[Chore] Clean up pytorch helper functions in vllm.utils (#26908)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2025-10-18 09:48:22 -07:00
committed by GitHub
parent 5c2acb270a
commit 6ac5e06f7c
119 changed files with 772 additions and 714 deletions
@@ -11,7 +11,7 @@ from tests.v1.attention.utils import full_cg_backend_configs as backend_configs
from vllm import LLM, SamplingParams
from vllm.config import CompilationConfig
from vllm.platforms import current_platform
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
@contextlib.contextmanager
@@ -20,7 +20,7 @@ from vllm.config import (
set_current_vllm_config,
)
from vllm.forward_context import BatchDescriptor, set_forward_context
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
# This import automatically registers `torch.ops.silly.attention`
from .. import silly_attention # noqa: F401
+1 -1
View File
@@ -19,7 +19,7 @@ from vllm.config import (
set_current_vllm_config,
)
from vllm.forward_context import BatchDescriptor, set_forward_context
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
# This import automatically registers `torch.ops.silly.attention`
from ..silly_attention import get_global_counter, reset_global_counter
+1 -1
View File
@@ -27,7 +27,7 @@ from vllm.config import (
set_current_vllm_config,
)
from vllm.forward_context import BatchDescriptor, set_forward_context
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
# This import automatically registers `torch.ops.silly.attention`
from .. import silly_attention # noqa: F401
+1 -1
View File
@@ -8,7 +8,7 @@ Centralizes custom operation definitions to avoid duplicate registrations.
import torch
from torch.library import Library
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
# Shared library for all compilation test operations
# Using "silly" namespace to match existing test expectations
+1 -1
View File
@@ -15,7 +15,7 @@ from vllm.config import (
set_current_vllm_config,
)
from vllm.forward_context import set_forward_context
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
def reference_fn(x: torch.Tensor):
+1 -1
View File
@@ -5,7 +5,7 @@ import dataclasses
import pytest
from vllm.config import CompilationMode
from vllm.utils import cuda_device_count_stateless
from vllm.utils.torch_utils import cuda_device_count_stateless
from ..utils import compare_all_settings
+1 -1
View File
@@ -8,7 +8,7 @@ from vllm.compilation.counter import compilation_counter
from vllm.compilation.fix_functionalization import FixFunctionalizationPass
from vllm.config import CompilationConfig, CUDAGraphMode, VllmConfig
from vllm.config.compilation import CompilationMode
from vllm.utils import _is_torch_equal_or_newer, is_torch_equal_or_newer
from vllm.utils.torch_utils import _is_torch_equal_or_newer, is_torch_equal_or_newer
def test_version():
+1 -1
View File
@@ -15,7 +15,7 @@ from vllm.config import (
set_current_vllm_config,
)
from vllm.forward_context import BatchDescriptor, set_forward_context
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
# This import automatically registers `torch.ops.silly.attention`
from . import silly_attention # noqa: F401
+1 -1
View File
@@ -12,7 +12,7 @@ from tests.quantization.utils import is_quant_method_supported
from vllm import LLM, SamplingParams
from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode, PassConfig
from vllm.platforms import current_platform
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
from ..utils import create_new_process_for_each_test
+1 -1
View File
@@ -15,8 +15,8 @@ from tests.v1.attention.utils import _Backend
from vllm import LLM, SamplingParams
from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode, PassConfig
from vllm.platforms import current_platform
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.flashinfer import has_flashinfer
from vllm.utils.torch_utils import is_torch_equal_or_newer
from ..utils import flat_product, multi_gpu_test
+1 -1
View File
@@ -60,8 +60,8 @@ from vllm.multimodal.utils import fetch_image
from vllm.outputs import RequestOutput
from vllm.sampling_params import BeamSearchParams
from vllm.transformers_utils.utils import maybe_model_redirect
from vllm.utils import set_default_torch_num_threads
from vllm.utils.collections import is_list_of
from vllm.utils.torch_utils import set_default_torch_num_threads
logger = init_logger(__name__)
+1 -1
View File
@@ -18,7 +18,7 @@ import pytest
from vllm.config.compilation import CompilationMode
from vllm.config.model import RunnerOption
from vllm.logger import init_logger
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
from ..models.registry import HF_EXAMPLE_MODELS
