Add fake balance for EP mode #962
Conversation
more robust balance management, from env control to option control, better logger info control, better format
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @STwangyingrui, 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 an experimental feature to improve the performance and stability of Expert Parallelism (EP) mode by implementing a 'fake balance' strategy for expert selection. This strategy, enabled via a new command-line flag, aims to distribute the workload more evenly across experts, leading to notable gains in both prefill and decoding throughput.
Highlights
- New Feature: Expert Balancing for EP Mode: Introduced a 'fake balance' mechanism specifically for Expert Parallelism (EP) mode. This aims to improve throughput and stabilize overheads by artificially balancing the distribution of tokens to experts.
- Configurable via Command Line: The new fake balance feature is controlled by a new command-line argument,
--enable_ep_fake_balance, allowing users to enable or disable this optimization. - Performance Improvements: Enabling this feature is reported to significantly increase prefill throughput by 35% and decoding throughput by 15%, while also stabilizing overheads.
- Expert Assignment Override: When enabled, the system will override the dynamically selected top-k experts with a pre-generated, balanced distribution of expert IDs, ensuring a more even load across experts.
- Conditional Logging for Debugging: Added logging in both prefill and decode stages to show the distribution of tokens to experts when the fake balance is active. Note that decode stage logging is only active when
cudagraphis disabled.
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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| 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 issue 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 is currently in preview and 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 to provide feedback.
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
-
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. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces a "fake balance" mode for expert parallelism. My main feedback is on the performance of the generate_balanced_tensor function in lightllm/utils/balance_utils.py. The current implementation has a performance bottleneck that can be addressed with a more efficient approach. I've provided a code suggestion for this improvement.
lightllm/utils/balance_utils.py
Outdated
| tensor = torch.zeros((num_tokens, self.num_selected), dtype=torch.int, device="cuda") | ||
| expert_load = torch.zeros(self.num_experts, dtype=torch.int, device="cuda") | ||
|
|
||
| for i in range(num_tokens): | ||
| available_experts = torch.arange(self.num_experts, device="cuda") | ||
| selected = [] | ||
| for _ in range(self.num_selected): | ||
| current_load = expert_load[available_experts] | ||
| min_load_indices = torch.where(current_load == current_load.min())[0] | ||
| if len(min_load_indices) > 1: | ||
| # If there are multiple least-loaded experts, select one randomly | ||
| chosen_index = torch.randint(0, len(min_load_indices), (1,), device="cuda").item() | ||
| chosen_expert_index = min_load_indices[chosen_index] | ||
| else: | ||
| chosen_expert_index = min_load_indices[0] | ||
| chosen_expert = available_experts[chosen_expert_index] | ||
| selected.append(chosen_expert) | ||
| # Remove the selected expert from the list of available experts | ||
| available_experts = torch.cat( | ||
| [available_experts[:chosen_expert_index], available_experts[chosen_expert_index + 1 :]] | ||
| ) | ||
| expert_load[chosen_expert] += 1 | ||
|
|
||
| tensor[i] = torch.tensor(selected, dtype=torch.int, device="cuda") |
There was a problem hiding this comment.
The current implementation of generate_balanced_tensor is inefficient due to the use of torch.cat inside a loop. This creates a new tensor and copies data in every iteration, which can be slow for large num_tokens or num_experts. A more performant approach would be to use a boolean mask to keep track of selected experts, avoiding the expensive torch.cat operation. This can significantly reduce the overhead.
tensor = torch.empty((num_tokens, self.num_selected), dtype=torch.int, device="cuda")
expert_load = torch.zeros(self.num_experts, dtype=torch.int, device="cuda")
for i in range(num_tokens):
selected_mask = torch.zeros(self.num_experts, dtype=torch.bool, device="cuda")
for j in range(self.num_selected):
# Use a large value for already selected experts to exclude them
load_view = torch.where(selected_mask, torch.iinfo(expert_load.dtype).max, expert_load)
min_load_indices = torch.where(load_view == load_view.min())[0]
if len(min_load_indices) > 1:
# If there are multiple least-loaded experts, select one randomly
rand_idx = torch.randint(0, len(min_load_indices), (1,), device="cuda").item()
chosen_expert = min_load_indices[rand_idx]
else:
chosen_expert = min_load_indices[0]
tensor[i, j] = chosen_expert
expert_load[chosen_expert] += 1
selected_mask[chosen_expert] = True
Add fake balance for EP mode, which is controled by option of --enable_ep_fake_balance.
Cost: EP8 batch128 input64 (40+ different seqlens) totally cost about 5 seconds.
Benefit: prefill throughput increase 35%, decoding throughput increase 15%, and the overheads become stable.