Add MFU logging support

[y2logic]

PR type

• New Feature

PR information

Implement a callback plugin to add MFU metrics to log.
Related to issue Add MFU (Model FLOPs Utilization) logging support #5791

Write the detail information belongs to this PR.

• Specify device TFLOPS by setting ENV DEVICE_TFLOPS. (Higher priority)
• Provide a quick test for estimating device compute capability.
• Provide a fallback strategy based on lookup tables, more data needed. (Community may have more reliable data)

Experiment results

[INFO:swift] model_parameter_info: PeftModelForCausalLM: 498.4319M Params (4.3991M Trainable [0.8826%]), 0.0000M Buffers.
[INFO:swift] Setting DEVICE_TFLOPS: None. You can adjust this hyperparameter through the environment variable: `DEVICE_TFLOPS`.
[INFO:swift] Estimating device TFLOPS baseline. Device: [cuda:0] dtype: [torch.float16]
[设备 cuda:0] 测试总耗时：8.3362s，平均耗时: 0.1389 s，dtype：torch.float16，性能: 7.9137 TFLOPS
[INFO:swift] Estimate test finished. [7.9137353585254075 TFLOPS] Device count: [1]
[INFO:swift] use_reentrant: True
[INFO:swift] The logging file will be saved in: /home/jovyan/y2logic/config/output/v8-20251105-011248/logging.jsonl
Train:   0%|                                                                                                   | 0/8 [00:00<?, ?it/s][INFO:swift] use_logits_to_keep: True
/opt/conda/lib/python3.11/site-packages/torch/utils/checkpoint.py:295: FutureWarning: `torch.cpu.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cpu', args...)` instead.
  with torch.enable_grad(), device_autocast_ctx, torch.cpu.amp.autocast(**ctx.cpu_autocast_kwargs):  # type: ignore[attr-defined]
{'loss': 1.92887628, 'grad_norm': 1.0251199, 'learning_rate': 1e-05, 'token_acc': 0.55102721, 'epoch': 0.13, 'MFU': 0.200113, 'global_step/max_steps': '1/8', 'percentage': '12.50%', 'elapsed_time': '23s', 'remaining_time': '2m 44s', 'memory(GiB)': 11.15, 'train_speed(iter/s)': 0.042512}
{'loss': 1.80279827, 'grad_norm': 0.94247454, 'learning_rate': 9.5e-06, 'token_acc': 0.5656051, 'epoch': 0.26, 'MFU': 0.202701, 'global_step/max_steps': '2/8', 'percentage': '25.00%', 'elapsed_time': '44s', 'remaining_time': '2m 14s', 'memory(GiB)': 11.15, 'train_speed(iter/s)': 0.044644}
{'loss': 1.72881317, 'grad_norm': 0.92661977, 'learning_rate': 8.12e-06, 'token_acc': 0.57771261, 'epoch': 0.38, 'MFU': 0.20362, 'global_step/max_steps': '3/8', 'percentage': '37.50%', 'elapsed_time': '1m 7s', 'remaining_time': '1m 51s', 'memory(GiB)': 11.15, 'train_speed(iter/s)': 0.044769}
{'loss': 1.74326658, 'grad_norm': 0.93214059, 'learning_rate': 6.11e-06, 'token_acc': 0.57969724, 'epoch': 0.51, 'MFU': 0.204517, 'global_step/max_steps': '4/8', 'percentage': '50.00%', 'elapsed_time': '1m 28s', 'remaining_time': '1m 28s', 'memory(GiB)': 11.15, 'train_speed(iter/s)': 0.045414}
{'loss': 1.88908231, 'grad_norm': 0.96761024, 'learning_rate': 3.89e-06, 'token_acc': 0.5600538, 'epoch': 0.64, 'MFU': 0.204593, 'global_step/max_steps': '5/8', 'percentage': '62.50%', 'elapsed_time': '1m 49s', 'remaining_time': '1m 5s', 'memory(GiB)': 11.15, 'train_speed(iter/s)': 0.045703}
{'loss': 1.7942878, 'grad_norm': 0.96311468, 'learning_rate': 1.88e-06, 'token_acc': 0.5789354, 'epoch': 0.77, 'MFU': 0.204409, 'global_step/max_steps': '6/8', 'percentage': '75.00%', 'elapsed_time': '2m 13s', 'remaining_time': '44s', 'memory(GiB)': 11.15, 'train_speed(iter/s)': 0.044814}
{'loss': 1.61809587, 'grad_norm': 0.84860581, 'learning_rate': 5e-07, 'token_acc': 0.60130132, 'epoch': 0.9, 'MFU': 0.203771, 'global_step/max_steps': '7/8', 'percentage': '87.50%', 'elapsed_time': '2m 37s', 'remaining_time': '22s', 'memory(GiB)': 11.15, 'train_speed(iter/s)': 0.044347}
{'loss': 1.79473865, 'grad_norm': 0.96563059, 'learning_rate': 0.0, 'token_acc': 0.57638475, 'epoch': 1.0, 'MFU': 0.203993, 'global_step/max_steps': '8/8', 'percentage': '100.00%', 'elapsed_time': '2m 56s', 'remaining_time': '0s', 'memory(GiB)': 11.15, 'train_speed(iter/s)': 0.045302}

Paste your experiment result here(if needed).

[gemini-code-assist]

Summary of Changes

Hello @y2logic, 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 introduces a significant enhancement to the training pipeline by integrating Model FLOPs Utilization (MFU) logging. This feature aims to provide developers and researchers with critical performance metrics, enabling them to better understand and optimize the efficiency of their hardware during model training. The implementation offers a flexible approach to determining theoretical device TFLOPS, either through direct user input via an environment variable or through an intelligent automatic estimation process, ensuring robust MFU reporting across different setups.

Highlights

• Introduction of PerfMetricsLogCallback: A new TrainerCallback named PerfMetricsLogCallback has been added to automatically calculate and log Model FLOPs Utilization (MFU) during the training process, providing insights into hardware efficiency.
• Flexible TFLOPS Configuration: Users can now specify the theoretical maximum TFLOPS of their device using the DEVICE_TFLOPS environment variable, which takes precedence over automatic estimation.
• Automatic TFLOPS Estimation: If the DEVICE_TFLOPS environment variable is not explicitly set, the system will automatically estimate the device's TFLOPS by running a quick matrix multiplication benchmark tailored to the device and data type.
• Device TFLOPS Lookup Table: A device_flops_map has been included, containing theoretical FLOPS values for various common GPU devices (e.g., H100, A100, RTX 3070 Ti, Ascend910), intended for future fallback strategies or reference.

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 PerfMetricsLogCallback to compute and log Model FLOPs Utilization (MFU), which is a valuable addition for performance monitoring. The implementation is well-structured, offering device TFLOPS estimation and configuration via environment variables. My review focuses on improving robustness, maintainability, and consistency. I've identified a bug in a helper function, suggested refactoring to reduce code duplication, recommended leveraging existing utilities, and pointed out some dead code. Addressing these points will enhance the quality and reliability of the new callback.