swift-megatron qwen3-235b-a22b

examples/train/megatron/qwen3_235b.sh

@@ -0,0 +1,40 @@
+# ppue 32 * 16 = 512 gpus
+MASTER_ADDR=$MASTER_ADDR \
+MASTER_PORT=$MASTER_PORT \
+NNODES=$NNODES \
+NODE_RANK=$NODE_RANK \
+NPROC_PER_NODE=$NPROC_PER_NODE \        
+megatron sft \
+    --load /mnt/sllmworkspace-p/common/sllmworks/model/Qwen3-235B-A22B-mcore \
+    --dataset /sllmworks/datasets/Chinese-DeepSeek-R1-Distill-data-110k-SFT \
+    --tensor_model_parallel_size 4 \
+    --expert_model_parallel_size 8 \
+    --pipeline_model_parallel_size 2 \
+    --moe_shared_expert_overlap true \
+    --moe_aux_loss_coeff 0.01 \
+    --micro_batch_size 2 \
+    --global_batch_size 128 \
+    --packing true \
+    --recompute_granularity full \
+    --recompute_method uniform \
+    --recompute_num_layers 1 \
+    --train_iters 20 \
+    --split_dataset_ratio 0 \
+    --finetune true \
+    --cross_entropy_loss_fusion true \
+    --lr 1e-05 \
+    --lr_warmup_iters 2 \
+    --min_lr 1e-06 \
+    --save /mnt/sllmworkspace-p/common/sllmworks/2641825084/checkpoint \
+    --save_interval 200 \
+    --max_length 8192 \
+    --num_workers 8 \
+    --dataset_num_proc 8 \
+    --no_save_optim true \
+    --no_save_rng true \
+    --sequence_parallel true \
+    --use_flash_attn false \
+    --attention_backend auto \
+    --tensorboard_dir /home/admin/logs/tfevent \
+    --max_epochs 1 \
+    --ckpt_format torch_dist