[Chore]:Extract math and argparse utilities to separate modules (#27188)

Signed-off-by: Yeshwanth Surya <yeshsurya@gmail.com>
Signed-off-by: Yeshwanth N <yeshsurya@gmail.com>
Signed-off-by: yeshsurya <yeshsurya@gmail.com>

benchmarks/benchmark_block_pool.py

@@ -5,7 +5,7 @@ import gc
 from benchmark_utils import TimeCollector
 from tabulate import tabulate
 
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.v1.core.block_pool import BlockPool
 
 

benchmarks/benchmark_long_document_qa_throughput.py

@@ -46,7 +46,7 @@ import time
 
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def test_long_document_qa(llm=None, sampling_params=None, prompts=None):

benchmarks/benchmark_ngram_proposer.py

@@ -19,7 +19,7 @@ from vllm.config import (
     VllmConfig,
 )
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.v1.spec_decode.ngram_proposer import NgramProposer
 from vllm.v1.worker.gpu_input_batch import InputBatch
 from vllm.v1.worker.gpu_model_runner import GPUModelRunner

benchmarks/benchmark_prefix_caching.py

@@ -37,7 +37,7 @@ from transformers import PreTrainedTokenizerBase
 
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 try:
     from vllm.transformers_utils.tokenizer import get_tokenizer

benchmarks/benchmark_prioritization.py

@@ -11,7 +11,7 @@ import time
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
 
 from vllm.engine.arg_utils import EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 # Select a equi-probable random priority

benchmarks/benchmark_serving_structured_output.py

@@ -51,7 +51,7 @@ except ImportError:
     from backend_request_func import get_tokenizer
 
 try:
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -15,7 +15,7 @@ from utils import make_rand_sparse_tensors
 from weight_shapes import WEIGHT_SHAPES
 
 from vllm import _custom_ops as ops
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -18,7 +18,8 @@ from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     w8a8_triton_block_scaled_mm,
 )
-from vllm.utils import FlexibleArgumentParser, cdiv
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils.math_utils import cdiv
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -10,7 +10,7 @@ import torch
 from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
 from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
 from vllm.triton_utils import triton
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
 
 

benchmarks/kernels/benchmark_activation.py

@@ -10,7 +10,7 @@ import vllm.model_executor.layers.activation  # noqa F401
 from vllm.model_executor.custom_op import CustomOp
 from vllm.platforms import current_platform
 from vllm.triton_utils import triton
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
 
 batch_size_range = [1, 16, 32, 64, 128]

benchmarks/kernels/benchmark_bitblas.py

@@ -28,7 +28,7 @@ except ImportError as e:
 
 from bitblas import Matmul, MatmulConfig, auto_detect_nvidia_target
 
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 parser = FlexibleArgumentParser(
     description="Benchmark BitBLAS int4 on a specific target."

benchmarks/kernels/benchmark_cutlass_fp4_moe.py

@@ -20,7 +20,7 @@ from vllm.model_executor.layers.fused_moe.config import (
 from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
 from vllm.scalar_type import scalar_types
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 WEIGHT_SHAPES_MOE = {
     "nvidia/DeepSeek-R1-FP4": [

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -14,7 +14,7 @@ from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_confi
 from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 # Weight shapes for different models: [num_experts, topk, hidden_size,
 # intermediate_size]

benchmarks/kernels/benchmark_device_communicators.py

@@ -39,7 +39,7 @@ from vllm.distributed.device_communicators.pynccl_allocator import (
 )
 from vllm.distributed.device_communicators.symm_mem import SymmMemCommunicator
 from vllm.logger import init_logger
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 logger = init_logger(__name__)
 

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -13,7 +13,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
     fused_experts,
     fused_topk,
 )
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = [
     "nm-testing/Mixtral-8x7B-Instruct-v0.1",

benchmarks/kernels/benchmark_layernorm.py

@@ -7,7 +7,7 @@ import torch
 
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
 
 

benchmarks/kernels/benchmark_lora.py

@@ -25,7 +25,7 @@ if HAS_TRITON:
     from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink
     from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
 
