1. Test Purpose

We plan to build a three-level CI pipeline for NPU, including PR, Nightly, and Manual pipelines, to ensure the code quality of SGLang on NPU hardware.

2. Test Scope

NPU test cases mainly include black-box test cases such as functionality and performance.
Since unit test (UT) cases are independent of the underlying hardware, we reuse the existing GPU test cases.

1. Functionality Test: Test the basic functions and features supported by NPU as listed in docs/platforms/ascend_npu_support_features.md.

2. Model Accuracy Test: Perform accuracy tests on models supported by NPU as listed in docs/platforms/ascend_npu_support_models.md.

3. Performance Test: Perform performance tests based on the performance best practices for NPU described in docs/platforms/ascend_npu_best_practice.md.

4. Stability Test: Short-term stability test on NPU hardware.

3. Pipeline Planning

1. PR Pipeline: Include key functions supported by NPU to prevent merged code from introducing critical issues and breaking basic functions. Triggered on each PR submission.
   workflow yaml file: pr-test-npu.yml

2. Nightly Pipeline: Include basic functions supported by NPU, accuracy and performance tests of major models supported by NPU. Triggered daily at 01:00 (Beijing Time).
   workflow yaml file: nightly-test-npu.yml

3. Manual Pipeline: Covers all function supported by NPU, accuracy and performance tests of all models supported by NPU. Manually triggered by maintainers, mainly for release verification to ensure high-quality delivery to users.
   workflow yaml file: To be started

4. Related PRs

• Optimize .github/workflows/nightly-test-npu.yml and add basic functional test cases : Add NPU basic function testcases #19382

• Add performance test workflow and test cases: Add ascend performance test cases. #18522