[Hardware][CPU] using current_platform.is_cpu (#9536)

tests/conftest.py

@@ -32,9 +32,10 @@ from vllm.inputs import (ExplicitEncoderDecoderPrompt, TextPrompt,
                          to_enc_dec_tuple_list, zip_enc_dec_prompts)
 from vllm.logger import init_logger
 from vllm.outputs import RequestOutput
+from vllm.platforms import current_platform
 from vllm.sampling_params import BeamSearchParams
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, cuda_device_count_stateless,
-                        identity, is_cpu)
+                        identity)
 
 logger = init_logger(__name__)
 
@@ -236,7 +237,8 @@ class HfRunner:
 
     def wrap_device(self, input: _T, device: Optional[str] = None) -> _T:
         if device is None:
-            return self.wrap_device(input, "cpu" if is_cpu() else "cuda")
+            return self.wrap_device(
+                input, "cpu" if current_platform.is_cpu() else "cuda")
 
         if hasattr(input, "device") and input.device.type == device:
             return input

tests/encoder_decoder/test_e2e_correctness.py

@@ -7,8 +7,8 @@ from typing import List, Optional, Tuple
 import pytest
 from transformers import AutoModelForSeq2SeqLM
 
+from vllm.platforms import current_platform
 from vllm.sequence import SampleLogprobs
-from vllm.utils import is_cpu
 
 from ..conftest import DecoderPromptType
 from ..models.utils import check_logprobs_close
@@ -35,7 +35,7 @@ def vllm_to_hf_output(
 @pytest.mark.parametrize("decoder_prompt_type", list(DecoderPromptType))
 @pytest.mark.parametrize("enforce_eager", [True, False])
 @pytest.mark.skipif(
-    is_cpu(),
+    current_platform.is_cpu(),
     reason="CPU backend is not currently supported with encoder/decoder models"
 )
 def test_encoder_decoder_e2e(

tests/kernels/test_attention_selector.py

@@ -19,7 +19,8 @@ def test_env(name: str, device: str, monkeypatch):
     override_backend_env_variable(monkeypatch, name)
 
     if device == "cpu":
-        with patch("vllm.attention.selector.is_cpu", return_value=True):
+        with patch("vllm.attention.selector.current_platform.is_cpu",
+                   return_value=True):
             backend = which_attn_to_use(16, torch.float16, torch.float16, 16,
                                         False)
         assert backend.name == "TORCH_SDPA"

tests/models/decoder_only/language/test_phimoe.py

@@ -5,7 +5,7 @@ Run `pytest tests/models/test_phimoe.py`.
 import pytest
 import torch
 
-from vllm.utils import is_cpu
+from vllm.platforms import current_platform
 
 from ....utils import large_gpu_test
 from ...utils import check_logprobs_close
@@ -70,7 +70,7 @@ def test_phimoe_routing_function():
         assert torch.equal(topk_ids, ground_truth[test_id]["topk_ids"])
 
 
-@pytest.mark.skipif(condition=is_cpu(),
+@pytest.mark.skipif(condition=current_platform.is_cpu(),
                     reason="This test takes a lot time to run on CPU, "
                     "and vllm CI's disk space is not enough for this model.")
 @large_gpu_test(min_gb=80)

tests/models/decoder_only/vision_language/test_fuyu.py

@@ -3,8 +3,8 @@ from typing import List, Optional, Tuple, Type
 import pytest
 
 from vllm.multimodal.utils import rescale_image_size
+from vllm.platforms import current_platform
 from vllm.sequence import SampleLogprobs
-from vllm.utils import is_cpu
 
 from ....conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
 from ...utils import check_logprobs_close
@@ -103,7 +103,7 @@ def run_test(
 
 
 target_dtype = "half"
-if is_cpu():
+if current_platform.is_cpu():
     target_dtype = "bfloat16"
 
 

tests/models/decoder_only/vision_language/test_internvl.py

@@ -7,7 +7,7 @@ from PIL.Image import Image
 from transformers import AutoConfig
 
 from vllm.multimodal.utils import rescale_image_size
-from vllm.utils import is_cpu
+from vllm.platforms import current_platform
 
 from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
                           _ImageAssets)
@@ -244,7 +244,7 @@ def run_awq_test(
 
 
 target_dtype = "half"
-if is_cpu():
+if current_platform.is_cpu():
     target_dtype = "bfloat16"
 
