Megatron SFT Qwen3 Omni is Slow

[mertunsall]

The speed seems to be extremely slow on 8xH100 with config (only difference is long context training)

PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
NPROC_PER_NODE=8 \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
megatron sft \
    --load 'Qwen3-Omni-30B-A3B-Instruct-mcore' \
    --dataset 'dataset_id' \
    --train_type lora \
    --lora_rank 8 \
    --lora_alpha 32 \
    --target_modules all-linear \
    --tensor_model_parallel_size 2 \
    --expert_model_parallel_size 2 \
    --context_parallel_size 2 \
    --moe_grouped_gemm true \
    --moe_shared_expert_overlap true \
    --moe_aux_loss_coeff 1e-3 \
    --micro_batch_size 1 \
    --global_batch_size 32 \
    --packing true \
    --freeze_llm false \
    --freeze_vit true \
    --freeze_aligner true \
    --split_dataset_ratio 0.01 \
    --recompute_granularity full \
    --recompute_method uniform \
    --recompute_num_layers 1 \
    --finetune true \
    --cross_entropy_loss_fusion true \
    --lr 4e-5 \
    --max_epochs 1 \
    --lr_warmup_fraction 0.05 \
    --min_lr 4e-6 \
    --save 'Qwen3-Omni-30B-A3B-Instruct-280925-lora' \
    --eval_interval 200 \
    --save_interval 200 \
    --max_length 32768 \
    --num_workers 100 \
    --dataset_num_proc 100 \
    --no_save_optim true \
    --no_save_rng true \
    --sequence_parallel true \
    --use_flash_attn true \
    --use_hf true

I get ~15 minutes for 5 iterations.

[Jintao-Huang]

Normal, simply caused by max_length & packing.

Increase DP number, reduce recomputation, and accelerate via communication overlap.

[mertunsall]

How would you suggest reducing recomputation and accelerating via comm overlap?

also, what would be your recommended setting for full finetuning of the 30B model? should 16xH100 in 2 nodes suffice?

[YouTwoMeToo]

Hello, could you tell me if you used the following script to convert your weight files from HF format to Megatron format? I used this script for conversion, but when I try to fine-tune using the moe.sh script, I keep getting an error message that the current parallel/model configuration is inconsistent with the sharding method recorded in the checkpoint.
swift export \ --model Qwen3-Omni-Instruct \ --to_mcore true \ --torch_dtype bfloat16 \ --output_dir Qwen3-Omni-Instruct-mcore