[QUESTION] why llava model context parallel split sequence use tp group?
Open
module: multimodal
question
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 4.5k
- Avg merge
- 4d 6h
- Merged PRs (30d)
- 271
Description
https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/models/multimodal/llava_model.py#L746
if tp=1 and cp=2,this code fails to split sequence as expected,
Contributor guide
Assessment
This issue has not been assessed yet.