[model] support Qwen3-235B-A22B-Instruct-250718

[Jintao-Huang]

Model Finetuning

https://github.com/modelscope/ms-swift/blob/main/examples/train/megatron/lora/qwen3_235b.sh

We introduce self-cognition finetuning of Qwen3-235B-A22B-Instruct-2507 using Megatron & LoRA integrated with ms-swift. You will need 8 GPUs with 80GiB memory each.

Before starting the finetuning process, please ensure your environment is properly set up.

For instructions on installing Megatron-related dependencies, please refer to the Megatron-SWIFT training documentation (Docker images are also available):
https://swift.readthedocs.io/en/latest/Instruction/Megatron-SWIFT-Training.html

git clone https://github.com/modelscope/ms-swift.git
cd ms-swift
pip install -e .

The finetuning dataset should be prepared in the following format (the "system" field is optional). You can specify it in the training script using --dataset <dataset_path>.

{"messages": [{"role": "user", "content": "Where is the capital of Zhejiang?"}, {"role": "assistant", "content": "The capital of Zhejiang is Hangzhou."}]}

1. Convert HF-format weights to Megatron format and test the conversion accuracy:

# 8 * 80GiB
# To test the conversion accuracy, please set --test_convert_precision true. Note that this requires 1.3T of memory resources.
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
swift export \
    --model Qwen/Qwen3-235B-A22B-Instruct-2507 \
    --to_mcore true \
    --torch_dtype bfloat16 \
    --output_dir Qwen3-235B-A22B-Instruct-2507-mcore

2. Self-cognition finetuning (full-parameter training) of Qwen3-235B-A22B-Instruct-2507-mcore requires 8 H20 GPUs with 90GiB memory each, and achieves a training speed of 3.5 seconds per iteration.

# Memory usage: 8 x 90GiB
# Training speed: 3.5s/it
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 Qwen3-235B-A22B-Instruct-2507-mcore \
    --dataset 'swift/Chinese-Qwen3-235B-2507-Distill-data-110k-SFT#2000' \
              'swift/self-cognition#1000' \
    --train_type lora \
    --lora_rank 8 \
    --lora_alpha 32 \
    --target_modules all-linear \
    --split_dataset_ratio 0.01 \
    --expert_model_parallel_size 8 \
    --moe_grouped_gemm true \
    --moe_shared_expert_overlap true \
    --moe_aux_loss_coeff 1e-3 \
    --micro_batch_size 2 \
    --global_batch_size 16 \
    --recompute_granularity full \
    --recompute_method uniform \
    --recompute_num_layers 1 \
    --max_epochs 1 \
    --finetune true \
    --cross_entropy_loss_fusion true \
    --lr 1e-4 \
    --lr_warmup_fraction 0.05 \
    --min_lr 1e-5 \
    --save megatron_output/Qwen3-235B-A22B-Instruct-2507 \
    --eval_interval 200 \
    --save_interval 200 \
    --max_length 2048 \
    --num_workers 8 \
    --dataset_num_proc 8 \
    --no_save_optim true \
    --no_save_rng true \
    --sequence_parallel true \
    --attention_backend flash \
    --model_author swift \
    --model_name swift-robot

Training memory usage:

Training log:

If you need to run it on 8 GPUs with 80GiB memory each, you can use the following configuration:

# Memory usage: 8 * 78GiB
# Training speed: 9.5s/it
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 Qwen3-235B-A22B-Instruct-2507-mcore \
    --dataset 'swift/Chinese-Qwen3-235B-2507-Distill-data-110k-SFT#2000' \
              'swift/self-cognition#1000' \
    --optimizer_cpu_offload true \
    --use_precision_aware_optimizer true \
    --train_type lora \
    --lora_rank 8 \
    --lora_alpha 32 \
    --target_modules all-linear \
    --split_dataset_ratio 0.01 \
    --expert_model_parallel_size 2 \
    --pipeline_model_parallel_size 4 \
    --decoder_first_pipeline_num_layers 23 \
    --decoder_last_pipeline_num_layers 23 \
    --moe_grouped_gemm true \
    --moe_shared_expert_overlap true \
    --moe_aux_loss_coeff 1e-3 \
    --micro_batch_size 8 \
    --global_batch_size 16 \
    --recompute_granularity full \
    --recompute_method uniform \
    --recompute_num_layers 1 \
    --max_epochs 1 \
    --finetune true \
    --cross_entropy_loss_fusion true \
    --lr 1e-4 \
    --lr_warmup_fraction 0.05 \
    --min_lr 1e-5 \
    --save megatron_output/Qwen3-235B-A22B-Instruct-2507 \
    --eval_interval 200 \
    --save_interval 200 \
    --max_length 2048 \
    --num_workers 8 \
    --dataset_num_proc 8 \
    --no_save_optim true \
    --no_save_rng true \
    --sequence_parallel true \
    --attention_backend flash \
    --model_author swift \
    --model_name swift-robot

3. Convert Megatron-format weights to HF format and test the conversion accuracy:

# 8 * 80GiB
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
swift export \
    --mcore_adapters megatron_output/Qwen3-235B-A22B-Instruct-2507/vx-xxx \
    --to_hf true \
    --torch_dtype bfloat16 \
    --output_dir megatron_output/Qwen3-235B-A22B-Instruct-2507/vx-xxx-hf

4. After training is completed, use the following command for inference:

# 8 * 80GiB
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
swift infer \
    --model output/vx-xxx/checkpoint-xxx \
    --infer_backend vllm \
    --stream true \
    --temperature 0 \
    --vllm_tensor_parallel_size 8 \
    --vllm_max_model_len 8192 \
    --max_new_tokens 2048

5. Push the model to ModelScope:

swift export \
    --model output/vx-xxx/checkpoint-xxx \
    --push_to_hub true \
    --hub_model_id '<your-model-id>' \
    --hub_token '<your-sdk-token>'

[gemini-code-assist]

Summary of Changes

[gemini-code-assist]

Code Review

This pull request adds support for the Qwen3-235B-A22B-Instruct-250718 model by updating the model registration and documentation. The core changes are correct. However, I have a couple of suggestions for improvement:

1. Code Organization: In swift/llm/model/model/qwen.py, the new model is added in a separate ModelGroup. It would be more consistent to add it to the existing ModelGroup for qwen3_moe models.

Addressing these points will improve the maintainability and clarity of the codebase.

[gemini-code-assist]

For better code organization and maintainability, it would be clearer to add the new model Qwen/Qwen3-235B-A22B-Instruct-250718 to the existing ModelGroup for qwen3_moe models, rather than creating a new ModelGroup for a single model. You can add it under the # instruct comment in the first ModelGroup.