoint的方法有两种，分别是uniform和block。对于uniform方法来说，每次会过self.recompute_num_layers个layer保存一次input activation。custom(l, l + self.recompute_num_layers), custom函数是用于执行自定义的前向计算。对于block方法，对于小于self.recompute_num_layers的layer执行input activation的checkpoint, 对于大于等于self.recompute_num_layers的layer还执行原有操作。

示例代码如下：

class ParallelTransformer(MegatronModule):

    def _checkpointed_forward(self, hidden_states, attention_mask,
                              encoder_output, enc_dec_attn_mask,
                              rotary_pos_emb, is_first_microbatch):
        def custom(start, end):
            def custom_forward(*args, **kwargs):
                x_, *args = args
                for index in range(start, end):
                    layer = self._get_layer(index)
                    x_ = layer(x_, *args, **kwargs)
                return x_
            return custom_forward
    
        if self.recompute_method == 'uniform':
            # Uniformly divide the total number of Transformer layers and
            # checkpoint the input activation of each divided chunk.
            # A method to further reduce memory usage reducing checkpoints.
            l = 0
            while l < self.num_layers:
                if self.transformer_impl == 'transformer_engine':
                    ...
                else:
                    hidden_states = tensor_parallel.checkpoint(
                        custom(l, l + self.recompute_num_layers),
                        self.distribute_saved_activations,
                        hidden_states, attention_mask,
                        encoder_output, enc_dec_attn_mask,
                        None, None, None, None, rotary_pos_emb)

                l += self.recompute_num_layers
        elif self.recompute_method == 'block':
            # Checkpoint the input activation of only a set number of individual
            # Transformer layers and skip the rest.
            # A method fully use the device memory removing redundant re-computation.
            for l in range(self.num_layers):
                if l < self.recompute_num_layers:
                    if self.transformer_impl == 'transformer_engine':
                        ...
                    else:
                        hidden_states = tensor_parallel.checkpoint(
                            custom(l, l + 1),
                            self.distribute_saved_activations,
                            hidden_states, attention_mask,
                            encoder_output, enc_dec_attn_mask,
                            None, None, None, None, rotary_pos_emb)
                else:
                    if self.transformer_impl == 'transformer_engine':
                        ...
                    else:
                        hidden_states = custom(l, l + 1)(
                            hidden_states, attention_mask,
                            encoder_output, enc_dec_attn_mask,
                            None, None, None, None, rotary_pos_emb)
        ...

对于selective颗粒度的重计算目前不支持重计算方法的选择。

    if args.recompute_granularity == 'selective':
        assert args.recompute_method is None, \
            'recompute method is not yet supported for ' \
            'selective recomputing granularity'

在计算ParallelTransformer时会只针对attention的input activation进行checkpoint。

class ParallelTransformer(MegatronModule):
    """Transformer class."""

    def __init__(...):
        ...
        self.checkpoint_core_attention = args.recompute_granularity == 'selective'
        ...

    def forward(self, hidden_states, attention_mask,
                encoder_output=None, inference_params=None,
                rotary_pos_emb=None):
        ...
        if not self.use_flash_attn:
            if self.checkpoint_core_attention:
                context_layer = self._checkpointed_attention_forward(
                    query_layer, key_layer, value_layer, attention_mask)
        ...
        
    def _checkpointed_attention_forward(self, query_layer, key_layer,
                                        value_layer, attention_mask,
                                        rotary_pos_emb=None):
        """Forward method with activation checkpointing."""
        def custom_forward(*inputs):
            query_layer = inputs[0]
            key_layer = inputs[1]
            value_layer = inputs[2]
            attention_mask = inputs[3]
            output_ = self.core_attention(query_layer, key_layer,
                                          value_layer, attention_mask)
            return output_

        q_pos_emb, k_pos_emb = (None, None) if rotary_pos_emb is None \
            else rotary_pos_emb

        hidden_states = tensor_parallel.checkpoint(
            custom_forward,
            False, query_layer, key_layer, value_layer, attention_mask,
            q_pos_emb, k_pos_emb)

