使用mindspore的ckpt_to_safetensors api 转化权重,
mindspore 2.4.0的daily dev版本,python3.10的aarch64的linux安装包:
https://repo.mindspore.cn/mindspore/mindspore/version/202410/20241021/master_20241021010025_07b984a9306bacc69d92287758d39a87c98f34b1_newest/unified/aarch64/mindspore-2.4.0-cp310-cp310-linux_aarch64.whl
成功将微调后的ckpt权重转化为safetensor权重,mindie-service启动报错:
尝试1:微调并合并的ckpt完整权重,进行safetensor转化,报错信息:Fatal Python error: PyThreadState_Get: the function must be called with the GIL held, but the GIL is released (the current Python thread state is NULL) Python runtime state: finalizing (tstate=0x0000ffff8001e3f0)
Daemon wait pid with 2423, status 134Daemon wait pid with 2419, status 134Daemon wait pid with 2420, status 134Daemon is killing..
尝试2:微调的分布式ckpt权重,分别进行safetensor转化,报错信息:[ERROR] [pid: 3159] weights.py-41: Key lm_head.weight was found in multiple files: /mnt/models/Baichuan2-13B-Chat/lora_weights/checkpoint_7.safetensors and /mnt/models/Baichuan2-13B-Chat/lora_weights/checkpoint_2.safetensors
mindie环境已验证,未微调权重可以挂起service服务
问题1:mindie是否存在支持ckpt挂起service的接口?
问题2:mindspore 2.4.0的daily dev版本的ckpt_to_safetensors api,在转化权重上,是否仍存在问题?
问题3:是否有其他途径,可以对微调过的ckpt权重,在mindie上挂起service服务?
mindspore 2.4.0的daily dev版本,python3.10的aarch64的linux安装包:
https://repo.mindspore.cn/mindspore/mindspore/version/202410/20241021/master_20241021010025_07b984a9306bacc69d92287758d39a87c98f34b1_newest/unified/aarch64/mindspore-2.4.0-cp310-cp310-linux_aarch64.whl
成功将微调后的ckpt权重转化为safetensor权重,mindie-service启动报错:
尝试1:微调并合并的ckpt完整权重,进行safetensor转化,报错信息:Fatal Python error: PyThreadState_Get: the function must be called with the GIL held, but the GIL is released (the current Python thread state is NULL) Python runtime state: finalizing (tstate=0x0000ffff8001e3f0)
Daemon wait pid with 2423, status 134Daemon wait pid with 2419, status 134Daemon wait pid with 2420, status 134Daemon is killing..
尝试2:微调的分布式ckpt权重,分别进行safetensor转化,报错信息:[ERROR] [pid: 3159] weights.py-41: Key lm_head.weight was found in multiple files: /mnt/models/Baichuan2-13B-Chat/lora_weights/checkpoint_7.safetensors and /mnt/models/Baichuan2-13B-Chat/lora_weights/checkpoint_2.safetensors
mindie环境已验证,未微调权重可以挂起service服务
问题1:mindie是否存在支持ckpt挂起service的接口?
问题2:mindspore 2.4.0的daily dev版本的ckpt_to_safetensors api,在转化权重上,是否仍存在问题?
问题3:是否有其他途径,可以对微调过的ckpt权重,在mindie上挂起service服务?