qwen3VL GRPO训练模型保存时优化器状态不一致

[jiayouxiaoding]

System Info

最新的VERL megatron0.13.1

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

(WorkerDict pid=12272) WARNING:2025-10-30 21:54:46,671:There is difference in the common state dict in different ranks. The differences are {4: ([('optimizer', 0, 'param_state', 170), ('optimizer', 0, 'param_state', 171), ('optimizer', 0, 'param_state', 172), ('optimizer', 0, 'param_state', 173), ('optimizer', 0, 'param_state', 174), ('optimizer', 0, 'param_state', 175), ('optimizer', 0, 'param_state', 176), ('optimizer', 0, 'param_state', 177), ('optimizer', 0, 'param_state', 178), ('optimizer', 0, 'param_state', 179), ('optimizer', 0, 'param_state', 180), ('optimizer', 0, 'param_state', 181), ('optimizer', 0, 'param_state', 182), ('optimizer', 0, 'param_state', 183), ('optimizer', 0, 'param_state', 184), ('optimizer', 0, 'param_state', 185), ('optimizer', 0, 'param_state', 186), ('optimizer', 0, 'param_state', 187), ('optimizer', 0, 'param_state', 188), ('optimizer', 0, 'param_state', 189), ('optimizer', 0, 'param_state', 190), ('optimizer', 0, 'param_state', 191), ('optimizer', 0, 'param_state', 192), ('optimizer', 0, 'param_state', 193), ('optimizer', 0, 'param_state', 194), ('optimizer', 0, 'param_state', 195), ('optimizer', 0, 'param_state', 196), ('optimizer', 0, 'param_state', 197), ('optimizer', 0, 'param_state', 198),

Expected behavior

保证可以成功的保存训练权重，并且保证保存的权重对结果没有影响

[ieellee]

fsdp有问题吗

[jiayouxiaoding]

fsdp有问题吗
之前qwen2.5vl时候用的fsdp没问题，这次qwen3vl用的megatron，只要保存模型时候就会有这个warning