Creation Process: SFT
SFT on approx 13 million tokens, SFW / NSFW RP, stories, creative instruct & chat data. Some of the SFW datasets are public and can be found in the model datasets list.
I've switched over from Axolotl to MS-Swift w/ Megatron to train MoE models now. There's a roughly 5-10x speedup in training the models, thanks to escaping the naive MoE implementation in TRL. The training time for this run took only 40 minutes, excluding environment setup time.
A low LR for GLM Air appears to be king. Going any higher, I've found it extremely easy to begin overcooking the model.
>
MS-Swift config
Not optimized for cost / performance efficiency, YMMV.
SFT (8*H200)
PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
NPROC_PER_NODE=8 \
WANDB_API_KEY=wandb_key \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
megatron sft \
--load '/workspace/glm-4.5-air-mcore' \
--dataset '/workspace/joined_dataset_cleaned_modified.jsonl' \
--load_from_cache_file true \
--train_type lora \
--lora_rank 256 \
--lora_alpha 16 \
--use-rslora true \
--target_modules all-linear \
--split_dataset_ratio 0.01 \
--moe_permute_fusion true \
--tensor_model_parallel_size 8 \
--expert_tensor_parallel_size 1 \
--expert_model_parallel_size 8 \
--moe_grouped_gemm true \
--moe_shared_expert_overlap true \
--moe_aux_loss_coeff 6e-5 \
--micro_batch_size 4 \
--global_batch_size 32 \
--recompute_granularity full \
--recompute_method uniform \
--recompute_num_layers 1 \
--max_epochs 2 \
--cross_entropy_loss_fusion true \
--lr 6e-6 \
--lr_warmup_fraction 0.05 \
--min_lr 6e-7 \
--save megatron_output/Iceblink-v3-SFT-3 \
--eval_interval 20 \
--save_interval 25 \
--finetune true \
--packing true \
--max_length 10280 \
--num_workers 8 \
--dataset_num_proc 8 \
--no_save_optim true \
--no_save_rng true \
--sequence_parallel true \
--wandb_project Megatron-Air-SFT \
--wandb_exp_name Iceblink-v3-SFT-3 \
--attention_backend flash
