[Chore] Separate out vllm.utils.platform_utils.py (#27374)

Signed-off-by: Jonathan <chenleejonathan@gmail.com>

tests/kernels/core/test_uva.py

@@ -3,7 +3,7 @@
 import pytest
 import torch
 
-from vllm.utils import is_uva_available
+from vllm.utils.platform_utils import is_uva_available
 from vllm.utils.torch_utils import get_cuda_view_from_cpu_tensor
 
 CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]

tests/v1/logits_processors/test_correctness.py

@@ -21,7 +21,7 @@ from tests.v1.sample.utils import (
 from vllm.config import VllmConfig
 from vllm.platforms import current_platform
 from vllm.sampling_params import SamplingParams
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.sample.logits_processor import (
     BatchUpdate,
     BatchUpdateBuilder,

tests/v1/sample/test_sampler.py

@@ -7,7 +7,7 @@ import torch
 
 from tests.v1.sample.utils import create_allowed_token_ids
 from vllm.platforms import current_platform
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.torch_utils import make_tensor_with_pad
 from vllm.v1.sample.logits_processor import LogitsProcessors
 from vllm.v1.sample.metadata import SamplingMetadata

tests/v1/worker/test_gpu_input_batch.py

@@ -10,7 +10,7 @@ import torch
 
 from vllm.platforms import current_platform
 from vllm.sampling_params import SamplingParams
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.torch_utils import make_tensor_with_pad
 from vllm.v1.pool.metadata import PoolingMetadata
 from vllm.v1.sample.logits_processor import LogitsProcessors

vllm/device_allocator/cumem.py

@@ -18,7 +18,7 @@ from typing import Any
 import torch
 
 from vllm.logger import init_logger
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 
 logger = init_logger(__name__)
 

vllm/lora/lora_weights.py

@@ -8,7 +8,7 @@ import torch
 import torch.types
 
 from vllm.lora.peft_helper import PEFTHelper
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 
 
 class LoRALayerWeights:

vllm/lora/models.py

@@ -31,8 +31,8 @@ from vllm.model_executor.models import SupportsLoRA, supports_multimodal
 from vllm.model_executor.models.interfaces import is_pooling_model
 from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.model_executor.models.utils import PPMissingLayer, WeightsMapper
-from vllm.utils import is_pin_memory_available
 from vllm.utils.cache import LRUCache
+from vllm.utils.platform_utils import is_pin_memory_available
 
 logger = init_logger(__name__)
 

vllm/model_executor/model_loader/utils.py

@@ -26,7 +26,7 @@ from vllm.model_executor.models.adapters import (
     try_create_mm_pooling_model_cls,
 )
 from vllm.model_executor.models.interfaces import SupportsQuant, supports_multimodal
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 
 logger = init_logger(__name__)
 

vllm/model_executor/models/qwen2_5_vl.py

@@ -72,7 +72,7 @@ from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
 from vllm.multimodal.parse import MultiModalDataItems
 from vllm.multimodal.processing import PromptReplacement, PromptUpdate
 from vllm.sequence import IntermediateTensors
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
 from .interfaces import (

vllm/model_executor/models/utils.py

@@ -22,8 +22,8 @@ from vllm.logger import init_logger
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.multimodal import NestedTensors
 from vllm.sequence import IntermediateTensors
-from vllm.utils import (
-    cdiv,
+from vllm.utils import cdiv
+from vllm.utils.platform_utils import (
     is_pin_memory_available,
     is_uva_available,
 )

vllm/usage/usage_lib.py

@@ -21,7 +21,7 @@ import torch
 import vllm.envs as envs
 from vllm.connections import global_http_connection
 from vllm.logger import init_logger
-from vllm.utils import cuda_get_device_properties
+from vllm.utils.platform_utils import cuda_get_device_properties
 from vllm.utils.torch_utils import cuda_device_count_stateless
 from vllm.version import __version__ as VLLM_VERSION
 

vllm/utils/__init__.py

@@ -27,12 +27,8 @@ from argparse import (
     _ArgumentGroup,
 )
 from collections import defaultdict
-from collections.abc import (
-    Callable,
-    Sequence,
-)
-from concurrent.futures.process import ProcessPoolExecutor
-from functools import cache, partial, wraps
+from collections.abc import Callable
+from functools import partial, wraps
 from typing import TYPE_CHECKING, Any, TypeVar
 
 import cloudpickle
@@ -44,6 +40,7 @@ import yaml
 import vllm.envs as envs
 from vllm.logger import enable_trace_function_call, init_logger
 from vllm.ray.lazy_utils import is_in_ray_actor
+from vllm.utils.platform_utils import cuda_is_initialized, xpu_is_initialized
 
 _DEPRECATED_MAPPINGS = {
     "cprofile": "profiling",
@@ -169,21 +166,6 @@ def round_down(x: int, y: int) -> int:
     return (x // y) * y
 
 
-@cache
-def is_pin_memory_available() -> bool:
-    from vllm.platforms import current_platform
-
-    return current_platform.is_pin_memory_available()
-
-
-@cache
-def is_uva_available() -> bool:
-    """Check if Unified Virtual Addressing (UVA) is available."""
-    # UVA requires pinned memory.
-    # TODO: Add more requirements for UVA if needed.
-    return is_pin_memory_available()
-
-
 # TODO: This function can be removed if transformer_modules classes are
 # serialized by value when communicating between processes
 def init_cached_hf_modules() -> None:
@@ -216,35 +198,6 @@ def enable_trace_function_call_for_thread(vllm_config: VllmConfig) -> None:
         enable_trace_function_call(log_path)
 
 
-def cuda_is_initialized() -> bool:
-    """Check if CUDA is initialized."""
-    if not torch.cuda._is_compiled():
-        return False
-    return torch.cuda.is_initialized()
-
-
-def xpu_is_initialized() -> bool:
-    """Check if XPU is initialized."""
-    if not torch.xpu._is_compiled():
-        return False
-    return torch.xpu.is_initialized()
-
-
-def cuda_get_device_properties(
-    device, names: Sequence[str], init_cuda=False
-) -> tuple[Any, ...]:
-    """Get specified CUDA device property values without initializing CUDA in
-    the current process."""
-    if init_cuda or cuda_is_initialized():
-        props = torch.cuda.get_device_properties(device)
-        return tuple(getattr(props, name) for name in names)
-
-    # Run in subprocess to avoid initializing CUDA as a side effect.
-    mp_ctx = multiprocessing.get_context("fork")
-    with ProcessPoolExecutor(max_workers=1, mp_context=mp_ctx) as executor:
-        return executor.submit(cuda_get_device_properties, device, names, True).result()
-
-
 def weak_bind(
     bound_method: Callable[..., Any],
 ) -> Callable[..., None]:

vllm/utils/platform_utils.py

@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import multiprocessing
+from collections.abc import Sequence
+from concurrent.futures.process import ProcessPoolExecutor
+from functools import cache
+from typing import Any
+
+import torch
+
+
+def cuda_is_initialized() -> bool:
+    """Check if CUDA is initialized."""
+    if not torch.cuda._is_compiled():
+        return False
+    return torch.cuda.is_initialized()
+
+
+def xpu_is_initialized() -> bool:
+    """Check if XPU is initialized."""
+    if not torch.xpu._is_compiled():
+        return False
+    return torch.xpu.is_initialized()
+
+
+def cuda_get_device_properties(
+    device, names: Sequence[str], init_cuda=False
+) -> tuple[Any, ...]:
+    """Get specified CUDA device property values without initializing CUDA in
+    the current process."""
+    if init_cuda or cuda_is_initialized():
+        props = torch.cuda.get_device_properties(device)
+        return tuple(getattr(props, name) for name in names)
+
+    # Run in subprocess to avoid initializing CUDA as a side effect.
+    mp_ctx = multiprocessing.get_context("fork")
+    with ProcessPoolExecutor(max_workers=1, mp_context=mp_ctx) as executor:
+        return executor.submit(cuda_get_device_properties, device, names, True).result()
+
+
+@cache
+def is_pin_memory_available() -> bool:
+    from vllm.platforms import current_platform
+
+    return current_platform.is_pin_memory_available()
+
+
+@cache
+def is_uva_available() -> bool:
+    """Check if Unified Virtual Addressing (UVA) is available."""
+    # UVA requires pinned memory.
+    # TODO: Add more requirements for UVA if needed.
+    return is_pin_memory_available()

vllm/v1/attention/backends/flashinfer.py

@@ -34,12 +34,13 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
 )
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
-from vllm.utils import cdiv, is_pin_memory_available
+from vllm.utils import cdiv
 from vllm.utils.flashinfer import (
     can_use_trtllm_attention,
     flashinfer_disable_q_quantization,
     use_trtllm_attention,
 )
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.attention.backends.utils import (
     AttentionCGSupport,
     AttentionMetadataBuilder,

vllm/v1/kv_offload/worker/cpu_gpu.py

@@ -7,7 +7,7 @@ import torch
 from vllm import _custom_ops as ops
 from vllm.attention import AttentionBackend
 from vllm.logger import init_logger
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
 from vllm.v1.kv_offload.worker.worker import (
     OffloadingHandler,

vllm/v1/pool/metadata.py

@@ -5,7 +5,7 @@ from dataclasses import dataclass
 import torch
 
 from vllm.pooling_params import PoolingParams
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 
 pin_memory = is_pin_memory_available()
 

vllm/v1/sample/ops/penalties.py

@@ -4,7 +4,7 @@
 import torch
 
 from vllm.model_executor.layers.utils import apply_penalties
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.torch_utils import make_tensor_with_pad
 
 

vllm/v1/sample/sampler.py

@@ -6,7 +6,7 @@ import torch
 import torch.nn as nn
 
 from vllm.config.model import LogprobsMode
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.outputs import LogprobsTensors, SamplerOutput
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.sample.ops.bad_words import apply_bad_words

vllm/v1/spec_decode/eagle.py

@@ -24,7 +24,7 @@ from vllm.model_executor.models.deepseek_v2 import DeepseekV32IndexerCache
 from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.platforms import current_platform
-from vllm.utils import is_pin_memory_available
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
 from vllm.v1.attention.backends.tree_attn import (
     TreeAttentionMetadata,

vllm/v1/worker/gpu_model_runner.py

@@ -72,13 +72,13 @@ from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
 from vllm.utils import (
     cdiv,
     check_use_alibi,
-    is_pin_memory_available,
     length_from_prompt_token_ids_or_embeds,
     round_up,
 )
 from vllm.utils.jsontree import json_map_leaves
 from vllm.utils.mem_constants import GiB_bytes
 from vllm.utils.mem_utils import DeviceMemoryProfiler
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.utils.torch_utils import (
     get_dtype_size,
     kv_cache_dtype_str_to_dtype,

vllm/v1/worker/tpu_model_runner.py

@@ -53,7 +53,8 @@ from vllm.multimodal.inputs import (
 from vllm.multimodal.utils import group_mm_kwargs_by_modality
 from vllm.sequence import IntermediateTensors
 from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
-from vllm.utils import LayerBlockType, cdiv, is_pin_memory_available, prev_power_of_2
+from vllm.utils import LayerBlockType, cdiv, prev_power_of_2
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.attention.backends.pallas import (
     TPU_STR_DTYPE_TO_TORCH_DTYPE,
     PallasAttentionBackend,