YutongWang

[Feature] Implement function of automatically algorithm performance tests for Sedna Edge-cloud Re-ID

1

**Why is this needed**: Person Re-ID based on Cloud-edge collaborative architecture is a new example feature that Sedna will soon support, which can continuously recognize, track and search the target...

INT8量化版本在设置加载显卡时运行报错:Triton Error [CUDA]: an illegal memory access was encountered

2

我在多卡机器上部署moss-moon-003-sft-int8模型，运行下面指令是： ``` model = AutoModelForCausalLM.from_pretrained("fnlp/moss-moon-003-sft-int8", trust_remote_code=True).half().cuda(1) inputs="blabla" for k in inputs: inputs[k] = inputs[k].cuda(1) ``` 在 `outputs = model.generate(**inputs, do_sample=True, temperature=0.7, top_p=0.8, repetition_penalty=1.02, max_new_tokens=256)` 处报错**Triton Error [CUDA]: an illegal memory...