Files
sglang/test
laoyao0822 f8fca72635 Expand prefill CP KV capacity by sharding persistent NSA KV
Prefill CP previously replicated NSA/MLA persistent KV on every CP rank, so CP8 consumed eight copies of KV memory while exposing only one rank of logical cache capacity. This change splits logical KV locs from per-rank physical storage, shards MLA latent KV and NSA index K/scale by deterministic page ownership, and keeps existing NSA attention kernels working through a full-view runtime materialization layer.

Mooncake PD transfer now sends each prefill CP rank's owned physical pages with explicit logical page positions so non-CP decode can reconstruct full-layout KV. The implementation is guarded by an explicit server flag and startup checks, and the design documentation records the implemented scope, debug environment, and Phase 3 boundary.

Constraint: Phase 2 must preserve existing NSA attention/index kernels via runtime full-view materialization
Constraint: Decode side remains non-CP and receives full KV through Mooncake
Rejected: Shard-aware NSA attention in this change | belongs to Phase 3 because it requires distributed topk/softmax/output contracts
Rejected: Request-contiguous CP ownership | unstable under chunked prefill and tied to attention split mode
Confidence: medium
Scope-risk: broad
Directive: Do not enable round-robin CP shared KV without wiring runtime materialization/PD transfer contracts for that split mode
Directive: Keep SGLANG_DEBUG_CP_SHARED_KV disabled for perf measurements; it intentionally enables CUDA-syncing diagnostics
Tested: Remote py_compile for shared-KV touched Python files in g0034 container
Tested: Remote pytest selected cp_shared/shared_kv/nsa suite: 37 passed, 34 deselected
Not-tested: Full GLM5 multi-node throughput/regression run after final doc update
Not-tested: Phase 3 shard-aware runtime, round-robin CP mode, and non-Mooncake PD backends
2026-04-26 04:11:17 +08:00
..

Run Unit Tests

SGLang uses the built-in library unittest as the testing framework.

Test Backend Runtime

cd sglang/test/srt

# Run a single file
python3 test_srt_endpoint.py

# Run a single test
python3 test_srt_endpoint.py TestSRTEndpoint.test_simple_decode

# Run a suite with multiple files
python3 run_suite.py --suite per-commit

Test Frontend Language

cd sglang/test/lang

# Run a single file
python3 test_choices.py

Adding or Updating Tests in CI

  • Create new test files under test/srt or test/lang depending on the type of test.
  • For nightly tests, place them in test/srt/nightly/. Use the NightlyBenchmarkRunner helper class in nightly_utils.py for performance benchmarking tests.
  • Ensure they are referenced in the respective run_suite.py (e.g., test/srt/run_suite.py) so they are picked up in CI. For most small test cases, they can be added to the per-commit-1-gpu suite. Sort the test cases alphabetically by name.
  • Ensure you added unittest.main() for unittest and sys.exit(pytest.main([__file__])) for pytest in the scripts. The CI run them via python3 test_file.py.
  • The CI will run some suites such as per-commit-1-gpu, per-commit-2-gpu, and nightly-1-gpu automatically. If you need special setup or custom test groups, you may modify the workflows in .github/workflows/.

CI Registry System

Tests in test/registered/ use a registry-based CI system for flexible backend/schedule configuration.

Registration Functions

from sglang.test.ci.ci_register import (
    register_cuda_ci,
    register_amd_ci,
    register_cpu_ci,
    register_npu_ci,
)

# Per-commit test (small 1-gpu, runs on 5090)
register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-small")

# Per-commit test (large 1-gpu, runs on H100)
register_cuda_ci(est_time=120, suite="stage-b-test-1-gpu-large")

# Per-commit test (2-gpu)
register_cuda_ci(est_time=200, suite="stage-b-test-2-gpu-large")

# Nightly-only test
register_cuda_ci(est_time=200, suite="nightly-1-gpu", nightly=True)

# Multi-backend test
register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-small")
register_amd_ci(est_time=120, suite="stage-a-test-1-gpu-small-amd")