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_TP_SIZES = [1]

benchmarks/kernels/benchmark_machete.py

@@ -33,7 +33,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
     quantize_weights,
 )
 from vllm.scalar_type import ScalarType, scalar_types
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-3-8b", "meta-llama/Llama-2-70b-hf"]
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024]

benchmarks/kernels/benchmark_marlin.py

@@ -44,7 +44,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
     sort_weights,
 )
 from vllm.scalar_type import ScalarType, scalar_types
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192]

benchmarks/kernels/benchmark_moe.py

@@ -22,7 +22,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
 from vllm.transformers_utils.config import get_config
 from vllm.triton_utils import triton
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 FP8_DTYPE = current_platform.fp8_dtype()
 

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -17,7 +17,7 @@ from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import (
 )
 from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 FP8_DTYPE = current_platform.fp8_dtype()
 

benchmarks/kernels/benchmark_mrope.py

@@ -39,7 +39,7 @@ import torch
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.platforms import current_platform
 from vllm.transformers_utils.config import get_config
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 

benchmarks/kernels/benchmark_paged_attention.py

@@ -9,7 +9,7 @@ import torch
 from vllm import _custom_ops as ops
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import (
     STR_DTYPE_TO_TORCH_DTYPE,
     create_kv_caches_with_random,

benchmarks/kernels/benchmark_quant.py

@@ -7,7 +7,7 @@ import torch
 
 from vllm import _custom_ops as ops
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
 
 

benchmarks/kernels/benchmark_reshape_and_cache.py

@@ -9,7 +9,7 @@ from tabulate import tabulate
 from vllm import _custom_ops as ops
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import (
     STR_DTYPE_TO_TORCH_DTYPE,
     create_kv_caches_with_random,

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

@@ -12,7 +12,7 @@ from vllm.attention.ops.triton_reshape_and_cache_flash import (
 )
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import (
     STR_DTYPE_TO_TORCH_DTYPE,
     create_kv_caches_with_random_flash,

benchmarks/kernels/benchmark_rope.py

@@ -8,7 +8,7 @@ import torch
 
 from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def benchmark_rope_kernels_multi_lora(

benchmarks/kernels/benchmark_trtllm_decode_attention.py

@@ -8,7 +8,7 @@ from datetime import datetime
 import flashinfer
 import torch
 
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
 FP8_DTYPE = torch.float8_e4m3fn

benchmarks/kernels/benchmark_trtllm_prefill_attention.py

@@ -8,7 +8,7 @@ from datetime import datetime
 import flashinfer
 import torch
 
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
 FP8_DTYPE = torch.float8_e4m3fn

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -18,7 +18,7 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
 )
 from vllm.platforms import current_platform
 from vllm.triton_utils import triton
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 mp.set_start_method("spawn", force=True)
 

benchmarks/kernels/graph_machete_bench.py

@@ -11,7 +11,7 @@ import regex as re
 import seaborn as sns
 from torch.utils.benchmark import Measurement as TMeasurement
 
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(

benchmarks/overheads/benchmark_hashing.py

@@ -5,7 +5,7 @@ import cProfile
 import pstats
 
 from vllm import LLM, SamplingParams
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 # A very long prompt, total number of tokens is about 15k.
 LONG_PROMPT = ["You are an expert in large language models, aren't you?"] * 1000

examples/offline_inference/audio_language.py

@@ -18,7 +18,7 @@ from transformers import AutoTokenizer
 from vllm import LLM, EngineArgs, SamplingParams
 from vllm.assets.audio import AudioAsset
 from vllm.lora.request import LoRARequest
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 audio_assets = [AudioAsset("mary_had_lamb"), AudioAsset("winning_call")]
 question_per_audio_count = {

examples/offline_inference/basic/chat.py

@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def create_parser():

examples/offline_inference/basic/classify.py

@@ -4,7 +4,7 @@
 from argparse import Namespace
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/basic/embed.py

@@ -4,7 +4,7 @@
 from argparse import Namespace
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/basic/generate.py

@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def create_parser():

examples/offline_inference/basic/reward.py

@@ -4,7 +4,7 @@
 from argparse import Namespace
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/basic/score.py

@@ -4,7 +4,7 @@
 from argparse import Namespace
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/encoder_decoder_multimodal.py

@@ -13,7 +13,7 @@ from typing import NamedTuple
 
 from vllm import LLM, EngineArgs, PromptType, SamplingParams
 from vllm.assets.audio import AudioAsset
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 class ModelRequestData(NamedTuple):

examples/offline_inference/llm_engine_example.py

@@ -8,7 +8,7 @@ for processing prompts with various sampling parameters.
 import argparse
 
 from vllm import EngineArgs, LLMEngine, RequestOutput, SamplingParams
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def create_test_prompts() -> list[tuple[str, SamplingParams]]:

examples/offline_inference/load_sharded_state.py

@@ -25,7 +25,7 @@ python load_sharded_state.py \
 import dataclasses
 
 from vllm import LLM, EngineArgs, SamplingParams
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/pooling/embed_jina_embeddings_v3.py

@@ -4,7 +4,7 @@
 from argparse import Namespace
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/pooling/embed_matryoshka_fy.py

@@ -4,7 +4,7 @@
 from argparse import Namespace
 
 from vllm import LLM, EngineArgs, PoolingParams
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/pooling/multi_vector_retrieval.py

@@ -4,7 +4,7 @@
 from argparse import Namespace
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/pooling/ner.py

@@ -5,7 +5,7 @@
 from argparse import Namespace
 
 from vllm import LLM, EngineArgs
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/profiling_tpu/profiling.py

@@ -13,7 +13,7 @@ from tqdm import tqdm
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
 from vllm.inputs import PromptType
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 DURATION_MS = int(os.getenv("VLLM_TPU_PROFILE_DURATION_MS", 3000))
 DELAY_MS = int(os.getenv("VLLM_TPU_PROFILE_DELAY_MS", 0))

examples/offline_inference/qwen2_5_omni/only_thinker.py

@@ -13,7 +13,7 @@ from vllm.assets.audio import AudioAsset
 from vllm.assets.image import ImageAsset
 from vllm.assets.video import VideoAsset
 from vllm.multimodal.image import convert_image_mode
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 class QueryResult(NamedTuple):

examples/offline_inference/save_sharded_state.py

@@ -30,7 +30,7 @@ from pathlib import Path
 
 from vllm import LLM, EngineArgs
 from vllm.model_executor.model_loader import ShardedStateLoader
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def parse_args():

examples/offline_inference/spec_decode.py

@@ -9,7 +9,7 @@ from vllm.inputs import TokensPrompt
 from vllm.v1.metrics.reader import Counter, Vector
 
 try:
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 

examples/offline_inference/vision_language.py

@@ -22,7 +22,7 @@ from vllm.assets.image import ImageAsset
 from vllm.assets.video import VideoAsset
 from vllm.lora.request import LoRARequest
 from vllm.multimodal.image import convert_image_mode
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 class ModelRequestData(NamedTuple):

examples/offline_inference/vision_language_multi_image.py

@@ -18,7 +18,7 @@ from transformers import AutoProcessor, AutoTokenizer
 from vllm import LLM, EngineArgs, SamplingParams
 from vllm.lora.request import LoRARequest
 from vllm.multimodal.utils import fetch_image
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 QUESTION = "What is the content of each image?"
 IMAGE_URLS = [

examples/offline_inference/vision_language_pooling.py

@@ -18,7 +18,7 @@ from PIL.Image import Image
 from vllm import LLM, EngineArgs
 from vllm.entrypoints.score_utils import ScoreMultiModalParam
 from vllm.multimodal.utils import fetch_image
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 ROOT_DIR = Path(__file__).parent.parent.parent
 EXAMPLES_DIR = ROOT_DIR / "examples"

examples/online_serving/openai_chat_completion_client_for_multimodal.py

@@ -26,7 +26,7 @@ import requests
 from openai import OpenAI
 from utils import get_first_model
 
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 # Modify OpenAI's API key and API base to use vLLM's API server.
 openai_api_key = "EMPTY"

examples/others/tensorize_vllm_model.py

@@ -16,7 +16,7 @@ from vllm.model_executor.model_loader.tensorizer import (
     tensorize_vllm_model,
     tensorizer_kwargs_arg,
 )
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 logger = logging.getLogger()
 

tests/engine/test_arg_utils.py

@@ -22,7 +22,7 @@ from vllm.engine.arg_utils import (
     optional_type,
     parse_type,
 )
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 @pytest.mark.parametrize(

tests/entrypoints/openai/test_cli_args.py

@@ -7,7 +7,7 @@ import pytest
 
 from vllm.entrypoints.openai.cli_args import make_arg_parser, validate_parsed_serve_args
 from vllm.entrypoints.openai.serving_models import LoRAModulePath
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 from ...utils import VLLM_PATH
 

