add npu fsdp example

[addsubmuldiv]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Write the detail information belongs to this PR.

Experiment results

Paste your experiment result here(if needed).

[gemini-code-assist]

Summary of Changes

Hello @addsubmuldiv, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request provides a comprehensive new example for fine-tuning the Qwen3-32B model using LoRA and Fully Sharded Data Parallel (FSDP) on Ascend NPU hardware. It aims to offer a clear, runnable demonstration of distributed training best practices, incorporating bfloat16 precision and Flash Attention 2 for efficient large language model training.

Highlights

• New FSDP Example for NPU: This pull request introduces a new example demonstrating how to perform LoRA fine-tuning of the Qwen3-32B model using Fully Sharded Data Parallel (FSDP) on NPU (Ascend) hardware.
• Accelerate Configuration: A dedicated fsdp.json configuration file is added, detailing accelerate settings for FSDP, including bf16 mixed precision, 8 processes, and specific FSDP parameters like TRANSFORMER_BASED_WRAP for Qwen3DecoderLayer and FULL_SHARD sharding.
• Training Script: A train.sh script is included to launch the example, setting up environment variables for NPU devices and configuring training parameters such as per_device_train_batch_size, gradient_accumulation_steps, gradient_checkpointing, and flash_attention_2.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces a new example for training on NPU with FSDP. The changes consist of an accelerate configuration file and a training script. My review provides suggestions to improve file formatting and script conventions, such as removing trailing whitespace, adding missing newlines, and including a shebang in the shell script for better portability and consistency.

[gemini-code-assist]

This JSON file has a couple of formatting issues that should be addressed for consistency:

• There are trailing whitespaces on lines 6, 8, 19, 20, and 23.
• The file is missing a newline character at the end.

I've provided a suggestion to fix both issues.

{
  "compute_environment": "LOCAL_MACHINE",
  "debug": false,
  "distributed_type": "FSDP",
  "downcast_bf16": "no",
  "mixed_precision": "bf16",
  "num_machines": 1,
  "num_processes": 8,
  "machine_rank": 0,
  "rdzv_backend": "static",
  "same_network": true,
  "use_cpu": false,
  "fsdp_config": {
    "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
    "fsdp_transformer_cls_names_to_wrap": "Qwen3DecoderLayer",
    "fsdp_sharding_strategy": "FULL_SHARD",
    "fsdp_backward_prefetch": "BACKWARD_PRE",
    "fsdp_forward_prefetch": true,
    "fsdp_limit_all_gathers": true,
    "fsdp_state_dict_type": "FULL_STATE_DICT",
    "fsdp_sync_module_states": true,
    "fsdp_cpu_ram_efficient_loading": true,
    "fsdp_use_orig_params": true,
    "fsdp_offload_params": false
  }
}

[gemini-code-assist]

To improve the script and adhere to standard practices, please consider the following:

1. Add a shebang (e.g., #!/bin/bash) at the top of the script. This ensures it's executed with the correct interpreter.
2. Add a newline character at the end of the file. This is a common convention for text files.

#!/bin/bash
# hardware: Atlas 900 A2 (910B1 8*64G)
export TASK_QUEUE_ENABLE=2
export CPU_AFFINITY_CONF=2
ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
accelerate launch --config_file "./examples/ascend/train/qwen3_lora_fsdp/fsdp.json" \
    swift/cli/sft.py \
    --model 'Qwen/Qwen3-32B' \
    --train_type lora \
    --dataset 'swift/self-cognition#1000' \
    --torch_dtype bfloat16 \
    --per_device_train_batch_size 10 \
    --gradient_accumulation_steps 2 \
    --gradient_checkpointing true \
    --gradient_checkpointing_kwargs '{"use_reentrant": false}' \
    --max_length 1200 \
    --num_train_epochs 2 \
    --eval_strategy no \
    --save_steps 500 \
    --logging_steps 1 \
    --dataloader_num_workers 8 \
    --dataset_num_proc 8 \
    --save_total_limit 2 \
    --save_only_model true \
    --output_dir output \
    --attn_impl 'flash_attention_2' \
    --packing true

[Copilot]

Pull Request Overview

This PR adds a new example demonstrating FSDP (Fully Sharded Data Parallel) training for Qwen3-32B model with LoRA on Ascend NPU hardware (Atlas 900 A2 with 910B1 8*64G). It provides a complete setup for distributed training on Huawei's Ascend NPU platform.

Key Changes

• Adds FSDP configuration file for Qwen3 model training with 8-way data parallelism
• Provides training script with NPU-specific environment settings and optimized hyperparameters
• Enables flash attention and gradient checkpointing for memory efficiency

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.

File	Description
examples/ascend/train/qwen3_lora_fsdp/fsdp.json	FSDP configuration with full sharding strategy, backward/forward prefetch optimizations, and Qwen3DecoderLayer wrapping policy
examples/ascend/train/qwen3_lora_fsdp/train.sh	Training script with NPU device visibility settings, accelerate launch configuration, and training hyperparameters for Qwen3-32B LoRA fine-tuning

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Remove trailing whitespace at the end of this line. Trailing whitespace can cause issues with version control and code formatting tools.

Suggested change

	"mixed_precision": "bf16",
	"mixed_precision": "bf16",

[Copilot]

Remove trailing whitespace at the end of this line. Trailing whitespace can cause issues with version control and code formatting tools.

Suggested change

	"fsdp_limit_all_gathers": true,
	"fsdp_limit_all_gathers": true,

[Copilot]

Remove trailing whitespace at the end of this line. Trailing whitespace can cause issues with version control and code formatting tools.

Suggested change

	"fsdp_state_dict_type": "FULL_STATE_DICT",
	"fsdp_state_dict_type": "FULL_STATE_DICT",

[Copilot]

Remove trailing whitespace at the end of this line. Trailing whitespace can cause issues with version control and code formatting tools.

Suggested change

	"fsdp_use_orig_params": true,
	"fsdp_use_orig_params": true,

[ji-huazhong]

Suggested change

	# hardware: Atlas 900 A2 (910B1 8*64G)
	# hardware: Atlas 900 A2