diff --git a/scripts/run_ci_metax.sh b/scripts/run_ci_metax.sh index 9e18f00b41..b8c86ae248 100644 --- a/scripts/run_ci_metax.sh +++ b/scripts/run_ci_metax.sh @@ -91,7 +91,6 @@ METAX_CI_CASELIST=( "tests/operators/test_update_inputs_v1.py" "tests/operators/test_set_value_by_flags_and_idx.py" "tests/operators/test_get_token_penalty_multi_scores.py" - "tests/operators/test_speculate_get_token_penalty_multi_scores.py" "tests/operators/test_token_penalty.py" "tests/operators/test_stop_generation_multi_ends.py" "tests/operators/test_get_padding_offset.py" diff --git a/tests/metax_ci/test_ernie_21b.py b/tests/metax_ci/test_ernie_21b.py index 970906fa2e..9e4375fa7f 100644 --- a/tests/metax_ci/test_ernie_21b.py +++ b/tests/metax_ci/test_ernie_21b.py @@ -29,9 +29,9 @@ class TestErnie21B(unittest.TestCase): max_model_len=256, quantization="wint8", load_choices="default_v1", - enable_prefix_caching=False, + # enable_prefix_caching=False, disable_custom_all_reduce=True, - graph_optimization_config={"use_cudagraph": False, "graph_opt_level": 0}, + # graph_optimization_config={"use_cudagraph": False, "graph_opt_level": 0}, ) cls.sampling_params = fastdeploy.SamplingParams(top_p=0.95, max_tokens=256, temperature=0.6) diff --git a/tests/metax_ci/test_ernie_28b_vl.py b/tests/metax_ci/test_ernie_28b_vl.py index 4c09433238..1aafa77026 100644 --- a/tests/metax_ci/test_ernie_28b_vl.py +++ b/tests/metax_ci/test_ernie_28b_vl.py @@ -36,8 +36,8 @@ class TestErnie28BVL(unittest.TestCase): max_model_len=32768, quantization="wint8", disable_custom_all_reduce=True, - enable_prefix_caching=False, - graph_optimization_config={"use_cudagraph": False, "graph_opt_level": 0}, + # enable_prefix_caching=False, + # graph_optimization_config={"use_cudagraph": False, "graph_opt_level": 0}, limit_mm_per_prompt={"image": 100}, reasoning_parser="ernie-45-vl", load_choices="default_v1",