Skip to content

CI: Initial Aiter benchmark tests - #2115

Closed
gyohuangxin wants to merge 0 commit into
mainfrom
benchmark_release
Closed

CI: Initial Aiter benchmark tests#2115
gyohuangxin wants to merge 0 commit into
mainfrom
benchmark_release

Conversation

@gyohuangxin

Copy link
Copy Markdown
Member

No description provided.

@gyohuangxin
gyohuangxin marked this pull request as ready for review February 26, 2026 08:14
@gyohuangxin
gyohuangxin requested review from a team and Copilot February 26, 2026 08:14
@gyohuangxin
gyohuangxin marked this pull request as draft February 26, 2026 08:14

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR introduces a new GitHub Actions workflow to run Aiter benchmark tests using a reusable workflow from the ROCm/ATOM repository. The workflow is configured to test the DeepSeek-R1-0528 model with specific parameters including FP8 KV cache and 8-way tensor parallelism.

Changes:

  • Added new CI workflow .github/workflows/aiter-benchmark.yaml that calls an external reusable workflow from ROCm/ATOM
  • Configured to run on pull requests and manual dispatch (marked as "for testing")
  • Sets up benchmark parameters for DeepSeek-R1-0528 model testing

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

name: Aiter Benchmark

on:
pull_request: # for testing

Copilot AI Feb 26, 2026

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The comment "for testing" suggests this is a temporary configuration. Most other CI workflows in the repository include a push trigger for the main branch in addition to pull_request. Consider whether the push trigger should be added for consistency with workflows like aiter-test.yaml, atom-test.yaml, vllm_benchmark.yaml, and sglang_downstream.yaml, which all trigger on both push and pull_request events.

Suggested change
pull_request: # for testing
push:
branches: [main]
paths-ignore:
- '**/*.md'
- 'docs/**'
- 'LICENSE'
- '.gitignore'
pull_request:

Copilot uses AI. Check for mistakes.
name: Aiter Benchmark
uses: ROCm/ATOM/.github/workflows/atom-benchmark.yaml@atom_benchmark_expose
with:
deepseek-r1-0528: true

Copilot AI Feb 26, 2026

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The parameter name 'deepseek-r1-0528' with a boolean value 'true' suggests this is a feature flag. However, this doesn't match the typical model naming conventions seen in other workflows. In atom-test.yaml, the model is referenced as "DeepSeek-R1-0528" (line 49-51) with corresponding model_path "deepseek-ai/DeepSeek-R1-0528". Verify that the reusable workflow 'ROCm/ATOM/.github/workflows/atom-benchmark.yaml@atom_benchmark_expose' expects this exact parameter name and format. If this is meant to specify a model, consider using a more descriptive parameter name like 'model_name' or 'enable_deepseek_r1_0528' to make the intent clearer.

Suggested change
deepseek-r1-0528: true
enable_deepseek_r1_0528: true

Copilot uses AI. Check for mistakes.
jobs:
aiter_benchmark:
name: Aiter Benchmark
uses: ROCm/ATOM/.github/workflows/atom-benchmark.yaml@atom_benchmark_expose

Copilot AI Feb 26, 2026

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The workflow references a reusable workflow at branch 'atom_benchmark_expose'. Verify that this branch exists in the ROCm/ATOM repository and is the intended stable reference. Using a branch name (rather than a tag or commit SHA) means the workflow behavior could change if the branch is updated. Consider whether this should reference a specific tag or commit for stability, especially if this is intended for production use beyond testing.

Suggested change
uses: ROCm/ATOM/.github/workflows/atom-benchmark.yaml@atom_benchmark_expose
uses: ROCm/ATOM/.github/workflows/atom-benchmark.yaml@0123456789abcdef0123456789abcdef01234567

Copilot uses AI. Check for mistakes.
with:
deepseek-r1-0528: true
extra_args: "--kv_cache_dtype fp8 -tp 8"
param_list: "1024,1024,128,0.8" No newline at end of file

Copilot AI Feb 26, 2026

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The parameter 'param_list' with value "1024,1024,128,0.8" is passed without any documentation or comments explaining what these comma-separated values represent. Based on typical benchmark parameters, these could be input_length, output_length, batch_size, and some ratio, but this is not clear. Consider adding an inline comment to document what these parameters mean, similar to how other workflows document their configurations.

Suggested change
param_list: "1024,1024,128,0.8"
param_list: "1024,1024,128,0.8" # input_length, output_length, batch_size, request_ratio

Copilot uses AI. Check for mistakes.
with:
deepseek-r1-0528: true
extra_args: "--kv_cache_dtype fp8 -tp 8"
param_list: "1024,1024,128,0.8" No newline at end of file

Copilot AI Feb 26, 2026

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

When calling a reusable workflow from an external repository (ROCm/ATOM), secrets are not automatically inherited and must be explicitly passed. Many workflows in this repository use secrets like DOCKER_PASSWORD and HF_TOKEN_TEST. If the reusable atom-benchmark.yaml workflow requires any secrets (e.g., for Docker login or model downloads), you need to add a 'secrets:' section to pass them explicitly, such as 'secrets: inherit' or by passing specific secrets.

Suggested change
param_list: "1024,1024,128,0.8"
param_list: "1024,1024,128,0.8"
secrets: inherit

Copilot uses AI. Check for mistakes.
@gyohuangxin
gyohuangxin marked this pull request as ready for review February 26, 2026 09:47
@gyohuangxin
gyohuangxin marked this pull request as draft February 27, 2026 02:12
@valarLip valarLip closed this Mar 18, 2026
@zufayu
zufayu deleted the benchmark_release branch August 11, 2026 06:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants