diff --git a/.buildkite/test-amd.yaml b/.buildkite/test-amd.yaml index 97281445af..3d04922a76 100644 --- a/.buildkite/test-amd.yaml +++ b/.buildkite/test-amd.yaml @@ -320,18 +320,6 @@ steps: #----------------------------------------------------- mi250 · models / language -----------------------------------------------------# -- label: ":amd: (MI250) Language Models (MTEB)" # TBD - timeout_in_minutes: 180 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250] - agent_pool: mi250_1 - optional: true - working_dir: "/vllm-workspace/tests" - source_file_dependencies: - - vllm/ - - tests/models/language/pooling_mteb_test - commands: - - pytest -v -s models/language/pooling_mteb_test - - label: ":amd: (MI250) Language Models (PPL)" # TBD timeout_in_minutes: 180 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250] @@ -1961,6 +1949,19 @@ steps: #----------------------------------------------------- mi300 · models / language -----------------------------------------------------# +- label: ":amd: (MI300) Language Models (MTEB)" # TBD + timeout_in_minutes: 180 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300] + dind: false + agent_pool: mi300_1 + optional: true + working_dir: "/vllm-workspace/tests" + source_file_dependencies: + - vllm/ + - tests/models/language/pooling_mteb_test + commands: + - pytest -v -s models/language/pooling_mteb_test + - label: ":amd: (MI300) Language Models (Extended Pooling)" # TBD timeout_in_minutes: 180 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300] diff --git a/tests/models/language/pooling_mteb_test/test_modernbert_fp8.py b/tests/models/language/pooling_mteb_test/test_modernbert_fp8.py index 51e9061b9b..8e062f0d76 100644 --- a/tests/models/language/pooling_mteb_test/test_modernbert_fp8.py +++ b/tests/models/language/pooling_mteb_test/test_modernbert_fp8.py @@ -5,9 +5,27 @@ import pytest from tests.models.utils import EmbedModelInfo from tests.quantization.utils import is_quant_method_supported +from vllm.platforms import current_platform from .mteb_embed_utils import mteb_test_embed_models + +def _fp8_scaled_mm_unsupported() -> bool: + """Whether the ModernBERT online FP8 test should be skipped. + + ``is_quant_method_supported("fp8")`` returns True on MI250 (gfx90a), but + serving a per-tensor FP8 linear layer on ROCm requires CDNA3+ or RDNA4, + so treat it as unsupported here. + """ + if not is_quant_method_supported("fp8"): + return True + if current_platform.is_rocm(): + from vllm.platforms.rocm import get_cdna_version, on_gfx12x + + return get_cdna_version() <= 2 and not on_gfx12x() + return False + + MODEL_INFO = EmbedModelInfo( "Alibaba-NLP/gte-modernbert-base", mteb_score=0.748193353, @@ -39,8 +57,8 @@ def _assert_modernbert_online_fp8(model) -> None: @pytest.mark.skipif( - not is_quant_method_supported("fp8"), - reason="FP8 is not supported on this GPU type.", + _fp8_scaled_mm_unsupported(), + reason="No FP8 ScaledMM kernel is available on this GPU type.", ) def test_modernbert_online_fp8_mteb(hf_runner, vllm_runner, monkeypatch) -> None: monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")