From d554c7777b20dd33b2bee65377eefd5ac831539e Mon Sep 17 00:00:00 2001 From: Shixiaowei02 <39303645+Shixiaowei02@users.noreply.github.com> Date: Tue, 23 Jun 2026 23:50:19 -0700 Subject: [PATCH] [https://nvbugs/6274934][fix] Skip DeepSeek-V3-Lite bf16 conditional disagg worker test on non-Hopper test_workers_conditional_disaggregation_deepseek_v3_lite_bf16 runs DeepSeek-V3-Lite, which uses MLA. The context-phase MLA FMHA kernels (headSize=192, headSizeV=128) are only available on Hopper (SM90), so on pre-Hopper GPUs AttentionOp::initialize() aborts with "Deepseek should be supported by fmha in context part." Add @skip_no_hopper, matching the sibling DeepSeek-V3-Lite disaggregation tests (test_workers_kv_cache_aware_router_deepseek_v3_lite_bf16 and test_disaggregated_deepseek_v3_lite_bf16_conditional). Signed-off-by: Shixiaowei02 <39303645+Shixiaowei02@users.noreply.github.com> --- tests/integration/defs/disaggregated/test_workers.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/integration/defs/disaggregated/test_workers.py b/tests/integration/defs/disaggregated/test_workers.py index 35149391bf32..74db49e44287 100644 --- a/tests/integration/defs/disaggregated/test_workers.py +++ b/tests/integration/defs/disaggregated/test_workers.py @@ -653,6 +653,7 @@ def test_workers_conditional_disaggregation(disaggregated_test_root, asyncio.run(tester.test_multi_round_request(prompts)) +@skip_no_hopper @pytest.mark.parametrize("deepseek_v3_model_root", ['DeepSeek-V3-Lite-bf16'], indirect=True) def test_workers_conditional_disaggregation_deepseek_v3_lite_bf16(