# Temporarily disabled test
register_cuda_ci(est_time=80, suite="stage-b-test-1-gpu-small", disabled="flaky - see #12345")

Choosing Between 1-GPU Suites (5090 vs H100)

When adding 1-GPU tests, choose the appropriate suite based on hardware compatibility:

Suite Runner GPU When to Use
stage-a-test-1-gpu-small 1-gpu-5090 RTX 5090 (32GB, SM120) Stage A per-commit smoke on 5090 (CUDA)
stage-a-test-1-gpu-small-amd AMD CI runners ROCm Stage A per-commit smoke (AMD)
stage-b-test-1-gpu-small 1-gpu-5090 RTX 5090 (32GB, SM120) 5090-compatible tests (preferred)
stage-b-test-1-gpu-large 1-gpu-h100 H100 (80GB, SM90) Large models or 5090-incompatible tests

Use stage-b-test-1-gpu-small (5090) whenever possible - this is the preferred suite for most 1-GPU tests.

Use stage-b-test-1-gpu-large (H100) if ANY of these apply:

  1. Architecture incompatibility (SM120/Blackwell):

    • FA3 attention backend (requires SM≤90)
    • MLA with FA3 backend
    • FP8/MXFP4 quantization (not supported on SM120)
    • Certain Triton kernels (shared memory limits)
  2. Memory requirements:

    • Models >30B params or large MoE
    • Tests requiring >32GB VRAM
  3. Known 5090 failures:

    • Weight update/sync tests
    • Certain spec decoding tests

If a test cannot run on 5090 due to any of the above, use stage-b-test-1-gpu-large which runs on H100.

Available Suites

Per-Commit (CUDA):

  • Stage A: stage-a-test-1-gpu-small (5090), stage-a-test-2, stage-a-test-cpu
  • Stage B: stage-b-test-1-gpu-small (5090), stage-b-test-1-gpu-large (H100), stage-b-test-2-gpu-large
  • Stage C (4-GPU): stage-c-test-4-gpu-h100, stage-c-test-4-gpu-b200, stage-c-test-4-gpu-gb200, stage-c-test-deepep-4-gpu-h100
  • Stage C (8-GPU): stage-c-test-8-gpu-h20, stage-c-test-8-gpu-h200, stage-c-test-8-gpu-b200, stage-c-test-deepep-8-gpu-h200

Per-Commit (AMD):

  • stage-a-test-1-gpu-small-amd, stage-b-test-1-gpu-small-amd, stage-b-test-2-gpu-large-amd

Nightly:

  • nightly-1-gpu, nightly-2-gpu, nightly-4-gpu, nightly-8-gpu, etc.

Running Tests with run_suite.py

# Run per-commit tests
python test/run_suite.py --hw cuda --suite stage-b-test-1-gpu-small

# Run nightly tests
python test/run_suite.py --hw cuda --suite nightly-1-gpu --nightly

# With auto-partitioning (for parallel CI jobs)
python test/run_suite.py --hw cuda --suite stage-b-test-1-gpu-small \
    --auto-partition-id 0 --auto-partition-size 4

Writing Elegant Test Cases

  • Learn from existing examples in sglang/test/srt.
  • Reduce the test time by using smaller models and reusing the server for multiple test cases. Launching a server takes a lot of time.
  • Use as few GPUs as possible. Do not run long tests with 8-gpu runners.
  • If the test cases take too long, considering adding them to nightly tests instead of per-commit tests.
  • Keep each test function focused on a single scenario or piece of functionality.
  • Give tests descriptive names reflecting their purpose.
  • Use robust assertions (e.g., assert, unittest methods) to validate outcomes.
  • Clean up resources to avoid side effects and preserve test independence.
  • Reduce the test time by using smaller models and reusing the server for multiple test cases.

Adding New Models to Nightly CI

  • For text models: extend global model lists variables in test_utils.py, or add more model lists
  • For vlms: extend the MODEL_THRESHOLDS global dictionary in test/srt/nightly/test_vlms_mmmu_eval.py