from ..utils import compare_two_settings, create_new_process_for_each_test
+1 -1
View File
@@ -11,10 +11,10 @@ import vllm.envs as envs
from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
from vllm.distributed.utils import StatelessProcessGroup
from vllm.utils import (
cuda_device_count_stateless,
get_open_port,
update_environment_variables,
)
from vllm.utils.torch_utils import cuda_device_count_stateless
from ..utils import multi_gpu_test
+4 -1
View File
@@ -3,7 +3,10 @@
import pytest
from vllm.utils import create_kv_caches_with_random, create_kv_caches_with_random_flash
from vllm.utils.torch_utils import (
create_kv_caches_with_random,
create_kv_caches_with_random_flash,
)
@pytest.fixture()
@@ -15,7 +15,7 @@ from tests.kernels.utils import make_alibi_bias
from vllm.attention.ops.chunked_prefill_paged_decode import chunked_prefill_paged_decode
from vllm.attention.ops.prefix_prefill import context_attention_fwd
from vllm.platforms import current_platform
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
NUM_HEADS = [64]
NUM_QUERIES_PER_KV = [1, 64]
+2 -1
View File
@@ -3,7 +3,8 @@
import pytest
import torch
from vllm.utils import get_cuda_view_from_cpu_tensor, is_uva_available
from vllm.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)]
@@ -13,8 +13,9 @@ import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.platforms import current_platform
from vllm.utils import cuda_device_count_stateless, has_deep_ep, has_deep_gemm, has_pplx
from vllm.utils import has_deep_ep, has_deep_gemm, has_pplx
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
from vllm.utils.torch_utils import cuda_device_count_stateless
from .modular_kernel_tools.common import (
Config,
+1 -1
View File
@@ -22,8 +22,8 @@ from vllm.utils import (
STR_BACKEND_ENV_VAR,
STR_FLASH_ATTN_VAL,
STR_XFORMERS_ATTN_VAL,
make_tensor_with_pad,
)
from vllm.utils.torch_utils import make_tensor_with_pad
# For now, disable "test_aot_dispatch_dynamic" since there are some
# bugs related to this test in PyTorch 2.4.
@@ -7,7 +7,7 @@ from huggingface_hub import snapshot_download
from transformers import AutoConfig, AutoModel, CLIPImageProcessor
from vllm.distributed import cleanup_dist_env_and_memory
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from ....conftest import ImageTestAssets
@@ -9,7 +9,7 @@ from transformers import AutoConfig, AutoModel, CLIPImageProcessor
from vllm.distributed import cleanup_dist_env_and_memory
from vllm.model_executor.models.radio import RadioModel
from vllm.transformers_utils.configs.radio import RadioConfig
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from ....conftest import ImageTestAssets
@@ -26,7 +26,6 @@ from vllm.distributed import (
init_distributed_environment,
initialize_model_parallel,
)
from vllm.model_executor.model_loader.utils import set_default_torch_dtype
from vllm.model_executor.models.interfaces import (
SupportsMultiModal,
supports_multimodal,
@@ -36,6 +35,7 @@ from vllm.multimodal.processing import BaseMultiModalProcessor, InputProcessingC
from vllm.multimodal.utils import group_mm_kwargs_by_modality
from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
from vllm.utils.collections import is_list_of
from vllm.utils.torch_utils import set_default_torch_dtype
from ...registry import _MULTIMODAL_EXAMPLE_MODELS, HF_EXAMPLE_MODELS
from ...utils import dummy_hf_overrides
+1 -1
View File
@@ -46,10 +46,10 @@ from vllm.platforms import current_platform
from vllm.transformers_utils.tokenizer import get_tokenizer
from vllm.utils import (
FlexibleArgumentParser,
cuda_device_count_stateless,
get_open_port,
)
from vllm.utils.mem_constants import GB_bytes
from vllm.utils.torch_utils import cuda_device_count_stateless
if current_platform.is_rocm():
from amdsmi import (
+6 -4
View File
@@ -24,11 +24,8 @@ from vllm.transformers_utils.detokenizer_utils import convert_ids_list_to_tokens
from vllm.utils import (
FlexibleArgumentParser,
bind_kv_cache,
common_broadcastable_dtype,
current_stream,
get_open_port,
get_tcp_uri,
is_lossless_cast,
join_host_port,
make_zmq_path,
make_zmq_socket,
@@ -37,6 +34,11 @@ from vllm.utils import (
split_zmq_path,
unique_filepath,
)
from vllm.utils.torch_utils import (
common_broadcastable_dtype,
current_stream,
is_lossless_cast,
)
from vllm.utils.mem_utils import MemorySnapshot, memory_profiling
from ..utils import create_new_process_for_each_test, flat_product
@@ -408,7 +410,7 @@ def test_bind_kv_cache_non_attention():
def test_bind_kv_cache_pp():
with patch("vllm.utils.cuda_device_count_stateless", lambda: 2):
with patch("vllm.utils.torch_utils.cuda_device_count_stateless", lambda: 2):