 

tests/models/decoder_only/vision_language/test_phi3v.py

@@ -10,8 +10,9 @@ from vllm.inputs import InputContext, token_inputs
 from vllm.model_executor.models.phi3v import _IMAGE_TOKEN_ID
 from vllm.multimodal import MultiModalRegistry
 from vllm.multimodal.utils import rescale_image_size
+from vllm.platforms import current_platform
 from vllm.sequence import SampleLogprobs
-from vllm.utils import is_cpu, is_hip
+from vllm.utils import is_hip
 
 from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
                           _ImageAssets)
@@ -49,7 +50,7 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
 
 
 target_dtype = "half"
-if is_cpu():
+if current_platform.is_cpu():
     target_dtype = "bfloat16"
 
 # ROCm Triton FA can run into shared memory issues with these models,

tests/models/utils.py

@@ -5,8 +5,8 @@ import torch
 
 from vllm.config import ModelConfig, TaskOption
 from vllm.inputs import InputContext
+from vllm.platforms import current_platform
 from vllm.sequence import Logprob, PromptLogprobs, SampleLogprobs
-from vllm.utils import is_cpu
 
 TokensText = Tuple[List[int], str]
 
@@ -270,7 +270,7 @@ def build_model_context(model_name: str,
     if tokenizer_name is None:
         tokenizer_name = model_name
     if dtype is None:
-        dtype = "bfloat16" if is_cpu() else "half"
+        dtype = "bfloat16" if current_platform.is_cpu() else "half"
 
     model_config = ModelConfig(
         model_name,

tests/worker/test_encoder_decoder_model_runner.py

@@ -5,8 +5,9 @@ import pytest
 import torch
 
 from vllm.engine.arg_utils import EngineArgs
+from vllm.platforms import current_platform
 from vllm.sequence import SamplingParams, SequenceData, SequenceGroupMetadata
-from vllm.utils import is_cpu, make_tensor_with_pad
+from vllm.utils import make_tensor_with_pad
 from vllm.worker.enc_dec_model_runner import EncoderDecoderModelRunner
 from vllm.worker.model_runner import _get_graph_batch_size
 
@@ -31,7 +32,7 @@ def _create_model_runner(model: str, *args,
     return model_runner
 
 
-@pytest.mark.skipif(condition=is_cpu(),
+@pytest.mark.skipif(condition=current_platform.is_cpu(),
                     reason="CPU backend is currently "
                     "unsupported for encoder/ "
                     "decoder models")
@@ -74,7 +75,7 @@ def test_empty_seq_group():
     assert return_seq_lens is None
 
 
-@pytest.mark.skipif(condition=is_cpu(),
+@pytest.mark.skipif(condition=current_platform.is_cpu(),
                     reason="CPU backend is currently "
                     "unsupported for encoder/ "
                     "decoder models")
@@ -264,7 +265,7 @@ def test_prepare_prompt(batch_size):
     assert torch.equal(actual, expected)
 
 
-@pytest.mark.skipif(condition=is_cpu(),
+@pytest.mark.skipif(condition=current_platform.is_cpu(),
                     reason="CPU backend is currently "
                     "unsupported for encoder/ "
                     "decoder models")

vllm/attention/backends/torch_sdpa.py

@@ -10,9 +10,9 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionMetadata, AttentionType)
 from vllm.attention.backends.utils import CommonAttentionState
 from vllm.attention.ops.paged_attn import PagedAttentionMetadata
-from vllm.utils import is_cpu
+from vllm.platforms import current_platform
 
-if is_cpu():
+if current_platform.is_cpu():
     try:
         from vllm.attention.ops.ipex_attn import PagedAttention
     except ImportError:

vllm/attention/ops/blocksparse_attention/interface.py

@@ -3,7 +3,7 @@ import math
 import torch
 
 from vllm.platforms import current_platform
-from vllm.utils import is_cpu, is_hip
+from vllm.utils import is_hip
 
 from .utils import (dense_to_crow_col, get_head_sliding_step,
                     get_sparse_attn_mask)
@@ -32,7 +32,7 @@ class LocalStridedBlockSparseAttn(torch.nn.Module):
     ):
         super().__init__()
         if use_spda is None:
-            use_spda = is_hip() or is_cpu() or not \
+            use_spda = is_hip() or current_platform.is_cpu() or not \
                        IS_COMPUTE_8_OR_ABOVE
         device = device or (torch.cuda.current_device()
                             if current_platform.is_cuda_alike() else "cpu")

