diff --git a/tests/evals/gsm8k/test_gsm8k_correctness.py b/tests/evals/gsm8k/test_gsm8k_correctness.py index ee40c65853..d14f41843b 100644 --- a/tests/evals/gsm8k/test_gsm8k_correctness.py +++ b/tests/evals/gsm8k/test_gsm8k_correctness.py @@ -9,16 +9,30 @@ pytest -s -v tests/evals/gsm8k/test_gsm8k_correctness.py \ --config-list-file=configs/models-small.txt """ +import importlib.metadata import shlex +from importlib.util import find_spec import pytest +import torch import yaml +from packaging import version from tests.utils import RemoteOpenAIServer from vllm.platforms import current_platform from .gsm8k_eval import evaluate_gsm8k +# MXFP4 via quark requires amd-quark >= 0.12 on torch >= 2.11. +# Earlier torch releases work with older quark versions. See +# https://github.com/amd/Quark/issues/34 +# TODO: Remove once amd-quark>=0.12.0 +QUARK_MXFP4_TORCH_COMPATIBLE = find_spec("quark") is not None and ( + version.parse(importlib.metadata.version("amd-quark")) >= version.parse("0.12.0") + if version.parse(torch.__version__.split("+")[0]) >= version.parse("2.11") + else True +) + def run_gsm8k_eval(eval_config: dict, server_url: str) -> dict: """Run GSM8K evaluation using our isolated script.""" @@ -88,16 +102,19 @@ def test_gsm8k_correctness(config_filename): "Skipping DeepSeek-V3.2 and DeepSeek-R1 on ROCm platforms " "due to agent pool disk space issues and pod evictions." ) - if current_platform.is_rocm() and ( - "Qwen3.5-35B-A3B-MXFP4-AITER-TP2" in config_filename.name - ): + if current_platform.is_rocm() and ("Qwen3.5-35B-A3B-MXFP4" in config_filename.name): from vllm.platforms.rocm import on_gfx950 - if not on_gfx950(): + if not on_gfx950() and "AITER-TP2" in config_filename.name: pytest.skip( "Skipping Qwen3.5-35B-A3B-MXFP4-AITER-TP2 on non-GFX950 platforms. " "The quantization scheme is not supported on non-GFX950 platforms." ) + if not QUARK_MXFP4_TORCH_COMPATIBLE: + pytest.skip( + "Skipping Qwen3.5-35B-A3B-MXFP4: amd-quark >= 0.12 is required " + "on torch >= 2.11." + ) # Parse server arguments from config (use shlex to handle quoted strings) server_args_str = eval_config.get("server_args", "") server_args = shlex.split(server_args_str) if server_args_str else [] diff --git a/tests/kernels/moe/test_ocp_mx_moe.py b/tests/kernels/moe/test_ocp_mx_moe.py index fb5ae527b2..e768947b26 100644 --- a/tests/kernels/moe/test_ocp_mx_moe.py +++ b/tests/kernels/moe/test_ocp_mx_moe.py @@ -13,9 +13,15 @@ from vllm._aiter_ops import is_aiter_found from vllm.platforms import current_platform from vllm.utils.flashinfer import has_flashinfer -QUARK_MXFP4_AVAILABLE = find_spec("quark") is not None and version.parse( - importlib.metadata.version("amd-quark") -) >= version.parse("0.8.99") +# MXFP4 via quark requires amd-quark >= 0.12 on torch >= 2.11. +# Earlier torch releases work with older quark versions. See +# https://github.com/amd/Quark/issues/34 +# TODO: Remove once amd-quark>=0.12.0 +QUARK_MXFP4_TORCH_COMPATIBLE = find_spec("quark") is not None and ( + version.parse(importlib.metadata.version("amd-quark")) >= version.parse("0.12.0") + if version.parse(torch.__version__.split("+")[0]) >= version.parse("2.11") + else True +) TRTLLM_GEN_MXFP4_AVAILABLE = ( current_platform.is_cuda() and current_platform.is_device_capability_family(100) @@ -87,7 +93,10 @@ def enable_pickle(monkeypatch): ModelCase("fxmarty/Llama-3.1-70B-Instruct-2-layers-mxfp6", tp=4), ], ) -@pytest.mark.skipif(not QUARK_MXFP4_AVAILABLE, reason="amd-quark>=0.9 is not available") +@pytest.mark.skipif( + not QUARK_MXFP4_TORCH_COMPATIBLE, + reason="MXFP4 via quark requires amd-quark >= 0.12 on torch >= 2.11.", +) def test_mxfp4_loading_and_execution_moe(vllm_runner, model_case: ModelCase): if torch.accelerator.device_count() < model_case.tp: pytest.skip(