[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
@@ -10,7 +10,7 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8,
)
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
def flashinfer_fused_moe_blockscale_fp8(
@@ -52,8 +52,8 @@ from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import OCP_MX_Sc
from vllm.model_executor.utils import maybe_disable_graph_partition
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils import direct_register_custom_op, is_torch_equal_or_newer
from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used
from vllm.utils.torch_utils import direct_register_custom_op, is_torch_equal_or_newer
from .rocm_aiter_fused_moe import is_rocm_aiter_moe_enabled
@@ -52,8 +52,9 @@ from vllm.model_executor.layers.quantization.base_config import (
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.platforms.interface import CpuArchEnum
from vllm.utils import cdiv, direct_register_custom_op, has_deep_ep, has_pplx, round_up
from vllm.utils import cdiv, has_deep_ep, has_pplx, round_up
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.worker.ubatching import dbo_current_ubatch_id
if current_platform.is_cuda_alike():
@@ -11,7 +11,7 @@ from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
)
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
class QuantMethod(IntEnum):
@@ -23,8 +23,9 @@ from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
mxfp8_e4m3_quantize,
)
from vllm.triton_utils import tl, triton
from vllm.utils import cdiv, is_torch_equal_or_newer
from vllm.utils import cdiv
from vllm.utils.flashinfer import flashinfer_fp4_quantize
from vllm.utils.torch_utils import is_torch_equal_or_newer
@triton.jit
+1 -1
View File
@@ -13,7 +13,7 @@ from vllm.model_executor.layers.batch_invariant import (
vllm_is_batch_invariant,
)
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
def is_rocm_aiter_rmsnorm_enabled() -> bool:
@@ -34,7 +34,7 @@ from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateShapeCalculator,
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.linear_attn import LinearAttentionMetadata
if TYPE_CHECKING:
@@ -37,7 +37,7 @@ from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
selective_state_update,
)
from vllm.model_executor.utils import set_weight_attrs
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.mamba1_attn import Mamba1AttentionMetadata
@@ -46,7 +46,7 @@ from vllm.model_executor.model_loader.weight_utils import (
sharded_weight_loader,
)
from vllm.model_executor.utils import set_weight_attrs
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadata
# Added by the IBM Team, 2024
@@ -6,7 +6,10 @@ import torch
from vllm.config.cache import MambaDType
from vllm.config.model import ModelDType
from vllm.distributed import divide
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, get_kv_cache_torch_dtype
from vllm.utils.torch_utils import (
STR_DTYPE_TO_TORCH_DTYPE,
get_kv_cache_torch_dtype,
)
class MambaStateDtypeCalculator:
@@ -27,7 +27,7 @@ from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
causal_conv1d_fn,
causal_conv1d_update,
)
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionMetadata
@@ -23,7 +23,7 @@ from vllm.model_executor.layers.quantization import (
QuantizationMethods,
)
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
class BitsAndBytesConfig(QuantizationConfig):
@@ -24,7 +24,7 @@ from vllm.model_executor.layers.quantization.base_config import QuantizationConf
from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
class FPQuantConfig(QuantizationConfig):
@@ -28,7 +28,7 @@ from vllm.model_executor.layers.quantization.base_config import (
)
from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
from vllm.model_executor.utils import set_weight_attrs
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
logger = init_logger(__name__)
@@ -7,7 +7,7 @@ import torch
import vllm.envs as envs
from vllm import _custom_ops as ops
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
from .cutlass import CutlassScaledMMLinearKernel
from .ScaledMMLinearKernel import ScaledMMLinearLayerConfig
@@ -49,10 +49,10 @@ from vllm.platforms import current_platform
from vllm.scalar_type import scalar_types
from vllm.utils import (
has_triton_kernels,
is_torch_equal_or_newer,
round_up,
)
from vllm.utils.flashinfer import has_flashinfer
from vllm.utils.torch_utils import is_torch_equal_or_newer
logger = init_logger(__name__)
@@ -45,7 +45,7 @@ try:
from aiter.ops.triton.gemm_afp4wfp4 import gemm_afp4wfp4
