Support using SwanLab for experiment tracking #98

xichengpro · 2025-06-09T16:05:00Z

A significant number of users are unable to access wandb due to network restrictions and are more accustomed to using the localized tool SwanLab. To improve the project's usability and local compatibility, this PR adds support for integrating with SwanLab.

garrett4wade

Thank you for your contribution! The integration of swanlab configuration is fine, but some logging utilities should be merged with existing ones using wandb.

garrett4wade · 2025-06-10T08:49:11Z

docs/tutorial/quickstart.md

@@ -99,10 +99,14 @@ python3 training/main_sync_ppo.py --help

 We recommend using Weights & Biases (wandb) for monitoring. Run `wandb login` or set the `WANDB_API_KEY` environment variable. Set `wandb.mode=True` in your configuration to upload training statistics.

+Alternatively, you can use SwanLab for monitoring. Run swanlab login or set the `SWANLAB_API_KEY` environment variable. Set `swanlab.mode=True` in your configuration to upload training statistics.


Could you add a hyper link to swanlab to let more people know about it?

Also, use "`" to quote `swanlab login`.

swanlab.mode should be set to online if it has the same API as wandb. The previous typo was fixed in #100

Can you try to merge the two lines about wandb and swanlab somehow?

Thank you for the feedback!

I've updated the documentation based on your suggestions:

Added official links to Weights & Biases and SwanLab for better user reference.

Used backticks to quote commands and parameters (e.g., wandb login, swanlab login).

Updated swanlab.mode usage to align with WandB's API convention, now using "local" and "cloud" instead of True.

Merged WandB and SwanLab descriptions into a single, concise statement for better readability.

Added a note about using swanlab.mode="local" if the server is unreachable.

garrett4wade · 2025-06-10T08:49:48Z

pyproject.toml

@@ -61,6 +61,8 @@ dependencies = [
    "colorlog",
    "psutil",
    "pynvml",
+    "swanlab==0.6.2",


">=" or "=="?

garrett4wade · 2025-06-10T08:50:25Z

realhf/base/logging.py

@@ -158,6 +158,20 @@ def log_wandb_tensorboard(data, step=None, summary_writer=None):
        for key, val in data.items():
            summary_writer.add_scalar(f"{key}", val, step)

+def log_swanlab_tensorboard(data, step=None, summary_writer=None):


This function should be merged with log_wandb_tensorboard.

Thank you for the feedback!
I've refactored the code and merged log_swanlab_tensorboard with log_wandb_tensorboard into a single function called log_swanlab_wandb_tensorboard.

garrett4wade · 2025-06-10T08:53:34Z

realhf/system/model_function_call.py

@@ -447,6 +448,11 @@ async def run_step(self, buf_indices, sample, buffer_id: int):
                    step=ctrl.step_info.global_step,
                    summary_writer=self.summary_writer,
                )
+                logging.log_swanlab_tensorboard(


Merge this with log_wandb_tensorboard.

I've refactored the code and merged log_swanlab_tensorboard with log_wandb_tensorboard into a single function called log_swanlab_wandb_tensorboard.

garrett4wade · 2025-06-10T08:54:00Z

requirements.txt

+prettytable
+swanlab==0.6.2


Please double-check the version requirement.

I've modified the dependency to use the latest version automatically.

- Added official links for better user reference - Used backticks to quote commands and parameters - Unified mode settings to use "online" / "cloud" convention - Merged WandB and SwanLab descriptions into a single concise statement - Added note on using `swanlab.mode="local"` when server connection is unavailable

…o log_swanlab_wandb_tensorboard - Unified logging logic for SwanLab, WandB, and TensorBoard to reduce code duplication

- Updated SwanLab version in pyproject.toml - Updated SwanLab version in requirements.txt

- Config now uses provided arguments first - Falls back to reading from config.yaml if no input is given

xichengpro · 2025-06-12T12:59:13Z

Thanks for the feedback! I've updated the code based on your suggestion. Kindly review it again at your convenience.

garrett4wade

Thank you again for your contribution! We are almost there.

As a kind reminder, please format the files such that the CI will pass:

pip install -e .
# clear any external packages installed locally
rm -rf ./sympy
rm -rf ./sglang
# Run formatting
isort . && black .

garrett4wade · 2025-06-13T02:56:29Z

realhf/base/logging.py

 _LATEST_WANDB_STEP = 0
+_LATEST_SWANLAB_STEP = 0


These two step variables are the same. Remaining a single _LATEST_LOG_STEP will be fine.

Thanks for the feedback!
I've merged _LATEST_WANDB_STEP and _LATEST_SWANLAB_STEP into _LATEST_LOG_STEP.

garrett4wade · 2025-06-13T02:58:17Z

docs/tutorial/quickstart.md

@@ -97,12 +97,15 @@ python3 training/main_sync_ppo.py --help

 ## Monitoring the Training Process

-We recommend using Weights & Biases (wandb) for monitoring. Run `wandb login` or set the `WANDB_API_KEY` environment variable. Set `wandb.mode=online` in your configuration to upload training statistics.
+ We recommend using [Weights & Biases (wandb)](https://github.com/wandb/wandb)  or [SwanLab](https://github.com/SwanHubX/SwanLab)  for monitoring—run `wandb login` or `swanlab login`, or set the corresponding environment variable API key (`WANDB_API_KEY` or `SWANLAB_API_KEY`). Set `wandb.mode="online"` or `swanlab.mode="cloud"` in your configuration to upload training statistics. If you cannot connect to the server, you can also use `swanlab.mode="local"` to save data locally without uploading.


