[Torch 2.10 Megatron GRPO] The AccumulateGrad node's stream does not match the stream of the node that produced the incoming gradient.

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

Hi, 用最新的镜像训练GRPO （Qwen3.5 35b-a3b）得到了这个warning：

镜像：modelscope-registry.cn-hangzhou.cr.aliyuncs.com/modelscope-repo/modelscope:ubuntu22.04-cuda12.8.1-py311-torch2.10.0-vllm0.17.0-modelscope1.34.0-swift4.0.1

The AccumulateGrad node's stream does not match the stream of the node that produced the incoming gradient.

请问会不会影响训练精度呢 （目前看好像是torch 2.10.0 + cuda 12.8会有的问题）参考报错

How to Reproduce / 如何复现

training_options=" \
    --model_type qwen3_5_moe \
    --freeze_llm false \
    --freeze_vit true \
    --freeze_aligner true \
    --add_non_thinking_prefix true \
    --loss_scale ignore_empty_think \
    --decoder_first_pipeline_num_layers 24 \
    --steps_per_generation 8 \
    --micro_batch_size 1 \
    --global_batch_size 64 \
    --num_generations 8"
    # --mtp_num_layers ${NUM_MTP_LAYER} \
    # --mtp_loss_scaling_factor 0.1 \
# VLLM options
vllm_options=" \
    --use_vllm true \
    --vllm_mode colocate \
    --vllm_tensor_parallel_size 16 \
    --vllm_gpu_memory_utilization 0.5 \
    --vllm_max_model_len 20480"
# Common training options
common_training_options=" \
    --rlhf_type grpo \
    --loss_type sapo \
    --max_length 16384 \
    --max_completion_length 4086 \
    --lr 1e-5 \
    --lr_warmup_fraction 0.05 \
    --min_lr 1e-6 \
    --train_type full \
    --reward_funcs format \
    --tau_pos 1 \
    --tau_neg 1.05 \
    --epsilon 0.2 \
    --epsilon_high 0.2 \
    --beta 0.001 \
    --finetune true \
    --packing false \
    --padding_free true \
    --dynamic_sample false \
    --num_train_epochs 2 \
    --overlong_filter false \
    --importance_sampling_level token"
    # > 注意：如果开启`overlong_filter`, kl 和 clip_ratio 指标会过滤超长的样本
    # --external_plugins ${CUSTOM_WORK_DIR}/scripts/base/grpo/latest_plugin.py \

# Checkpoint options
checkpoint_options=" \
    --recompute_granularity full \
    --recompute_method uniform \
    --recompute_num_layers 1 \
    --attention_backend flash \
    --save_strategy epoch \
    --save_steps 500 \
    --logging_steps 1 \
    --log_completions true \
    --dataloader_num_workers 32 \
    --output_dir ${OUTPUT_DIR}/${OUT_NAME} \
    --no_save_optim \
    --no_save_rng"
# Offload options
offload_options=" \
    --offload_bridge true \
    --sleep_level 2 \
    --offload_model true \
    --offload_optimizer true \
    --optimizer_cpu_offload true"

Additional Information / 补充信息

No response

The report names no ms-swift files or tests. Start by reproducing the GRPO configuration with the listed Torch 2.10.0, CUDA 12.8.1, vLLM 0.17.0 and ms-swift 1.34.0 image, then compare the warning with the linked PyTorch Lightning report. Done means determining whether training precision is affected and documenting whether the issue belongs in ms-swift or an upstream dependency.