[CI]Enable mxfp4 lora test for ROCm platform (#43802)
Signed-off-by: Qiang Li <qiang.li2@amd.com>
qli88 committed
Jun 18, 2026 at 11:59 UTC
25faa1f4cc2ec5d0db50b2b2b04c43f58d8a0931
1 file changed
+25
-11
tests/lora/test_gptoss_tp.py
+25
-11
@@ -70,17 +70,20 @@ def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
70
assert generated_texts[i].startswith(EXPECTED_LORA_OUTPUT[i])
71
72
73
-@pytest.mark.skipif(
74
- not current_platform.is_cuda(),
75
- reason=(
76
- "Mxfp4 LoRA on ROCm is blocked by a spawn compatibility issue. "
77
- "The fused_moe_lora Triton kernel crashes in spawned subprocesses, "
78
- "and vLLM forces spawn mode when HIP is initialized before "
79
- "multiprocessing. Fixing this requires either making the LoRA "
80
- "Triton kernel spawn-safe or pre-warming the kernel cache."
81
- ),
73
+# TODO: make the Mxfp4MoeBackend.TRITON spawn-safe.
74
+# For now just use TRITON_UNFUSED kernel
75
+@pytest.mark.parametrize(
76
+ "mxfp4_use_marlin",
77
+ [
78
+ False,
79
+ pytest.param(
80
+ True,
81
+ marks=pytest.mark.skipif(
82
+ current_platform.is_rocm(), reason="marlin not supported"
83
+ ),
84
+ ),
85
+ ],
86
)
83
-@pytest.mark.parametrize("mxfp4_use_marlin", [True, False])
87
@pytest.mark.parametrize("specialize_active_lora", [True, False])
88
def test_gpt_oss_lora(
89
gptoss20b_lora_files,
@@ -109,7 +112,18 @@ def test_gpt_oss_lora(
112
113
@multi_gpu_test(num_gpus=2)
114
@pytest.mark.parametrize("fully_sharded_loras", [False, True])
112
-@pytest.mark.parametrize("mxfp4_use_marlin", [True, False])
115
+@pytest.mark.parametrize(
116
+ "mxfp4_use_marlin",
117
+ [
118
+ False,
119
+ pytest.param(
120
+ True,
121
+ marks=pytest.mark.skipif(
122
+ current_platform.is_rocm(), reason="marlin not supported"
123
+ ),
124
+ ),
125
+ ],
126
+)
127
def test_gpt_oss_lora_tp2(
128
gptoss20b_lora_files,
129
fully_sharded_loras,