[CI]Enable mxfp4 lora test for ROCm platform (#43802)

Signed-off-by: Qiang Li <qiang.li2@amd.com>
This commit is contained in:
qli88
2026-06-18 16:59:09 +00:00
committed by GitHub
parent 4583630b56
commit 25faa1f4cc
+25 -11
View File
@@ -70,17 +70,20 @@ def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
assert generated_texts[i].startswith(EXPECTED_LORA_OUTPUT[i]) assert generated_texts[i].startswith(EXPECTED_LORA_OUTPUT[i])
@pytest.mark.skipif( # TODO: make the Mxfp4MoeBackend.TRITON spawn-safe.
not current_platform.is_cuda(), # For now just use TRITON_UNFUSED kernel
reason=( @pytest.mark.parametrize(
"Mxfp4 LoRA on ROCm is blocked by a spawn compatibility issue. " "mxfp4_use_marlin",
"The fused_moe_lora Triton kernel crashes in spawned subprocesses, " [
"and vLLM forces spawn mode when HIP is initialized before " False,
"multiprocessing. Fixing this requires either making the LoRA " pytest.param(
"Triton kernel spawn-safe or pre-warming the kernel cache." True,
), marks=pytest.mark.skipif(
current_platform.is_rocm(), reason="marlin not supported"
),
),
],
) )
@pytest.mark.parametrize("mxfp4_use_marlin", [True, False])
@pytest.mark.parametrize("specialize_active_lora", [True, False]) @pytest.mark.parametrize("specialize_active_lora", [True, False])
def test_gpt_oss_lora( def test_gpt_oss_lora(
gptoss20b_lora_files, gptoss20b_lora_files,
@@ -109,7 +112,18 @@ def test_gpt_oss_lora(
@multi_gpu_test(num_gpus=2) @multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize("fully_sharded_loras", [False, True]) @pytest.mark.parametrize("fully_sharded_loras", [False, True])
@pytest.mark.parametrize("mxfp4_use_marlin", [True, False]) @pytest.mark.parametrize(
"mxfp4_use_marlin",
[
False,
pytest.param(
True,
marks=pytest.mark.skipif(
current_platform.is_rocm(), reason="marlin not supported"
),
),
],
)
def test_gpt_oss_lora_tp2( def test_gpt_oss_lora_tp2(
gptoss20b_lora_files, gptoss20b_lora_files,
fully_sharded_loras, fully_sharded_loras,