tests/kernels/attention/test_deepgemm_attention.py

@@ -6,7 +6,6 @@ import pytest
 import torch
 
 from vllm.platforms import current_platform
-from vllm.utils import cdiv
 from vllm.utils.deep_gemm import (
     _ceil_to_ue8m0,
     calc_diff,
@@ -16,6 +15,7 @@ from vllm.utils.deep_gemm import (
     get_paged_mqa_logits_metadata,
 )
 from vllm.utils.import_utils import has_deep_gemm
+from vllm.utils.math_utils import cdiv
 
 
 def kv_cache_cast_to_fp8(x: torch.Tensor) -> torch.Tensor:

tests/kernels/attention/test_flashinfer_trtllm_attention.py

@@ -10,7 +10,7 @@ from tests.kernels.quantization.nvfp4_utils import (
     get_nvfp4_global_scale,
 )
 from vllm.platforms import current_platform
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 if not current_platform.is_device_capability(100):
     pytest.skip(

tests/kernels/attention/test_mla_decode_cpu.py

@@ -7,7 +7,7 @@ from torch import Tensor
 
 import vllm._custom_ops as ops
 from vllm.platforms import current_platform
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 
 
 def ref_mla(

tests/kernels/attention/test_triton_decode_attention.py

@@ -5,7 +5,7 @@ import pytest
 import torch
 
 from vllm.attention.ops.triton_decode_attention import decode_attention_fwd
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 
 
 @pytest.mark.parametrize("B", [3, 5])

tests/kernels/moe/test_cutlass_grouped_gemm.py

@@ -13,8 +13,8 @@ from tests.kernels.moe.utils import per_token_cast_to_fp8
 from tests.kernels.utils import baseline_scaled_mm
 from vllm import _custom_ops as ops
 from vllm.platforms import current_platform
-from vllm.utils import cdiv
 from vllm.utils.deep_gemm import per_block_cast_to_fp8
+from vllm.utils.math_utils import cdiv
 
 
 @pytest.mark.parametrize(

tests/kernels/moe/test_gpt_oss_triton_kernels.py

@@ -27,7 +27,7 @@ from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import (
     triton_kernel_moe_forward,
 )
 from vllm.model_executor.layers.utils import shuffle_weight
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 
 def deshuffle(w: torch.Tensor):

tests/kernels/moe/test_moe_align_block_size.py

@@ -13,7 +13,7 @@ from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
     moe_align_block_size,
 )
 from vllm.platforms import current_platform
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 NUM_TOKENS = [1, 3, 256, 2256, 4096]
 NUM_EXPERTS = [32, 160, 256, 257]

tests/kernels/moe/test_pplx_cutlass_moe.py

@@ -13,7 +13,7 @@ from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassBatchedExper
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
 from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
 from vllm.platforms import current_platform
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 
 from ...utils import multi_gpu_test
 from .parallel_utils import ProcessGroupInfo, parallel_launch

tests/kernels/moe/test_pplx_moe.py

@@ -45,7 +45,7 @@ from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
     TopKWeightAndReduceDelegate,
 )
 from vllm.platforms import current_platform
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 from ...utils import multi_gpu_test
 from .parallel_utils import ProcessGroupInfo, parallel_launch

tests/kernels/moe/test_silu_mul_fp8_quant_deep_gemm.py

@@ -8,7 +8,7 @@ from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
     persistent_masked_m_silu_mul_quant,
 )
 from vllm.platforms import current_platform
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 
 fp8_dtype = torch.float8_e4m3fn
 

tests/kernels/moe/utils.py

@@ -16,8 +16,8 @@ from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
 )
 from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
 from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
-from vllm.utils import round_up
 from vllm.utils.deep_gemm import per_block_cast_to_fp8
+from vllm.utils.math_utils import round_up
 
 
 def triton_moe(

tests/kernels/quant_utils.py

@@ -6,7 +6,7 @@ import torch
 
 from vllm.model_executor.layers.quantization.utils.quant_utils import group_broadcast
 from vllm.platforms import current_platform
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 # Using the default value (240.0) from pytorch will cause accuracy
 # issue on dynamic quantization models. Here use 224.0 for rocm.

