DeepSeek-V4 训练支持

目前Megatron-SWIFT支持了DeepSeek-V4的微调与RL支持,包括MTP、FP8等特性。(FP4 blockwise训练暂时不支持,会在加载权重时自动转成FP8/BF16)

你需要使用Megatron-Core dev分支以及mcore-bridge、ms-swift main分支。

pip install git+https://github.com/NVIDIA/Megatron-LM.git@dev
pip install git+https://github.com/modelscope/mcore-bridge.git
pip install git+https://github.com/modelscope/ms-swift.git

精度对齐

目前Megatron-Core对DeepSeek-V4的支持算子实现有误,存在精度误差(后续可能更新),具体查看这个issue。你需要修改部分代码:

  • 修改这行,修改为mixed = torch.bmm(h_res_batched.transpose(-1, -2), residual_batched).view(s, b, n, C)

  • 修改这行,修改为h_res_batched = h_res.transpose(-1, -2).contiguous().view(s * b, n, n)

  • 此外为了支持精度对齐测试(FP32),你需注释掉这几行

修改完代码后,测试以下代码,确认无实现错误(测试transformers/megatron forward对齐情况):

创建mini版本的模型,我们将创建4层:

import os
import torch
from modelscope.hub.file_download import model_file_download
from safetensors.torch import safe_open
from swift import safe_snapshot_download

from mcore_bridge.utils import Fp8Dequantizer, SafetensorLazyLoader, StreamingSafetensorSaver

model_id = 'deepseek-ai/DeepSeek-V4-Flash-Base'
# Some models have the first few layers as dense and the rest as MoE; set this value accordingly
model_dir = safe_snapshot_download(model_id, download_model=False)

loader = SafetensorLazyLoader(model_dir)
state_dict = loader.get_state_dict()
saver = StreamingSafetensorSaver(save_dir=model_dir)
fp8_dequantizer = Fp8Dequantizer()  # Used to convert fp8 weights to bf16


def _open_file(self, filename: str):
    if filename not in self._file_handles:
        file_path = os.path.join(self.hf_model_dir, filename)
        tmp_dir = os.path.join(self.hf_model_dir, 'tmp')
        if not os.path.exists(file_path):
            file_path = os.path.join(tmp_dir, filename)
        if not os.path.exists(file_path):
            file_path = model_file_download(
                model_id=model_id,
                file_path=filename,
                local_dir=tmp_dir,
            )
        self._file_handles[filename] = safe_open(file_path, framework='pt')
    return self._file_handles[filename]


SafetensorLazyLoader._open_file = _open_file  # monkey patch (lazy downloading)
new_state_dict = {}
for k, v in state_dict.items():
    if k.startswith('layers.'):
        idx = int(k[len('layers.'):].split('.', 1)[0])
        if idx >= 4:
            continue
    if k.endswith('.scale'):
        continue
    elif k.endswith('.weight'):
        weight_scale_inv = k.replace('.weight', '.scale')
        if weight_scale_inv in state_dict:
            v = fp8_dequantizer.convert(v.load(), state_dict[weight_scale_inv].load()).to(torch.bfloat16)
    new_state_dict[k] = v if isinstance(v, torch.Tensor) else v.load()

for k, v in new_state_dict.items():
    saver.add_tensor(k, v)
saver.finalize()

然后修改config.json

  • num_hidden_layers修改为4

  • compress_ratios修改为[0, 0, 4, 128, 0]

  • 删除quantization_config

然后创建test.py,使用以下命令运行:CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 test.py。更多参考自定义Megatron模型文档

import os

os.environ['SWIFT_TEST_CONVERT_PRECISION'] = '1'

from swift.megatron import MegatronExportArguments, megatron_export_main
from swift import safe_snapshot_download
model_id = 'deepseek-ai/DeepSeek-V4-Flash-Base'

model_dir = safe_snapshot_download(model_id, download_model=False)

if __name__ == '__main__':
    megatron_export_main(
        MegatronExportArguments(
            model=model_dir,
            to_mcore=True,
            attention_backend='flash',
            tensor_model_parallel_size=1,
            pipeline_model_parallel_layout='Et*3|t*1mL',
            pipeline_model_parallel_size=2,
            expert_model_parallel_size=2,
            mtp_num_layers=1,
            test_convert_precision=True,
        ))

