[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>

vllm/models/common/ops/__init__.py

@@ -2,10 +2,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Ops shared across model implementations."""
 
-from .fused_allreduce_rms_norm import fused_allreduce_rms_norm
 from .fused_qk_rmsnorm import fused_q_kv_rmsnorm
 
 __all__ = [
-    "fused_allreduce_rms_norm",
     "fused_q_kv_rmsnorm",
 ]

vllm/models/deepseek_v32/amd/model.py

@@ -34,7 +34,7 @@ from vllm.model_executor.models.utils import (
     make_empty_intermediate_tensors_factory,
     make_layers,
 )
-from vllm.models.common.ops import fused_allreduce_rms_norm
+from vllm.models.common.ops.fused_allreduce_rms_norm import fused_allreduce_rms_norm
 from vllm.sequence import IntermediateTensors
 
 from .rocm import DeepseekV32MLAAttention

vllm/models/deepseek_v32/nvidia/model.py

@@ -38,7 +38,7 @@ from vllm.model_executor.models.utils import (
     make_layers,
     sequence_parallel_chunk,
 )
-from vllm.models.common.ops import fused_allreduce_rms_norm
+from vllm.models.common.ops.fused_allreduce_rms_norm import fused_allreduce_rms_norm
 from vllm.models.deepseek_v32.attention import DeepseekV32Attention
 from vllm.sequence import IntermediateTensors
 

vllm/models/kimi_k3/nvidia/dspark_mla.py

@@ -20,7 +20,7 @@ from vllm.model_executor.models.utils import (
     get_draft_quant_config,
     maybe_prefix,
 )
-from vllm.models.common.ops import fused_allreduce_rms_norm
+from vllm.models.common.ops.fused_allreduce_rms_norm import fused_allreduce_rms_norm
 from vllm.models.kimi_k3.nvidia.mla import MultiHeadLatentAttention
 from vllm.models.kimi_k3.nvidia.model import KimiMLP
 from vllm.utils.torch_utils import is_quantized_kv_cache