[BUG] [MLA] k_pos_emb in MLA is not reduced across TP ranks if linear_kv_down_proj is ColumnParallelLinear
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 4.5k
- Avg merge
- 4d 6h
- Merged PRs (30d)
- 271
Description
**Describe the bug**
https://github.com/NVIDIA/Megatron-LM/blob/e96a358f60c82b8ac8d965d91c3cc4ad0230a4e0/megatron/core/transformer/multi_latent_attention.py#L443
`gather_from_tensor_model_parallel_region` uses `split` for backward. The grad of `kv_compressed` is already reduced by `ColumnParallelLinear` but the grad of `k_pos_emb` is never reduced across TP ranks.
This part of code was introduced by https://github.com/NVIDIA/Megatron-LM/commit/1c425a83136c4f5b5841051fe9a501e3c36a750b
**Analysis**
The `expand` operation [here](https://github.com/NVIDIA/Megatron-LM/blob/e96a358f60c82b8ac8d965d91c3cc4ad0230a4e0/megatron/core/transformer/multi_latent_attention.py#L573-L577) is actually a simplified forward path for:
**k_pos_emb** *(nheads=1, duplicated)* --expand--> **k_pos_emb** *(nheads=N, duplicated)* --scatter--> **k_pos_emb** *(nheads=N/TP, sharded)*
During the backward, the grad of `k_pos_emb` is different across TP ranks and a sum-reduce might be needed.
Contributor guide
Assessment
This issue has not been assessed yet.