Coverage Overview
Replicates sgl-project/sglang's CI Coverage Overview workflow output, with our AMD-vs-NVIDIA exclusion rules applied.AMD vs NVIDIA Coverage
-
AMD Tests
-
enabled / - totalNVIDIA Tests
-
AMD-comparable, enabled.All Registrations
-
- disabledBackend Summary
| Backend | Total comparable | Enabled | NV-only | NV-only dup | Disabled | Total | Per-Commit | Nightly | PR % | Nightly % |
|---|
AMD Enabled ÷ NVIDIA-comparable Enabled.
PR % / Nightly % split that ratio by cadence, taking the
NVIDIA row's counts as the target
(AMD Per-Commit ÷ NVIDIA Per-Commit, and the same for
Nightly). This is where the headline comes from: AMD runs more nightly tests
than NVIDIA, so Nightly % is above 100% and carries the overall figure up
while PR % — what a PR actually gates on — sits well below it.
Being - registrations short of the PR target is
not the same as - per-commit NVIDIA tests going
unrun on AMD (NVIDIA Tests Not Run on AMD): AMD also
runs - per-commit tests NVIDIA doesn't,
mostly under test/registered/amd/, and those offset the count.
Click Show details for how each column is derived.
The percentage on the headline card is the AMD row's Enabled
divided by the NVIDIA row's Enabled.
The denominator is NVIDIA-comparable: it already drops the
NV-only registrations — NVIDIA tests AMD doesn't run
(hardware-specific suites, SM120/Marlin-LoRA, NVIDIA-only kernel
benchmarks), listed under
AMD-Excluded NVIDIA Tests.
Per-Commit / Nightly split the AMD-comparable Enabled
count by cadence.
The greyed +N beside them is the NV-only share of
that cadence, so
Per-Commit + Nightly + both +N = Enabled + NV-only.
Total comparable = Enabled + Disabled, the
AMD-comparable registrations. Disabled tests count toward
Total but are excluded from coverage on both sides, so
Total = Enabled + NV-only + Disabled and only the
Enabled columns drive the coverage percentage.
NV-only dup = NV-only − unique files: NV-only
counts registrations, but the
AMD-Excluded NVIDIA Tests panel lists
unique files, so files with multiple registrations (plus
NV-only multimodal_gen) make NV-only larger than the panel's entry count.
Backend Summary
Test Tiers (CUDA vs AMD)
why "overall" < "core"| Tier | Cadence | NVIDIA | NV-only | AMD | Not on AMD | Coverage |
|---|
register_amd_ci(...) — the gap Coverage can hide, since Coverage divides totals rather than measuring overlap (AMD's nightly surface is larger than NVIDIA's, which is also why core can exceed 100%). Listed under NVIDIA Tests Not Run on AMD. It credits AMD for running a file at any cadence, so - further registrations that NVIDIA gates every PR on, but AMD only covers nightly, are not counted.
Kernel/Total AMD shows registered (+sgl-kernel existing); Coverage shows registered% (potential% incl. sgl-kernel).
core = base-* / stage-* suites plus nightly registrations currently classified in the core bucket; use Backend Summary's Nightly column to split cadence.
extra = extra-* suites (incl. KV-canary) — run 2×/day on schedule + opt-in PRs (run-ci-extra label), not every PR.
kernel = JIT/sgl-kernel tests (upstream PR #27644, merged 2026-06-09), incl. kernel perf benchmarks (*-kernel-benchmark-*), which AMD runs via jit-kernel-unit-test-amd. NVIDIA = test/registered/kernels registrations; AMD = - registered + - run directly from sgl-kernel/tests/ via the sgl-kernel-unit-test-amd job (not register_amd_ci, so not in test/registered). Indicative, not exact. Note: the nightly-kernel-1-gpu suite is registered nightly=True but not yet wired into the nightly pipeline — it runs via the per-PR jit_kernel job.
Tests by Folder
| Folder | CUDA | AMD | NPU | CPU | Total |
|---|
AMD-Excluded NVIDIA Tests (-)
Click to collapse
NVIDIA registrations that the dashboard's AMD-coverage view rules out before
comparing AMD vs NVIDIA totals, all defined in compare_suites.py:
hardware-specific suites (NVIDIA_HARDWARE_SPECIFIC_SUITES),
SM120/Marlin-LoRA tests gated by _NVIDIA_ONLY_FILE_PATTERNS,
NVIDIA-only kernel perf benchmarks
(_CUDA_ONLY_BENCHMARK_FILES plus NVFP4/MXFP8/FP4-indexer/SM90),
and the per-file curated list AMD_EXCLUSION_REASONS — tests whose
own source gates them to NVIDIA (an is_hip bail-out, a compute-capability
requirement, a hardcoded fa3/fa4/flashinfer backend, or a CUDA-only library
under test). Contrast with
NVIDIA Tests Not Run on AMD, which AMD could run.
Grouped by category — click a group to expand.
NVIDIA Tests Not Run on AMD (-)
Click to expand
NVIDIA registrations the dashboard counts as AMD-comparable (so they are
in the coverage denominator) whose file has no enabled
register_amd_ci(...). Unlike
AMD-Excluded NVIDIA Tests, nothing about the
hardware or libraries blocks these — a registration would close each one.
This is the gap the headline percentage can hide: it divides totals, and
AMD's nightly surface is larger than NVIDIA's. Grouped by tier, then folder.
Disabled Tests (-)
Click to expand| File | Folder | Backend | Suite | Reason |
|---|
Coverage Trends (Last 30 Days)
Coverage detail by day
day-over-day change & the PRs behind itRecently Added Tests
...Roadmap
The NVIDIA Tests Not Run on AMD list
(- files) is a backlog, not a verdict:
nothing about the hardware blocks those tests, they just have no
register_amd_ci(...). The roadmap page sorts them into
phases — kernel suites, 1-GPU per-commit, multi-GPU, extra, nightly —
and shows the coverage each phase would buy, the runner time it costs,
and the target AMD suite per file.