Implementation:Vllm project Vllm Test AMD Pipeline Config

Knowledge Sources	vllm
Domains	CI/CD, Testing, AMD GPU
Last Updated	2026-02-08 00:00 GMT

Overview

Defines the Buildkite test pipeline template for running the vLLM test suite on AMD ROCm GPU hardware.

Description

This YAML file specifies test steps for AMD GPU verification, including labels, commands, timeouts, source file dependencies, and hardware mirroring options. It is fed into a Jinja template at the vllm-project/buildkite-ci repository to generate the final pipeline YAML. The file covers fast-check tests, correctness tests, model-specific tests, distributed tests, and more, all targeting AMD MI-series GPUs.

Usage

Contributors modify this file to add, remove, or update test steps that run on AMD hardware. The pipeline is automatically triggered by Buildkite CI on commits and pull requests. Tests can be conditionally executed based on source file dependencies, and individual steps can be marked as optional, soft-fail, or fast-check-only.

Code Reference

Source Location

• Repository: vllm
• File: .buildkite/test-amd.yaml
• Lines: 1-1694

Signature

steps:
  - label: "Pytorch Nightly Dependency Override Check"
    soft_fail: true
    source_file_dependencies: [...]
    commands: [...]

  - label: "Async Engine, Inputs, Utils, Worker Test"
    timeout_in_minutes: 50
    mirror_hardwares: [amdexperimental]
    source_file_dependencies: [...]
    commands: [...]

  - label: "Basic Correctness Test"
    timeout_in_minutes: 30
    fast_check: true
    torch_nightly: true
    commands: [...]

Import

# This file is consumed by the Jinja template engine in Buildkite CI:
# https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2
#
# It is not imported directly; Buildkite processes this file to generate
# the final pipeline configuration for AMD hardware tests.

I/O Contract

Inputs

Name	Type	Required	Description
label	string	Yes	The display name of the test step (emojis allowed)
commands	list[string]	Yes	Shell commands to execute for the test (incompatible with command)
command	string	Yes	Single shell command to execute (incompatible with commands)
timeout_in_minutes	int	No	Timeout for the step; uses default if not specified
mirror_hardwares	list[string]	No	Additional hardware targets (e.g., [amdexperimental])
agent_pool	string	No	Specific agent pool to run on (e.g., mi325_1)
grade	string	No	Test grade classification (e.g., Blocking)
source_file_dependencies	list[string]	No	File path prefixes that trigger this test; if empty, test always runs
fast_check	bool	No	Whether to run on each commit in the fastcheck pipeline
torch_nightly	bool	No	Whether to run against PyTorch nightly builds
optional	bool	No	Never run by default unless manually unblocked or scheduled nightly
soft_fail	bool	No	Allow step to fail without failing the entire pipeline
no_gpu	bool	No	Run the test without GPU resources (CPU-only)
num_gpus	int	No	Number of GPUs required (default: 1)
parallelism	int	No	Number of parallel shards for the step

Outputs

Name	Type	Description
Test results	Buildkite step results	Pass/fail status for each test step displayed in Buildkite UI
Test logs	stdout/stderr	Console output from pytest and other test runners

Usage Examples

# Adding a new AMD test step:
- label: "My New AMD Test"
  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
    - vllm/my_module/
    - tests/my_test/
  commands:
    - pytest -v -s my_test/

Related Pages

• Environment:Vllm_project_Vllm_Buildkite
• Environment:Vllm_project_Vllm_AMD_ROCm