Running Inference on Multiple GPU #132
Description
Im trying to run the Deepseek-vl2 on 2 A6000 GPUs (48GB). When I modify the model loading from,
vl_gpt: DeepseekVLV2ForCausalLM = AutoModelForCausalLM.from_pretrained(
model_path,
trust_remote_code=True,
torch_dtype=dtype,
)
to,
vl_gpt: DeepseekVLV2ForCausalLM = AutoModelForCausalLM.from_pretrained(
model_path,
trust_remote_code=True,
torch_dtype=dtype,
device_map="auto"
)
Its throwing error that tensors should be in the same device;
File "DeepSeek-VL2/deepseek_vl2/models/modeling_deepseek.py", line 108, in forward
return self.weight * hidden_states.to(input_dtype)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1!
Please let me know if there is a way to run the multigpu inference. I have already tried the method split_model function in an earlier issue, but still didnt work.