# 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。 ```shell 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 # Megatron-LM在以下commit hash下进行测试 # pip install git+https://github.com/NVIDIA/Megatron-LM.git@c6449f0b23be397449f21c0967c5fc90785e55ea ``` ## 精度对齐 - 为了支持精度对齐测试(FP32),你需注释掉[这几行](https://github.com/NVIDIA/Megatron-LM/blob/bd381ac364b5139840f0cba6389db54f2c092e90/megatron/core/transformer/experimental_attention_variant/dsa.py#L41-L43)。 修改完代码后,测试以下代码,确认无精度对齐问题(测试transformers/megatron forward对齐情况): 创建mini版本的模型,我们将创建4层: ```python 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](https://swift.readthedocs.io/zh-cn/latest/Megatron-SWIFT/Custom-Model.html)。 ```python 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](../../resources/deepseek_v4/precision.png) ## 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。 ```shell 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](../../resources/deepseek_v4/memory.png) 训练日志与损失: ![loss](../../resources/deepseek_v4/loss.png) 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 \ ``` - Packing/CP的支持:需安装mcore-bridge/ms-swift main分支。参考这两个PR:[ms-swift#9705](https://github.com/modelscope/ms-swift/pull/9705)、[mcore-bridge#140](https://github.com/modelscope/mcore-bridge/pull/140)。若要使用CP,你需要额外设置(需结合packing一起使用`--packing true`,并注意这个PR的合并[megatron-core#5706](https://github.com/NVIDIA/Megatron-LM/pull/5706)): ``` --sequence_packing_scheduler dp_balanced \ --cp_partition_mode contiguous \ ``` - 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](https://github.com/modelscope/ms-swift/blob/main/examples/megatron/fp8/lora.sh)。 ``` --fp8_recipe blockwise \ --fp8_format e4m3 \ --fp8_param_gather true \ ``` 推理训练后的模型: ```shell 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](../../resources/deepseek_v4/infer_result.png) 跑通vLLM推理: - 如果要使用vllm推理,你可以参考[这里的文档](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4-Flash)。你需要FP4/FP8精度的权重。 - 此外你需要copy原始的'config.json'文件,并修改'expert_dtype'(与训练后的config.json一致)。因为,使用transformers的`config.save_pretrained`保存的文件与原始文件不同,vllm不兼容保存后的文件。 - 如果遇到tilelang问题,可以查看[这个issue](https://github.com/modelscope/ms-swift/issues/9494)。 - mcore-bridge DeepSeek-V4 Fp8修复:[PR](https://github.com/modelscope/mcore-bridge/pull/133)。 这里先做量化(这里的量化会导致LoRA增量信息丢失,这里只作为例子,建议使用FP8全参数训练并导出FP8权重): ```shell CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \ NPROC_PER_NODE=8 \ megatron export \ --model megatron_output/DeepSeek-V4-Flash/vx-xxx/checkpoint-xxx-merged \ --output_dir megatron_output/DeepSeek-V4-Flash/vx-xxx/checkpoint-xxx-merged-FP8 \ --to_hf true \ --fp8_recipe blockwise \ --fp8_format e4m3 \ --fp8_param_gather true \ --mtp_num_layers 1 \ --expert_model_parallel_size 8 ``` vLLM启动命令: ```shell vllm serve megatron_output/DeepSeek-V4-Flash/vx-xxx/checkpoint-xxx-merged-FP8 \ --trust-remote-code \ --kv-cache-dtype fp8 \ --block-size 256 \ --enable-expert-parallel \ --tensor-parallel-size 8 \ --max-model-len 8192 \ --tokenizer-mode deepseek_v4 \ --tool-call-parser deepseek_v4 \ --enable-auto-tool-choice \ --reasoning-parser deepseek_v4 ```