Use only one GPU for MLA CI tests (#2858)

2025-01-13 03:55:33 -08:00
parent 4536d72446
commit 67008f4b32
4 changed files with 39 additions and 8 deletions
--- a/test/srt/run_suite.py
+++ b/test/srt/run_suite.py
@@ -22,6 +22,8 @@ suites = {
        "test_json_constrained.py",
        "test_large_max_new_tokens.py",
        "test_metrics.py",
+        "test_mla.py",
+        "test_mla_fp8.py",
        "test_no_chunked_prefill.py",
        "test_no_overlap_scheduler.py",
        "test_openai_server.py",