From 433c999dd061d6f72beab3d35005535c8c5435d7 Mon Sep 17 00:00:00 2001 From: pllimax Date: Mon, 14 Sep 2026 22:23:47 +0800 Subject: [PATCH] [NPU][CI] Fix sglang.test.ascend import failure in multi-node e2e pods (#39403) --- python/sglang/test/ascend/e2e/run_npu_testcase.sh | 4 ++++ .../npu/accuracy/glm5_2/test_npu_glm_5_2_w8a8_16p_gpqa.py | 8 ++++++-- .../npu/basic_function/dp_attn/test_npu_dp_attention.py | 6 +++++- 3 files changed, 15 insertions(+), 3 deletions(-) diff --git a/python/sglang/test/ascend/e2e/run_npu_testcase.sh b/python/sglang/test/ascend/e2e/run_npu_testcase.sh index 10ed4b3f5..acfce16f4 100644 --- a/python/sglang/test/ascend/e2e/run_npu_testcase.sh +++ b/python/sglang/test/ascend/e2e/run_npu_testcase.sh @@ -6,6 +6,10 @@ if [ ! -f "${sglang_source_path}/${test_case}" ];then exit 0 fi +# Leave the image WORKDIR (/sgl-workspace): the sglang repo cloned there can +# shadow the installed package on sys.path. +cd "${sglang_source_path}" + echo "NPU info:" npu-smi info diff --git a/test/registered/npu/accuracy/glm5_2/test_npu_glm_5_2_w8a8_16p_gpqa.py b/test/registered/npu/accuracy/glm5_2/test_npu_glm_5_2_w8a8_16p_gpqa.py index 749741b1f..0ec5a6b91 100644 --- a/test/registered/npu/accuracy/glm5_2/test_npu_glm_5_2_w8a8_16p_gpqa.py +++ b/test/registered/npu/accuracy/glm5_2/test_npu_glm_5_2_w8a8_16p_gpqa.py @@ -83,6 +83,8 @@ GLM_5_2_W8A8_16P_TWO_NODE_OTHER_ARGS = [ 1, "--speculative-num-draft-tokens", 4, + "--speculative-draft-model-quantization", + "unquant", ] GLM_5_2_W8A8_16P_TWO_NODE_MODEL_CONFIG = { @@ -103,9 +105,11 @@ class TestNPUGLM_5_2_W8A8_16P_GPQA(TestNpuAccuracyMultiNodePdMixTestCaseBase): # generation_config = {"max_tokens": 131072, "temperature": 1.0} eval_batch_size = 32 generation_config = { - "max_tokens": 65536, + "max_tokens": 131072, + "top_p": 0.95, "temperature": 1.0, - "timeout": 1200, + "timeout": 7200, + "retries": 2, "stream": True, } diff --git a/test/registered/npu/basic_function/dp_attn/test_npu_dp_attention.py b/test/registered/npu/basic_function/dp_attn/test_npu_dp_attention.py index dc3489181..960fc4206 100644 --- a/test/registered/npu/basic_function/dp_attn/test_npu_dp_attention.py +++ b/test/registered/npu/basic_function/dp_attn/test_npu_dp_attention.py @@ -23,7 +23,11 @@ from sglang.test.test_utils import ( popen_launch_server, ) -register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3") +register_npu_ci( + est_time=400, + suite="base-b-test-4-npu-a3", + disabled="Temporarily disable this test case due to DTS2026091464350", +) register_npu_ci(est_time=2400, suite="nightly-4-npu-a3", nightly=True)