Qwen3-VL-30B-A3B-Instruct官方最佳实践OOM

Question Description / 问题描述

环境：8 x A100 80G
ms-swift：4.0

使用官方文档当中的最佳实践进行测试

# 8 * 80GiB
PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
OMP_NUM_THREADS=14 \
NPROC_PER_NODE=8 \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
IMAGE_MAX_TOKEN_NUM=1024 \
VIDEO_MAX_TOKEN_NUM=128 \
FPS_MAX_FRAMES=16 \
megatron sft \
    --model Qwen/Qwen3-VL-30B-A3B-Instruct \
    --save_safetensors true \
    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#10000' \
              'AI-ModelScope/LaTeX_OCR:human_handwrite#5000' \
              'swift/VideoChatGPT:Generic#2000' \
    --load_from_cache_file true \
    --split_dataset_ratio 0.01 \
    --moe_permute_fusion true \
    --tensor_model_parallel_size 4 \
    --expert_model_parallel_size 8 \
    --moe_grouped_gemm true \
    --moe_shared_expert_overlap true \
    --moe_aux_loss_coeff 1e-6 \
    --micro_batch_size 1 \
    --global_batch_size 4 \
    --recompute_granularity full \
    --recompute_method uniform \
    --recompute_num_layers 1 \
    --num_train_epochs 1 \
    --finetune true \
    --cross_entropy_loss_fusion true \
    --lr 1e-5 \
    --lr_warmup_fraction 0.05 \
    --min_lr 1e-6 \
    --output_dir megatron_output/Qwen3-VL-30B-A3B-Instruct \
    --eval_steps 500 \
    --save_steps 500 \
    --max_length 4096 \
    --packing true \
    --dataloader_num_workers 8 \
    --dataset_num_proc 8 \
    --no_save_optim true \
    --no_save_rng true \
    --sequence_parallel true \
    --moe_expert_capacity_factor 2 \
    --attention_backend flash

出现OOM

[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 27.00 GiB. GPU 0 has a total capacity of 79.25 GiB of which 20.24 GiB is free. Process 2654119 has 59.01 GiB memory in use. Of the allocated memory 57.48 GiB is allocated by PyTorch, and 464.88 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

为什么我本地无法顺利训练这个最佳实践呢？

No source file, test, or entry point is named. Start by reproducing the documented megatron sft command on the stated 8 x A100 80G setup and inspect the reported allocation; determine whether the configuration or the official guidance causes the OOM. Done means documenting the root cause and a verified configuration or correction.