Please mention wandb.mode=offline together with swanlab.mode=local.

Thanks for the feedback!
I've added note on using wandb.mode="offline" together with swanlab.mode="local".

…EST_LOG_STEP

- Updated SwanLab version in requirements.txt

xichengpro · 2025-06-13T03:35:42Z

Thank you again for your contribution! We are almost there.

As a kind reminder, please format the files such that the CI will pass:
pip install -e .
# clear any external packages installed locally
rm -rf ./sympy
rm -rf ./sglang
# Run formatting
isort . && black .

Thank you for the reminder!

I've formatted the code using isort and black as requested, and all files should now conform to the project's style guidelines. The CI should now pass successfully.

Let me know if there's anything else I can improve!

garrett4wade · 2025-06-16T02:01:35Z

@GurrenLagann97 Can you provide another review?

GurrenLagann97

Seems perfect

Zeyi-Lin · 2025-06-16T15:08:36Z

🎉😄Thanks for your contribution to the SwanLab and AReaL community. @xichengpro

* Support using SwanLab for experiment tracking * docs: improve WandB and SwanLab integration documentation - Added official links for better user reference - Used backticks to quote commands and parameters - Unified mode settings to use "online" / "cloud" convention - Merged WandB and SwanLab descriptions into a single concise statement - Added note on using `swanlab.mode="local"` when server connection is unavailable * refactor: update default value of api_key * fix: correct help description from WandB to SwanLab in SwanLabConfig * refactor: merge log_swanlab_tensorboard and log_wandb_tensorboard into log_swanlab_wandb_tensorboard - Unified logging logic for SwanLab, WandB, and TensorBoard to reduce code duplication * chore: update swanlab version in dependency config files - Updated SwanLab version in pyproject.toml - Updated SwanLab version in requirements.txt * refactor: enhance SwanLab config handling for logging purposes - Config now uses provided arguments first - Falls back to reading from config.yaml if no input is given * docs: add note on using when server connection is unavailable * refactor: merge _LATEST_WANDB_STEP and _LATEST_SWANLAB_STEP into _LATEST_LOG_STEP * Format code with black and isort * chore: update swanlab version in dependency config files - Updated SwanLab version in requirements.txt * refactor: rename swanlab_wandb_data to log_data --------- Co-authored-by: dubingnan <[email protected]>

Support using SwanLab for experiment tracking

606650b

garrett4wade reviewed Jun 10, 2025

View reviewed changes

xichengpro and others added 6 commits June 12, 2025 14:28

Merge branch 'inclusionAI:main' into main

2cdc6f5

refactor: update default value of api_key

982c999

fix: correct help description from WandB to SwanLab in SwanLabConfig

61b83d5

refactor: merge log_swanlab_tensorboard and log_wandb_tensorboard int…

79430be

…o log_swanlab_wandb_tensorboard - Unified logging logic for SwanLab, WandB, and TensorBoard to reduce code duplication

chore: update swanlab version in dependency config files

7094932

- Updated SwanLab version in pyproject.toml - Updated SwanLab version in requirements.txt

xichengpro force-pushed the main branch 2 times, most recently from 8474630 to f86d103 Compare June 12, 2025 11:45

refactor: enhance SwanLab config handling for logging purposes

937a4e0

- Config now uses provided arguments first - Falls back to reading from config.yaml if no input is given

xichengpro force-pushed the main branch from f86d103 to 937a4e0 Compare June 12, 2025 11:47

garrett4wade reviewed Jun 13, 2025

View reviewed changes

bingnandu added 5 commits June 13, 2025 11:09

docs: add note on using when server connection is unavailable

fb6783d

refactor: merge _LATEST_WANDB_STEP and _LATEST_SWANLAB_STEP into _LAT…

4ca25a2

…EST_LOG_STEP

Format code with black and isort

dc85673

chore: update swanlab version in dependency config files

cae1950

- Updated SwanLab version in requirements.txt

refactor: rename swanlab_wandb_data to log_data

781ae3f

garrett4wade approved these changes Jun 16, 2025

View reviewed changes

garrett4wade requested a review from GurrenLagann97 June 16, 2025 02:02

GurrenLagann97 approved these changes Jun 16, 2025

View reviewed changes

garrett4wade merged commit bb14f02 into inclusionAI:main Jun 16, 2025
1 check passed

		@@ -99,10 +99,14 @@ python3 training/main_sync_ppo.py --help

		We recommend using Weights & Biases (wandb) for monitoring. Run `wandb login` or set the `WANDB_API_KEY` environment variable. Set `wandb.mode=True` in your configuration to upload training statistics.

		Alternatively, you can use SwanLab for monitoring. Run swanlab login or set the `SWANLAB_API_KEY` environment variable. Set `swanlab.mode=True` in your configuration to upload training statistics.

		prettytable
		swanlab==0.6.2

Support using SwanLab for experiment tracking #98

Support using SwanLab for experiment tracking #98

Uh oh!

Conversation

xichengpro commented Jun 9, 2025

Uh oh!

garrett4wade left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

xichengpro commented Jun 12, 2025

Uh oh!

garrett4wade left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

xichengpro Jun 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

xichengpro commented Jun 13, 2025

Uh oh!

garrett4wade commented Jun 16, 2025

Uh oh!

GurrenLagann97 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Zeyi-Lin commented Jun 16, 2025

Uh oh!

Uh oh!

xichengpro Jun 13, 2025 •

edited

Loading