Skip to content

[Tracking] SGLang CI/CD Test Coverage Improvements Initiative #20514

@nvpohanh

Description

@nvpohanh

SGLang CI/CD Test Coverage Improvements Initiative

This is the top-level tracking issue for a comprehensive initiative to improve SGLang's CI test coverage, with a focus on Blackwell GPU support.

Motivation

As SGLang expands to support more models, more hardware (especially Blackwell GPUs), and more deployment configurations, our CI test coverage needs to keep pace. Currently:

  • Many tests run on Hopper but not on Blackwell
  • Kernel-level tests mostly cover Triton-based kernels, leaving external backends (FlashInfer, DeepGemm, DeepEP, etc.) untested at the kernel level
  • E2E accuracy tests are too expensive to scale across many configurations
  • Key model/config combinations used in production may lack dedicated CI coverage

Plan

Proposed Test Hierarchy

Image

Backlog: Hopper vs Blackwell Test Gaps

#20507

Tracks 112 test files (390 methods) that run on Hopper but not on Blackwell. 109 targeted for nightly, 3 for weekly. Tests span across all coverage layers. No active work planned — kept as a backlog for reference.

Roadmaps

Q2 2026 (April–June)

#20847

Level April May June
Full E2E + Disagg First test: DSR1 FP8 1P1D DEP8 Expand features/configs Expand to more models + transfer layers
Full E2E + Agg Key models: DSR1, Qwen3.5, GLM5 More models: Minimax-M2.5, Qwen3-Coder TBD
Reduced Layers PoC: DSR1 (4 layers) on H200/B200 Expand: 3 models x many configs More models
Kernel Tests

Weekly Progress

2026-03-18

2026-03-13

  • Created all four tracking issues and DLFW epics
  • Identified 112 test files for Part 1 Hopper-to-Blackwell migration
  • Identified 60 kernel test items for Part 2
  • Defined test design for Part 3 reduced-layers tests (8 models)
  • Defined 32 model/config items for Part 4

Metadata

Metadata

Assignees

Labels

cicontinue integration relatednvidia

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions