Does the vllm 0.6 is ok?

[catsled]

for some reason, i have to use the vllm==0.6 to train grpo, but it will meet the

after that

[rank0]:   File "/usr/local/lib/python3.10/site-packages/vllm/model_executor/models/qwen2.py", line 289, in forward
[rank0]:     hidden_states = self.embed_tokens(input_ids)
[rank0]:   File "/usr/local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:   File "/usr/local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:   File "/usr/local/lib/python3.10/site-packages/vllm/model_executor/layers/vocab_parallel_embedding.py", line 413, in forward
[rank0]:     output_parallel = self.linear_method.embedding(self,
[rank0]:   File "/usr/local/lib/python3.10/site-packages/vllm/model_executor/layers/vocab_parallel_embedding.py", line 57, in embedding
[rank0]:     return F.embedding(input_, layer.weight)
[rank0]:   File "/usr/local/lib/python3.10/site-packages/torch/nn/functional.py", line 2292, in embedding
[rank0]:     return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
[rank0]: RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:7 and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

[vagitablebirdcode]

I have the same trouble，also if use vllm < 0.6.5, the GRPOTrainer will cause an error that lack of vllm.worker.worker.Worker._assert_memory_footprint_increased_during_profiling
when use vllm >= 0.6.5 and vllm < 0.7, there cause a conflict of multi-device use of vllm and other utils.

[XZ-X]

I encountered the same issue. Using vllm >= 0.7 solves the problem for me.