tests/kernels/quantization/test_cutlass_scaled_mm.py

@@ -13,7 +13,7 @@ import torch
 from tests.kernels.utils import baseline_scaled_mm, opcheck, to_fp8, to_int8
 from vllm import _custom_ops as ops
 from vllm.platforms import current_platform
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 
 MNK_FACTORS = [
     (1, 256, 128),

tests/v1/attention/test_attention_backends.py

@@ -18,7 +18,7 @@ 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 cdiv
+from vllm.utils.math_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,

tests/v1/attention/test_mla_backends.py

@@ -22,7 +22,7 @@ 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 cdiv
+from vllm.utils.math_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

tests/v1/attention/test_sparse_mla_backends.py

@@ -23,7 +23,7 @@ from tests.v1.attention.utils import (
 from vllm import _custom_ops as ops
 from vllm.attention.ops import flashmla
 from vllm.model_executor.layers.linear import ColumnParallelLinear
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 from vllm.v1.attention.backends.mla.flashmla_sparse import FlashMLASparseBackend
 from vllm.v1.attention.backends.mla.indexer import split_prefill_chunks
 

tests/v1/engine/test_engine_args.py

@@ -8,7 +8,7 @@ import pytest
 from vllm.config import VllmConfig
 from vllm.engine.arg_utils import EngineArgs
 from vllm.usage.usage_lib import UsageContext
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 
 def test_prefix_caching_from_cli():

vllm/attention/layers/cross_attention.py

@@ -16,7 +16,7 @@ from vllm.attention.layer import Attention
 from vllm.attention.selector import get_attn_backend
 from vllm.config import CacheConfig, VllmConfig
 from vllm.logger import init_logger
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 from vllm.v1.attention.backends.utils import (
     CommonAttentionMetadata,
     subclass_attention_backend,

vllm/attention/ops/pallas_kv_cache_update.py

@@ -7,7 +7,7 @@ import jax
 from jax.experimental import pallas as pl
 from jax.experimental.pallas import tpu as pltpu
 
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 
 
 def _kv_cache_update_kernel(

vllm/attention/ops/rocm_aiter_paged_attn.py

@@ -6,7 +6,7 @@ import torch
 
 from vllm.attention.ops.paged_attn import PagedAttention
 from vllm.platforms import current_platform
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 
 FP8_DTYPE = current_platform.fp8_dtype()
 

vllm/benchmarks/datasets.py

@@ -58,7 +58,7 @@ except ImportError:
     librosa = PlaceholderModule("librosa")
 
 try:
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 

vllm/compilation/fusion_attn.py

@@ -19,7 +19,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
     kStaticTensorScale,
 )
 from vllm.platforms import current_platform
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 from .fusion import QUANT_OPS, empty_bf16, empty_fp32, empty_i32
 from .fx_utils import is_func

vllm/engine/arg_utils.py

@@ -82,7 +82,8 @@ from vllm.transformers_utils.config import (
     maybe_override_with_speculators,
 )
 from vllm.transformers_utils.utils import check_gguf_file
-from vllm.utils import FlexibleArgumentParser, is_in_ray_actor
+from vllm.utils import is_in_ray_actor
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.mem_constants import GiB_bytes
 from vllm.utils.network_utils import get_ip
 from vllm.v1.sample.logits_processor import LogitsProcessor

vllm/entrypoints/api_server.py

@@ -26,7 +26,8 @@ from vllm.entrypoints.utils import with_cancellation
 from vllm.logger import init_logger
 from vllm.sampling_params import SamplingParams
 from vllm.usage.usage_lib import UsageContext
-from vllm.utils import FlexibleArgumentParser, random_uuid, set_ulimit
+from vllm.utils import random_uuid, set_ulimit
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.version import __version__ as VLLM_VERSION
 
 logger = init_logger("vllm.entrypoints.api_server")

vllm/entrypoints/cli/benchmark/main.py

@@ -9,7 +9,7 @@ from vllm.entrypoints.cli.types import CLISubcommand
 from vllm.entrypoints.utils import VLLM_SUBCMD_PARSER_EPILOG
 
 if typing.TYPE_CHECKING:
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 else:
     FlexibleArgumentParser = argparse.ArgumentParser
 

vllm/entrypoints/cli/collect_env.py

@@ -8,7 +8,7 @@ from vllm.collect_env import main as collect_env_main
 from vllm.entrypoints.cli.types import CLISubcommand
 
 if typing.TYPE_CHECKING:
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 else:
     FlexibleArgumentParser = argparse.ArgumentParser
 