当出现以下结果时,则表示对齐没有问题,可以进行训练了。 精度对齐

LoRA训练

BF16精度LoRA训练脚本如下,最后会保存LoRA增量权重和Merge-LoRA后的BF16完整权重。

PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
NPROC_PER_NODE=8 \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
megatron sft \
    --model deepseek-ai/DeepSeek-V4-Flash \
    --save_safetensors true \
    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#1000' \
              'AI-ModelScope/alpaca-gpt4-data-en#1000' \
              'swift/self-cognition#1000' \
    --model_author swift \
    --model_name swift-robot \
    --merge_lora true \
    --load_from_cache_file true \
    --add_non_thinking_prefix true \
    --loss_scale ignore_empty_think \
    --split_dataset_ratio 0.01 \
    --tuner_type lora \
    --lora_rank 16 \
    --lora_alpha 32 \
    --tensor_model_parallel_size 1 \
    --expert_model_parallel_size 8 \
    --micro_batch_size 4 \
    --global_batch_size 32 \
    --padding_free false \
    --group_by_length true \
    --recompute_granularity full \
    --recompute_method uniform \
    --recompute_num_layers 1 \
    --moe_permute_fusion true \
    --moe_grouped_gemm true \
    --moe_shared_expert_overlap true \
    --moe_aux_loss_coeff 1e-3 \
    --num_train_epochs 1 \
    --finetune true \
    --cross_entropy_loss_fusion true \
    --lr 1e-4 \
    --lr_warmup_fraction 0.05 \
    --min_lr 1e-5 \
    --output_dir megatron_output/DeepSeek-V4-Flash \
    --eval_steps 200 \
    --save_steps 200 \
    --max_length 4096 \
    --dataloader_num_workers 8 \
    --dataset_num_proc 8 \
    --no_save_optim true \
    --no_save_rng true \
    --sequence_parallel true \
    --mtp_num_layers 1 \
    --attention_backend flash

显存占用: 显存占用

训练日志与损失: loss

提示:

  • 如果你要设置pp并行,你需要额外设置pipeline_model_parallel_layout。例如:

--pipeline_model_parallel_size 2 \
--pipeline_model_parallel_layout 'Et*22|t*21mL' \
  • 全参数训练也是支持的,你需要降低learning_rate,并提高并行数。参考64卡训练例子:

--lr 1e-5 \
--min_lr 1e-6 \
--tensor_model_parallel_size 1 \
--expert_model_parallel_size 8 \
--pipeline_model_parallel_size 8 \
--pipeline_model_parallel_layout Et*5|t*5|t*6|t*6|t*6|t*5|t*5|t*5mL \
  • 暂时不支持padding_freepacking,但可以通过group_by_length加速。暂时不支持TP,待Megatron-Core支持。

  • FP8训练:你可以设置以下参数开启FP8训练,并最终将权重保存成FP8权重。推荐使用全参数训练。如果要使用LoRA + FP8,你需要只保存LoRA权重(设置--merge_lora false),并使用BF16权重进行Merge-LoRA(FP8 精度有限,LoRA delta 会被舍入为 0)。参考这个例子

--fp8_recipe blockwise \
--fp8_format e4m3 \
--fp8_param_gather true \

推理训练后的模型:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
swift infer \
    --model megatron_output/DeepSeek-V4-Flash/vx-xxx/checkpoint-xxx-merged \
    --infer_backend transformers \
    --enable_thinking false \
    --max_new_tokens 2048

推理结果:

result