[Bugfix][CI] Prevent common ops imports from initializing CUDA (#50639)
Signed-off-by: Andreas Karatzas <akaratza@amd.com> Co-authored-by: OpenAI Codex <noreply@openai.com>
Andreas Karatzas committed
Jul 31, 2026 at 23:13 UTC
f7097a92fd5374ff8a44167b7bc16eec42393559
4 files changed
+3
-5
vllm/models/common/ops/__init__.py
-2
@@ -2,10 +2,8 @@
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""Ops shared across model implementations."""
4
5
-from .fused_allreduce_rms_norm import fused_allreduce_rms_norm
5
from .fused_qk_rmsnorm import fused_q_kv_rmsnorm
6
7
__all__ = [
9
- "fused_allreduce_rms_norm",
8
"fused_q_kv_rmsnorm",
9
]
vllm/models/deepseek_v32/amd/model.py
+1
-1
@@ -34,7 +34,7 @@ from vllm.model_executor.models.utils import (
34
make_empty_intermediate_tensors_factory,
35
make_layers,
36
)
37
-from vllm.models.common.ops import fused_allreduce_rms_norm
37
+from vllm.models.common.ops.fused_allreduce_rms_norm import fused_allreduce_rms_norm
38
from vllm.sequence import IntermediateTensors
39
40
from .rocm import DeepseekV32MLAAttention
vllm/models/deepseek_v32/nvidia/model.py
+1
-1
@@ -38,7 +38,7 @@ from vllm.model_executor.models.utils import (
38
make_layers,
39
sequence_parallel_chunk,
40
)
41
-from vllm.models.common.ops import fused_allreduce_rms_norm
41
+from vllm.models.common.ops.fused_allreduce_rms_norm import fused_allreduce_rms_norm
42
from vllm.models.deepseek_v32.attention import DeepseekV32Attention
43
from vllm.sequence import IntermediateTensors
44
vllm/models/kimi_k3/nvidia/dspark_mla.py
+1
-1
@@ -20,7 +20,7 @@ from vllm.model_executor.models.utils import (
20
get_draft_quant_config,
21
maybe_prefix,
22
)
23
-from vllm.models.common.ops import fused_allreduce_rms_norm
23
+from vllm.models.common.ops.fused_allreduce_rms_norm import fused_allreduce_rms_norm
24
from vllm.models.kimi_k3.nvidia.mla import MultiHeadLatentAttention
25
from vllm.models.kimi_k3.nvidia.model import KimiMLP
26
from vllm.utils.torch_utils import is_quantized_kv_cache