From e8deda53a15eb016525dfeed4485ae41cb6ac1ed Mon Sep 17 00:00:00 2001
From: xvxuopop <127376094+xvxuopop@users.noreply.github.com>
Date: Thu, 18 Dec 2025 21:27:00 +0800
Subject: [PATCH] [example] add Qwen3 series examples (#9624)

Co-authored-by: UsernameFull <tohowtodoit@gmail.com>
---
 examples/ascend/qwen3_full_sft_fsdp2.yaml     | 45 +++++++++++++++++
 examples/ascend/qwen3moe_full_sft_fsdp.yaml   | 46 ++++++++++++++++++
 .../ascend/qwen3vlmoe_full_sft_fsdp2.yaml     | 48 +++++++++++++++++++
 3 files changed, 139 insertions(+)
 create mode 100644 examples/ascend/qwen3_full_sft_fsdp2.yaml
 create mode 100644 examples/ascend/qwen3moe_full_sft_fsdp.yaml
 create mode 100644 examples/ascend/qwen3vlmoe_full_sft_fsdp2.yaml

diff --git a/examples/ascend/qwen3_full_sft_fsdp2.yaml b/examples/ascend/qwen3_full_sft_fsdp2.yaml
new file mode 100644
index 000000000..04351ddd6
--- /dev/null
+++ b/examples/ascend/qwen3_full_sft_fsdp2.yaml
@@ -0,0 +1,45 @@
+# Start FSDP2 fine-tuning
+# accelerate launch \
+#     --config_file examples/accelerate/fsdp2_config.yaml \
+#     src/train.py examples/ascend/qwen3_full_sft_fsdp2.yaml
+# Change `num_processes` in fsdp2_config.yaml to 16 in A3
+
+### model
+model_name_or_path: Qwen/Qwen3-8B
+trust_remote_code: true
+use_v1_kernels: true
+flash_attn: fa2
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+
+### dataset
+dataset: alpaca_en_demo
+template: qwen3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/Qwen3-8B/full/sft
+logging_steps: 1
+save_steps: 500
+max_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 8
+gradient_accumulation_steps: 1
+learning_rate: 1.0e-5
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 1800
+resume_from_checkpoint: null
diff --git a/examples/ascend/qwen3moe_full_sft_fsdp.yaml b/examples/ascend/qwen3moe_full_sft_fsdp.yaml
new file mode 100644
index 000000000..659918b10
--- /dev/null
+++ b/examples/ascend/qwen3moe_full_sft_fsdp.yaml
@@ -0,0 +1,46 @@
+# Start FSDP fine-tuning
+# accelerate launch \
+#     --config_file examples/accelerate/fsdp_config.yaml \
+#     src/train.py examples/ascend/qwen3moe_full_sft_fsdp.yaml
+# Change `num_processes` in fsdp_config.yaml to 16 in A3
+
+### model
+model_name_or_path: Qwen/Qwen3-30B-A3B-Instruct-2507
+trust_remote_code: true
+use_v1_kernels: true
+flash_attn: fa2
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+disable_gradient_checkpointing: false
+
+### dataset
+dataset: alpaca_zh
+template: qwen3
+cutoff_len: 1024
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/Qwen3-30B-A3B-Instruct-2507/full/sft
+logging_steps: 1
+save_steps: 500
+max_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: true
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 4
+gradient_accumulation_steps: 1
+learning_rate: 1.0e-4
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+resume_from_checkpoint: null
+seed: 1234
diff --git a/examples/ascend/qwen3vlmoe_full_sft_fsdp2.yaml b/examples/ascend/qwen3vlmoe_full_sft_fsdp2.yaml
new file mode 100644
index 000000000..d841804d6
--- /dev/null
+++ b/examples/ascend/qwen3vlmoe_full_sft_fsdp2.yaml
@@ -0,0 +1,48 @@
+# Start FSDP2 fine-tuning
+# accelerate launch \
+#     --config_file examples/accelerate/fsdp2_config.yaml \
+#     src/train.py examples/ascend/qwen3vlmoe_full_sft_fsdp2.yaml
+# Change `num_processes` in fsdp2_config.yaml to 16 in A3
+
+### model
+model_name_or_path: Qwen/Qwen3-VL-30B-A3B-Instruct
+image_max_pixels: 262144
+video_max_pixels: 16384
+trust_remote_code: true
+use_v1_kernels: true
+flash_attn: fa2
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+disable_gradient_checkpointing: false
+
+### dataset
+dataset: llava_1k_en, llava_1k_zh
+template: qwen3_vl
+cutoff_len: 1024
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 4
+
+### output
+output_dir: saves/Qwen3-VL-30B-A3B-Instruct/full/sft
+logging_steps: 1
+save_steps: 500
+max_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: true
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 1
+learning_rate: 1.0e-4
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+resume_from_checkpoint: null
+seed: 1234