from aiter.ops.triton.quant import dynamic_mxfp4_quant
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
if is_rocm_aiter_fp4_asm_gemm_enabled():
from aiter import gemm_a4w4, per_1x32_f4_quant_hip
@@ -28,13 +28,13 @@ from vllm.model_executor.parameter import (
)
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils import direct_register_custom_op
from vllm.utils.deep_gemm import (
fp8_gemm_nt,
is_deep_gemm_e8m0_used,
is_deep_gemm_supported,
should_use_deepgemm_for_fp8_linear,
)
from vllm.utils.torch_utils import direct_register_custom_op
logger = init_logger(__name__)
@@ -7,7 +7,7 @@ import torch
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op, is_torch_equal_or_newer
from vllm.utils.torch_utils import direct_register_custom_op, is_torch_equal_or_newer
logger = init_logger(__name__)
@@ -3,7 +3,7 @@
import torch
from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import OCP_MX_BLOCK_SIZE
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
def _quant_dequant_mxfp6(
@@ -12,8 +12,8 @@ from vllm.config import CompilationMode, get_current_vllm_config
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.flashinfer import flashinfer_scaled_fp8_mm, has_flashinfer
from vllm.utils.torch_utils import direct_register_custom_op
# Input scaling factors are no longer optional in _scaled_mm starting
# from pytorch 2.5. Allocating a dummy tensor to pass as input_scale
@@ -10,7 +10,7 @@ import torch
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
if current_platform.is_cuda():
from vllm.vllm_flash_attn.layers.rotary import apply_rotary_emb
@@ -5,7 +5,7 @@ import torch
import vllm.envs as envs
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
def is_rocm_triton_rotary_embedding_enabled() -> bool:
+1 -1
View File
@@ -9,7 +9,7 @@ import torch
from vllm import _custom_ops as ops
from vllm import envs
from vllm.platforms import CpuArchEnum, current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
def shuffle_weight(w: torch.Tensor) -> torch.Tensor:
@@ -11,8 +11,8 @@ from vllm.logger import init_logger
from vllm.model_executor.model_loader.utils import (
initialize_model,
process_weights_after_loading,
set_default_torch_dtype,
)
from vllm.utils.torch_utils import set_default_torch_dtype
logger = init_logger(__name__)
@@ -32,7 +32,7 @@ from vllm.model_executor.layers.linear import (
RowParallelLinear,
)
from vllm.model_executor.model_loader.base_loader import BaseModelLoader
from vllm.model_executor.model_loader.utils import ParamMapping, set_default_torch_dtype
from vllm.model_executor.model_loader.utils import ParamMapping
from vllm.model_executor.model_loader.weight_utils import (
download_safetensors_index_file_from_hf,
download_weights_from_hf,
@@ -48,6 +48,7 @@ from vllm.model_executor.utils import (
set_weight_attrs,
)
from vllm.platforms import current_platform
from vllm.utils.torch_utils import set_default_torch_dtype
logger = init_logger(__name__)
@@ -15,13 +15,13 @@ from vllm.model_executor.model_loader.base_loader import BaseModelLoader
from vllm.model_executor.model_loader.utils import (
initialize_model,
process_weights_after_loading,
set_default_torch_dtype,
)
from vllm.model_executor.model_loader.weight_utils import (
get_gguf_extra_tensor_names,
get_gguf_weight_type_map,
gguf_quant_weights_iterator,
)
from vllm.utils.torch_utils import set_default_torch_dtype
class GGUFModelLoader(BaseModelLoader):
@@ -22,8 +22,8 @@ from vllm.model_executor.model_loader.tensorizer import (
from vllm.model_executor.model_loader.utils import (
get_model_architecture,
initialize_model,
set_default_torch_dtype,
)
from vllm.utils.torch_utils import set_default_torch_dtype
logger = init_logger(__name__)
+1 -1
View File
@@ -14,8 +14,8 @@ from vllm.model_executor.model_loader.default_loader import DefaultModelLoader
from vllm.model_executor.model_loader.utils import (
initialize_model,
process_weights_after_loading,
set_default_torch_dtype,
)
from vllm.utils.torch_utils import set_default_torch_dtype
logger = init_logger(__name__)
-10
View File
@@ -2,7 +2,6 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Utilities for selecting and loading models."""
import contextlib
import inspect
import warnings
from contextlib import contextmanager
@@ -32,15 +31,6 @@ from vllm.utils import is_pin_memory_available
logger = init_logger(__name__)
@contextlib.contextmanager
def set_default_torch_dtype(dtype: torch.dtype):
"""Sets the default torch dtype to the given dtype."""
