胡译文

icon indicating an email [email protected]

icon company Renmin University of China icon location Beijing icon location Undergraduate majored in AI and Fintech. Developed a highly capable 2.4B LLM. Participated in the Olympiad in Informatics. WeChat: yiwen_hu

Results 13 comments of 胡译文

installing pointnet2_ops_lib

you should install via http protocol. just simply replace "git+git://" with "git+http://"

[BUG] No `universal_checkpoint_info` in the Accelerate+Deepspeed Checkpoint

Hello @xylian86 , I was previously using the HF trainer. Why doesn't the universal checkpoint support the HF trainer? Is there any way to load the universal checkpoint? Do I...

Support dashscope API for Qwen models

Hi @krrishdholakia , I found this PR very useful and would love to see it merged. Are there any updates or changes needed? Thanks for your hard work!

推理速度如何？

请问和同等大小的Transformers架构模型，推理速度差别是多少呢

ZeNotes does not work with group library

Thank you for replying! I moved the papers to another group library and everything worked all fine.

Failed to load universal_checkpoint with deepspeed integreation

Here's my deepspeed config json: ```json { "bf16": { "enabled": "auto" }, "zero_optimization": { "stage": 2, "allgather_partitions": true, "allgather_bucket_size": 1e8, "overlap_comm": true, "reduce_scatter": true, "reduce_bucket_size": 1e8, "contiguous_gradients": true }, "gradient_accumulation_steps":...

Failed to load universal_checkpoint with deepspeed integreation

Another related issue: https://github.com/microsoft/DeepSpeed/issues/5405

Failed to load universal_checkpoint with deepspeed integreation

Hello @ArthurZucker and @muellerz. I am able to create a pull request to address the issue. I have resolved the issue by deleting all the “rng_state” files as it had...

Failed to load universal_checkpoint with deepspeed integreation

We can skip these rng_state and add a warning.

Mitigation to HuggingFace Trainer

I’m currently working on an MoE model and looking to implement expert parallelism. Writing EP/EP+TP/EP+DP from scratch with torch distributed communication is pretty challenging, especially if I want good training...