Motivation

This PR aims to comprehensively improve the test coverage of the Ascend (NPU) backend for the SGLang framework. A large number of targeted test cases have been added, and some existing test files have been optimized and adjusted. In total,48 files are involved (35 new files, 13 optimized files). The detailed changes are as follows:

1、Test Infrastructure Optimization
Adjusted the nightly-test-npu.yml workflow configuration, added the disaggregation_utils.py base class for PD disaggregation scenarios, and improved the test_ascend_utils.py test base class to support dependency management and model constant configuration for the new test cases.

2、Core Function Test Enhancement
Added dedicated test cases for basic functions including multi-level radix cache (L1/L2), expert parallelism, and speculative inference, covering validation for different input scenarios, models (DeepSeek, Qwen3, etc.), and performance metrics (TTFT).

3、Interface Layer Test Improvement
Fully supplemented Ascend platform API-related tests, covering scenarios such as chat/completions interfaces, function calling, stop matching, penalty, and profile performance analysis.

4、Model Test Iteration
Added or adjusted Ascend adaptation tests for various models including Gemma3, Phi4 multimodal, BGE reranker, and InternLM2 reward model; optimized existing test cases for models such as C4AI Command-R and DeepSeek V3.2. Slightly adjusted model test code for InternLM2, GLM4, etc., removed redundant logic, and improved the accuracy and integrity of test cases.

Modifications

1、Update nightly-test-npu.yml: Add test case dependency on third-party package installation.

2、Add disaggregation_utils.py：PD disaggregation scenario base class file.

3、Update the constant file test_ascend_utils.py, add model constants, and update model-related test cases.

4、Add model test cases, covering the models:
LLM Models, Reward Models, Rerank Models.

5、Add API test cases, covering the APIs:
/health; /health_generate; /ping; /model_info; /server_info; /get_load; /v1/models; /v1/models/{model:path}; /generate; /v1/chat/completions.

6、Add test cases for basic function:
L1 and L2 radix cache, deepep, speculative_inference.

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.