断点续训参数设置问题

Checklist / 检查清单

• I have searched existing issues, and this is a new question or discussion topic. / 我已经搜索过现有的 issues，确认这是一个新的问题与讨论。

Question Description / 问题描述

请问megatron pt的断点续训，第一次训练和断点续训的参数分别应该如何设置？

第一次训练能够正常训练，参数如下：

PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
NNODES=2 \
NODE_RANK=0 \
MASTER_ADDR=xxxx \
MASTER_PORT=29500 \
NPROC_PER_NODE=8 \
megatron pt \
    --model /data/disk1/Qwen/Qwen3-30B-A3B-Base \
    --no_save_optim true \
    --no_save_rng true \
    --dataset /data/disk1/dataset3.jsonl \
    --load_from_cache_file true \
    --tensor_model_parallel_size 1 \
    --expert_model_parallel_size 8 \
    --pipeline_model_parallel_size 1 \
    --moe_expert_capacity_factor 2 \
    --micro_batch_size 1 \
    --global_batch_size 128 \
    --recompute_granularity full \
    --recompute_num_layers 2 \
    --recompute_method uniform \
    --use_distributed_optimizer true \
    --use_flash_attn true \
    --cross_entropy_loss_fusion true \
    --lr 2e-5 \
    --finetune true \
    --lr_warmup_fraction 0.05 \
    --min_lr 2e-6 \
    --max_epochs 1 \
    --save /data/disk1/Qwen3-30B-A3B-Base-save \
    --save_interval 10 \
    --max_length 32768 \
    --num_workers 16 \
    --dataset_num_proc 2 \
    --model_author swift \
    --model_name swift-robot \
    --packing true \
    --overlap_grad_reduce true \
    --overlap_param_gather true \
    --moe_grouped_gemm true \
    --moe_permute_fusion true \
    --moe_shared_expert_overlap true \
    --log_throughput true

断点续训失败，参数如下：

PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
NNODES=2 \
NODE_RANK=1 \
MASTER_ADDR=xxxx \
MASTER_PORT=29500 \
NPROC_PER_NODE=8 \
megatron pt \
    --load /data/disk1/Qwen3-30B-A3B-Base-save/v0-20260205-202217/checkpoint-10  \ #这里做了更改
    --no_save_optim true \
    --no_save_rng true \
    --dataset /data/disk1/dataset3.jsonl \
    --load_from_cache_file true \
    --tensor_model_parallel_size 1 \
    --expert_model_parallel_size 8 \
    --pipeline_model_parallel_size 1 \
    --moe_expert_capacity_factor 2 \
    --micro_batch_size 1 \
    --global_batch_size 128 \
    --recompute_granularity full \
    --recompute_num_layers 2 \
    --recompute_method uniform \
    --use_distributed_optimizer true \
    --use_flash_attn true \
    --cross_entropy_loss_fusion true \
    --lr 2e-5 \
    --finetune false \                        #这里做了更改
    --lr_warmup_fraction 0.05 \
    --min_lr 2e-6 \
    --max_epochs 1 \
    --save /data/disk1/Qwen3-30B-A3B-Base-save \
    --save_interval 10 \
    --max_length 32768 \
    --num_workers 16 \
    --dataset_num_proc 2 \
    --model_author swift \
    --model_name swift-robot \
    --packing true \
    --overlap_grad_reduce true \
    --overlap_param_gather true \
    --moe_grouped_gemm true \
    --moe_permute_fusion true \
    --moe_shared_expert_overlap true \
    --log_throughput true

Start with the megatron pt entry point and compare the working command with the resume command, especially --load and --finetune. The issue does not include the failure output, so first reproduce the resume attempt using checkpoint-10; done means training resumes successfully from that checkpoint.