# this test runs with 1 GPU, but we simulate 2 GPUs
cfg = VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=2))
with set_current_vllm_config(cfg):
@@ -18,7 +18,8 @@ from tests.v1.attention.utils import (
from vllm.attention.backends.registry import _Backend
from vllm.config import ModelConfig
from vllm.platforms import current_platform
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, cdiv, is_torch_equal_or_newer
from vllm.utils import cdiv
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, is_torch_equal_or_newer
from vllm.v1.attention.backends.utils import (
CommonAttentionMetadata,
set_kv_cache_layout,
+2 -1
View File
@@ -22,7 +22,8 @@ from vllm import _custom_ops as ops
from vllm.attention.backends.registry import _Backend
from vllm.attention.ops.flashmla import is_flashmla_dense_supported
from vllm.config.vllm import set_current_vllm_config
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, cdiv
from vllm.utils import cdiv
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.attention.backends.utils import CommonAttentionMetadata
from vllm.v1.kv_cache_interface import FullAttentionSpec
+1 -1
View File
@@ -15,7 +15,7 @@ from vllm.inputs import PromptType
from vllm.outputs import RequestOutput
from vllm.platforms import current_platform
from vllm.sampling_params import RequestOutputKind
from vllm.utils import set_default_torch_num_threads
from vllm.utils.torch_utils import set_default_torch_num_threads
from vllm.v1.engine.async_llm import AsyncLLM
from vllm.v1.metrics.loggers import (
AggregatedLoggingStatLogger,
+1 -1
View File
@@ -12,7 +12,7 @@ from transformers import AutoTokenizer
from vllm import SamplingParams
from vllm.engine.arg_utils import EngineArgs
from vllm.platforms import current_platform
from vllm.utils import set_default_torch_num_threads
from vllm.utils.torch_utils import set_default_torch_num_threads
from vllm.v1.engine import EngineCoreRequest
from vllm.v1.engine.core import EngineCore
from vllm.v1.executor.abstract import Executor, UniProcExecutor
+1 -1
View File
@@ -21,7 +21,7 @@ from vllm.distributed.kv_events import BlockStored, KVEventBatch, ZmqEventPublis
from vllm.engine.arg_utils import EngineArgs
from vllm.platforms import current_platform
from vllm.usage.usage_lib import UsageContext
from vllm.utils import set_default_torch_num_threads
from vllm.utils.torch_utils import set_default_torch_num_threads
from vllm.v1.engine import EngineCoreRequest
from vllm.v1.engine.core import EngineCore
from vllm.v1.engine.core_client import AsyncMPClient, EngineCoreClient, SyncMPClient
+2 -1
View File
@@ -7,7 +7,8 @@ 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, make_tensor_with_pad
from vllm.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
from vllm.v1.sample.sampler import Sampler
+1 -1
View File
@@ -9,7 +9,7 @@ import regex as re
import torch
from vllm import CompletionOutput
from vllm.utils import make_tensor_with_pad
from vllm.utils.torch_utils import make_tensor_with_pad
from vllm.v1.sample.logits_processor import BatchUpdate, LogitsProcessor
from vllm.v1.sample.metadata import SamplingMetadata
+1 -1
View File
@@ -12,7 +12,7 @@ from tests.v1.shutdown.utils import (
from vllm import LLM, SamplingParams
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.sampling_params import RequestOutputKind
from vllm.utils import cuda_device_count_stateless
from vllm.utils.torch_utils import cuda_device_count_stateless
from vllm.v1.engine.async_llm import AsyncLLM
MODELS = ["hmellor/tiny-random-LlamaForCausalLM"]
+1 -1
View File
@@ -14,7 +14,7 @@ from tests.v1.shutdown.utils import (
from vllm import LLM, AsyncEngineArgs, SamplingParams
from vllm.distributed import get_tensor_model_parallel_rank
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.utils import cuda_device_count_stateless
from vllm.utils.torch_utils import cuda_device_count_stateless
from vllm.v1.engine.async_llm import AsyncLLM
from vllm.v1.engine.exceptions import EngineDeadError
+1 -1
View File
@@ -13,7 +13,7 @@ from vllm import LLM
from vllm.distributed import get_tensor_model_parallel_rank
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.utils import cuda_device_count_stateless
from vllm.utils.torch_utils import cuda_device_count_stateless
from vllm.v1.engine.async_llm import AsyncLLM
MODELS = ["hmellor/tiny-random-LlamaForCausalLM"]
+2 -1
View File
@@ -10,7 +10,8 @@ import torch
from vllm.platforms import current_platform
from vllm.sampling_params import SamplingParams
from vllm.utils import is_pin_memory_available, make_tensor_with_pad
from vllm.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
from vllm.v1.sample.metadata import SamplingMetadata