SFT on Qwen3 30B-a3b is quite slow

### Could you provide numbers of the expected speed (s/iter) and check if it's possible improve it's implementation?

In our experience, performing SFT with Qwen3 30B-a3b with the the Megatron backend is quite slow.
Here are our performance numbers:
|Nodes|GPUs | Max Seq Length | Global Batch Size | Micro Batch Size | Expert parallelism | Tensor parallelism | s/iter|
|--|--|--|--|--|--|--|--|
| 4 | 8 | 4k | 512 | 8 | 8| 1| **~13** |
| 4 | 8 | 32k | 512| 1 | 8 | 2 | **~35** |

We're using H200s and for the experiment with max_seq_len=32k we're doing sequence packing.

**Environment details**

- Image: `nvcr.io/nvidia/pytorch:25.06-py3`