minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD] register kv_canary + mock_model e2e tests to extra-a (1-gpu-small + 2-gpu-large) (#28850)

Browse Source
This commit is contained in:
Michael
2026-06-22 19:06:31 -07:00
committed by GitHub
parent 84338df6f0
commit 28d5627fd8
19 changed files with 87 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/run_suite.py
+7 -5
View File
@@ -46,15 +46,17 @@ PER_COMMIT_SUITES = {
# extra-a: label-gated PR opt-in suites in pr-test-amd-extra.yml
# (mirror of CUDA extra-a; tests stay tagged per-commit but only
# dispatch when the PR carries the `run-ci-extra` label). 1-gpu-small
# carries the mock-model / kv_canary *unit* tests; 1-gpu-large carries
# the subset of model e2e tests validated to pass on mi325 (quant
# fp8kv-triton, sessions streaming-session EAGLE3, spec standalone
# triton-backend variant). The rest of CUDA
# extra-a tests fail on ROCm (missing flash_attn.cute/flash_ops
# carries the mock-model / kv_canary unit + single-GPU canary e2e
# tests; 1-gpu-large carries the subset of model e2e tests validated
# to pass on mi325 (quant fp8kv-triton, sessions streaming-session
# EAGLE3, spec standalone triton-backend variant); 2-gpu-large carries
# the multi-GPU (TP/PP/PD) mock-model + kv_canary e2e tests. The rest
# of CUDA extra-a tests fail on ROCm (missing flash_attn.cute/flash_ops
# kernels, OOM, or accuracy regressions — e.g. gemma4-mtp-31b dips
# below the gsm8k floor on the topk=3 leg) and stay CUDA-only for now.
"extra-a-test-1-gpu-small-amd",
"extra-a-test-1-gpu-large-amd",
"extra-a-test-2-gpu-large-amd",
],
HWBackend.MUSA: [],
HWBackend.CUDA: [