old_dtype = torch.get_default_dtype()
torch.set_default_dtype(dtype)
yield
torch.set_default_dtype(old_dtype)
def initialize_model(
vllm_config: VllmConfig,
*,
+2 -1
View File
@@ -6,7 +6,8 @@ from typing import TYPE_CHECKING
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.model_executor.models import ModelRegistry
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, cdiv, round_up
from vllm.utils import cdiv, round_up
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.kv_cache_interface import FullAttentionSpec, MambaSpec
if TYPE_CHECKING:
+1 -1
View File
@@ -79,8 +79,8 @@ from vllm.model_executor.model_loader.weight_utils import (
from vllm.model_executor.models.utils import sequence_parallel_chunk
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.utils import direct_register_custom_op
from vllm.utils.deep_gemm import fp8_mqa_logits, fp8_paged_mqa_logits
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.mla.indexer import (
DeepseekV32IndexerBackend,
DeepseekV32IndexerMetadata,
+1 -1
View File
@@ -18,7 +18,6 @@ from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.model_loader.utils import set_default_torch_dtype
from vllm.model_executor.models.transformers.utils import replace_linear_class
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (
@@ -51,6 +50,7 @@ from vllm.transformers_utils.processors.deepseek_vl2 import DeepseekVLV2Processo
from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
from vllm.utils.collections import is_list_of
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from vllm.utils.torch_utils import set_default_torch_dtype
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
from .utils import (
+1 -1
View File
@@ -51,8 +51,8 @@ from vllm.multimodal.processing import (
from vllm.multimodal.profiling import BaseDummyInputsBuilder
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.utils import set_default_torch_num_threads
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from vllm.utils.torch_utils import set_default_torch_num_threads
from .interfaces import (
MultiModalEmbeddings,
+1 -1
View File
@@ -49,7 +49,6 @@ from vllm.model_executor.layers.resampler import (
Resampler2,
get_2d_sincos_pos_embed,
)
from vllm.model_executor.model_loader.utils import set_default_torch_dtype
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.model_executor.models.minicpm import MiniCPMForCausalLM
from vllm.model_executor.models.module_mapping import MultiModelKeys
@@ -88,6 +87,7 @@ from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.utils.collections import flatten_2d_lists
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from vllm.utils.torch_utils import set_default_torch_dtype
from .idefics2_vision_model import Idefics2VisionTransformer
from .interfaces import (
+1 -1
View File
@@ -64,7 +64,7 @@ from vllm.model_executor.models.utils import (
)
from vllm.model_executor.utils import set_weight_attrs
from vllm.sequence import IntermediateTensors
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadata
+1 -1
View File
@@ -71,7 +71,7 @@ from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs import Qwen3NextConfig
from vllm.triton_utils import tl, triton
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
from .interfaces import (
@@ -29,7 +29,7 @@ from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.models.interfaces import MixtureOfExperts
from vllm.model_executor.models.utils import maybe_prefix
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
from vllm.utils.torch_utils import direct_register_custom_op
from .utils import log_replacement
+4 -2
View File
@@ -24,11 +24,13 @@ from vllm.multimodal import NestedTensors
from vllm.sequence import IntermediateTensors
from vllm.utils import (
cdiv,
direct_register_custom_op,
get_cuda_view_from_cpu_tensor,
is_pin_memory_available,
is_uva_available,
)
from vllm.utils.torch_utils import (
direct_register_custom_op,
get_cuda_view_from_cpu_tensor,
)
logger = init_logger(__name__)
+1 -1
View File
@@ -34,7 +34,6 @@ from vllm.model_executor.layers.linear import (
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.model_loader.utils import set_default_torch_dtype
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (
@@ -53,6 +52,7 @@ from vllm.multimodal.profiling import BaseDummyInputsBuilder
from vllm.transformers_utils.processor import cached_get_processor
from vllm.utils.jsontree import json_map_leaves
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from vllm.utils.torch_utils import set_default_torch_dtype
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsTranscription
from .utils import (
+1 -1
View File
@@ -7,7 +7,7 @@ from typing import Any
import torch
from vllm.utils import is_torch_equal_or_newer
from vllm.utils.torch_utils import is_torch_equal_or_newer
def set_random_seed(seed: int) -> None: