A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 15:00 UTC
[CI/Build][AMD] Skip tests in test_fusions_e2e and test_dbo_dp_ep_gsm8k that require non-existing imports for ROCm (#30417)
Signed-off-by: Randall Smith <ransmith@amd.com> Co-authored-by: Randall Smith <ransmith@amd.com>
rasmith committed
Dec 11, 2025 at 18:24 UTC
48661d275fb44b969112a7bd8586dfd9f498e2e3
2 files changed
+27
-1
tests/compile/distributed/test_fusions_e2e.py
+25
-1
@@ -138,6 +138,17 @@ elif current_platform.is_rocm():
138
CUSTOM_OPS_FP8 = ["-quant_fp8", "+quant_fp8"]
139
140
141
+def has_cuda_graph_wrapper_metadata() -> bool:
142
+ from importlib import import_module
143
+
144
+ try:
145
+ module = import_module("torch._inductor.utils")
146
+ module.CUDAGraphWrapperMetadata # noqa B018
147
+ except AttributeError:
148
+ return False
149
+ return True
150
+
151
+
152
@pytest.mark.parametrize(
153
"model_name, model_kwargs, backend, matches, custom_ops",
154
# Test attention+quant_fp8 fusion with custom and torch impls of QuantFP8
@@ -145,7 +156,20 @@ CUSTOM_OPS_FP8 = ["-quant_fp8", "+quant_fp8"]
156
# quant_fp4 only has the custom impl
157
+ list(flat_product(MODELS_FP4, [""])),
158
)
148
-@pytest.mark.parametrize("inductor_graph_partition", [True, False])
159
+@pytest.mark.parametrize(
160
+ "inductor_graph_partition",
161
+ [
162
+ pytest.param(
163
+ True,
164
+ marks=pytest.mark.skipif(
165
+ not has_cuda_graph_wrapper_metadata(),
166
+ reason="This test requires"
167
+ "torch._inductor.utils.CUDAGraphWrapperMetadata to run",
168
+ ),
169
+ ),
170
+ False,
171
+ ],
172
+)
173
def test_attn_quant(
174
model_name: str,
175
model_kwargs: dict[str, Any],
tests/v1/distributed/test_dbo.py
+2
@@ -13,6 +13,7 @@ import torch
13
14
from tests.evals.gsm8k.gsm8k_eval import evaluate_gsm8k
15
from tests.utils import RemoteOpenAIServer
16
+from vllm.utils.import_utils import has_deep_ep
17
18
# Detect Blackwell / B200 (compute capability 10.x)
19
try:
@@ -44,6 +45,7 @@ DEEPEP_BACKENDS = [
45
]
46
47
48
+@pytest.mark.skipif(not has_deep_ep(), reason="These tests require deep_ep to run")
49
@pytest.mark.parametrize("all2all_backend", DEEPEP_BACKENDS)
50
@pytest.mark.xfail(
51
IS_BLACKWELL,