vllm/attention/selector.py

@@ -10,7 +10,7 @@ import vllm.envs as envs
 from vllm.attention.backends.abstract import AttentionBackend
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import STR_BACKEND_ENV_VAR, is_cpu, is_hip, is_openvino, is_xpu
+from vllm.utils import STR_BACKEND_ENV_VAR, is_hip, is_openvino, is_xpu
 
 logger = init_logger(__name__)
 
@@ -121,7 +121,7 @@ def get_attn_backend(
             ROCmFlashAttentionBackend)
         return ROCmFlashAttentionBackend
     elif backend == _Backend.TORCH_SDPA:
-        assert is_cpu(), RuntimeError(
+        assert current_platform.is_cpu(), RuntimeError(
             "Torch SDPA backend is only used for the CPU device.")
         logger.info("Using Torch SDPA backend.")
         from vllm.attention.backends.torch_sdpa import TorchSDPABackend
@@ -183,7 +183,7 @@ def which_attn_to_use(
         if backend_by_env_var is not None:
             selected_backend = backend_name_to_enum(backend_by_env_var)
 
-    if is_cpu():
+    if current_platform.is_cpu():
         if selected_backend != _Backend.TORCH_SDPA:
             logger.info("Cannot use %s backend on CPU.", selected_backend)
         return _Backend.TORCH_SDPA

vllm/distributed/parallel_state.py

@@ -37,7 +37,7 @@ from torch.distributed import Backend, ProcessGroup
 import vllm.envs as envs
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import is_cpu, supports_custom_op
+from vllm.utils import supports_custom_op
 
 
 @dataclass
@@ -1139,7 +1139,7 @@ def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
         import ray  # Lazy import Ray
         ray.shutdown()
     gc.collect()
-    if not is_cpu():
+    if not current_platform.is_cpu():
         torch.cuda.empty_cache()
 
 

vllm/model_executor/custom_op.py

@@ -7,7 +7,7 @@ import vllm.envs as envs
 from vllm.compilation.levels import CompilationLevel
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import is_cpu, is_hip, is_xpu, print_warning_once
+from vllm.utils import is_hip, is_xpu, print_warning_once
 
 logger = init_logger(__name__)
 
@@ -74,7 +74,7 @@ class CustomOp(nn.Module):
 
         if is_hip():
             return self.forward_hip
-        elif is_cpu():
+        elif current_platform.is_cpu():
             return self.forward_cpu
         elif current_platform.is_tpu():
             return self.forward_tpu

vllm/model_executor/models/utils.py

@@ -21,7 +21,7 @@ from vllm.model_executor.models import ModelRegistry
 from vllm.multimodal.base import NestedTensors
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
-from vllm.utils import is_cpu, is_pin_memory_available
+from vllm.utils import is_pin_memory_available
 
 logger = init_logger(__name__)
 
@@ -515,7 +515,7 @@ def get_vit_attn_backend() -> _Backend:
                     "so we use xformers backend instead. You can run "
                     "`pip install flash-attn` to use flash-attention backend.")
                 selected_backend = _Backend.XFORMERS
-        elif is_cpu():
+        elif current_platform.is_cpu():
             selected_backend = _Backend.TORCH_SDPA
         else:
             selected_backend = _Backend.XFORMERS

vllm/utils.py

@@ -318,15 +318,6 @@ def is_hip() -> bool:
     return torch.version.hip is not None
 
 
-@lru_cache(maxsize=None)
-def is_cpu() -> bool:
-    from importlib.metadata import PackageNotFoundError, version
-    try:
-        return "cpu" in version("vllm")
-    except PackageNotFoundError:
-        return False
-
-
 @lru_cache(maxsize=None)
 def is_openvino() -> bool:
     from importlib.metadata import PackageNotFoundError, version
@@ -798,7 +789,7 @@ def is_pin_memory_available() -> bool:
     elif is_neuron():
         print_warning_once("Pin memory is not supported on Neuron.")
         return False
-    elif is_cpu() or is_openvino():
+    elif current_platform.is_cpu() or is_openvino():
         return False
     return True