        return hidden_states

2.3 --distribute-saved-activations

只支持设置TP>1, 并且是full粒度的重计算, 对于uniform和block重计算都支持。

    # Activation recomputing.
    if args.distribute_saved_activations:
        assert args.tensor_model_parallel_size > 1, 'can distribute ' \
            'recomputed activations only across tensor model ' \
            'parallel groups'
        assert args.recompute_granularity == 'full', \
            'distributed recompute activations is only '\
            'application to full recompute granularity'
        assert args.recompute_method is not None, \
            'for distributed recompute activations to work you '\
            'need to use a recompute method '
        assert TORCH_MAJOR >= 1 and TORCH_MINOR >= 10, \
            'distributed recompute activations are supported for pytorch ' \
            'v1.10 and above (Nvidia Pytorch container >= 21.07). Current ' \
            'pytorch version is v%s.%s.' % (TORCH_MAJOR, TORCH_MINOR)

实现是在megatron/core/tensor_parallel/random.py文件的CheckpointFunction中，前向计算时使用no_grad()的上下文，不保存grad，在存activation的时候会把tensor展成一维的，每个rank只存自己的那一部分。

class CheckpointFunction(torch.autograd.Function):
    def forward(ctx, run_function, distribute_saved_activations, *args):
        ...

        with torch.no_grad():
            outputs = run_function(*args)
            
        # Divide hidden states across model parallel group and only keep
        # the chunk corresponding to the current rank.
        if distribute_saved_activations:
            ctx.input_0_shape = args[0].data.shape
            safely_set_viewless_tensor_data(
                args[0],
                split_tensor_into_1d_equal_chunks(args[0].data, new_buffer=True))
        ...
        
        # Store everything.
        ctx.save_for_backward(*args)

        return outputs

反向会在model_group中先通过gather_split_1d_tensor函数进行all_gather操作, 再进行backward的计算。

class CheckpointFunction(torch.autograd.Function):
    def backward(ctx, *args):
        if not torch.autograd._is_checkpoint_valid():
            raise RuntimeError("Checkpointing is not compatible with .grad(), "
                               "please use .backward() if possible")
        inputs = ctx.saved_tensors
        if ctx.distribute_saved_activations:
            safely_set_viewless_tensor_data(
                inputs[0],
                gather_split_1d_tensor(inputs[0].data).view(ctx.input_0_shape))
        ...
        # Compute the forward pass.
        detached_inputs = detach_variable(inputs)
        with torch.enable_grad():
            outputs = ctx.run_function(*detached_inputs)
        ...
        torch.autograd.backward(outputs, args)
        grads = tuple(inp.grad if isinstance(inp, torch.Tensor) else inp
                      for inp in detached_inputs)
        return (None, None) + grads

3. 配置选择

3.0 对比说明

随着模型大小的增加，sequence parallel和recompute都会节省内存，将内存需求减少约5倍。

条形图表示每层的前向、反向和重计算时间细分。基线代表没有重计算和序列并行时的情况。这些技术有效地减少了所有激活被重计算而不是保存时产生的开销。最大模型的开销从36%下降到仅为2%。

3.1 selective方式

只设--recompute-activations相当于只用了selective, 只对core_attention的input activation进行缓存。selective优点是新增的计算量相比较其他方案较少，性价比高。

--recompute-activations

3.2 block方式

使用block策略，前N层每一层都保存对应的input activation。可在pipeline并行中配合使用。--distribute-saved-activations可看情况使用。N等于pipeline stage中的layer数，可以最大限度使用重计算，节省显存。

--recompute-granularity full \
--recompute-method block \
--recompute-num-layers [N] \
--distribute-saved-activations \

3.3 uniform方式

使用uniform策略，每N个layer进行一次分组，每组会缓存输入input activation用于后续重计算，默认N为1表示对所有层都会，N等于1可以最大限度使用重计算，节省显存。在使用uniform基础上增加使用--distribute-saved-activations, 存activation时可以一个TP组进行分开存储。

--recompute-granularity full \
--recompute-method uniform \
--recompute-num-layers 1 \
--distribute-saved-activations \

4. 参考

• NVIDIA AI 平台大幅提高大型语言模型的性能