vllm/entrypoints/cli/main.py

@@ -20,7 +20,7 @@ def main():
     import vllm.entrypoints.cli.run_batch
     import vllm.entrypoints.cli.serve
     from vllm.entrypoints.utils import VLLM_SUBCMD_PARSER_EPILOG, cli_env_setup
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 
     CMD_MODULES = [
         vllm.entrypoints.cli.openai,

vllm/entrypoints/cli/openai.py

@@ -13,7 +13,7 @@ from openai.types.chat import ChatCompletionMessageParam
 from vllm.entrypoints.cli.types import CLISubcommand
 
 if TYPE_CHECKING:
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 else:
     FlexibleArgumentParser = argparse.ArgumentParser
 

vllm/entrypoints/cli/run_batch.py

@@ -11,7 +11,7 @@ from vllm.entrypoints.utils import VLLM_SUBCMD_PARSER_EPILOG
 from vllm.logger import init_logger
 
 if typing.TYPE_CHECKING:
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 else:
     FlexibleArgumentParser = argparse.ArgumentParser
 

vllm/entrypoints/cli/types.py

@@ -5,7 +5,7 @@ import argparse
 import typing
 
 if typing.TYPE_CHECKING:
-    from vllm.utils import FlexibleArgumentParser
+    from vllm.utils.argparse_utils import FlexibleArgumentParser
 else:
     FlexibleArgumentParser = argparse.ArgumentParser
 

vllm/entrypoints/openai/cli_args.py

@@ -29,7 +29,7 @@ from vllm.entrypoints.constants import (
 from vllm.entrypoints.openai.serving_models import LoRAModulePath
 from vllm.entrypoints.openai.tool_parsers import ToolParserManager
 from vllm.logger import init_logger
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 logger = init_logger(__name__)
 

vllm/entrypoints/openai/run_batch.py

@@ -32,7 +32,8 @@ from vllm.entrypoints.openai.serving_models import BaseModelPath, OpenAIServingM
 from vllm.entrypoints.openai.serving_score import ServingScores
 from vllm.logger import init_logger
 from vllm.reasoning import ReasoningParserManager
-from vllm.utils import FlexibleArgumentParser, random_uuid
+from vllm.utils import random_uuid
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.version import __version__ as VLLM_VERSION
 
 logger = init_logger(__name__)

vllm/entrypoints/utils.py

@@ -31,7 +31,7 @@ from vllm.entrypoints.openai.serving_models import LoRAModulePath
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.transformers_utils.tokenizers import MistralTokenizer
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 logger = init_logger(__name__)
 

vllm/model_executor/layers/fla/ops/layernorm_guard.py

@@ -21,7 +21,7 @@ import torch.nn.functional as F
 from einops import rearrange
 
 from vllm.triton_utils import tl, triton
-from vllm.utils import cdiv, next_power_of_2
+from vllm.utils.math_utils import cdiv, next_power_of_2
 
 from .utils import input_guard
 

vllm/model_executor/layers/fused_moe/config.py

@@ -14,9 +14,9 @@ from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import (
     OCP_MX_Scheme,
 )
 from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
-from vllm.utils import cdiv
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
 from vllm.utils.import_utils import has_triton_kernels
+from vllm.utils.math_utils import cdiv
 
 logger = init_logger(__name__)
 

vllm/model_executor/layers/fused_moe/deep_gemm_utils.py

@@ -10,8 +10,8 @@ import torch
 import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from vllm.model_executor.layers.fused_moe.utils import count_expert_num_tokens
 from vllm.triton_utils import tl, triton
-from vllm.utils import round_up
 from vllm.utils.deep_gemm import get_mk_alignment_for_contiguous_layout
+from vllm.utils.math_utils import round_up
 
 
 def expert_num_tokens_round_up_and_sum(

vllm/model_executor/layers/fused_moe/deepep_ht_prepare_finalize.py

@@ -12,7 +12,7 @@ from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
     TopKWeightAndReduceDelegate,
 )
 from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 from vllm.v1.worker.ubatching import (
     dbo_current_ubatch_id,
     dbo_enabled,

vllm/model_executor/layers/fused_moe/layer.py

@@ -55,9 +55,9 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_utils 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, round_up
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
 from vllm.utils.import_utils import has_deep_ep, has_pplx
+from vllm.utils.math_utils import cdiv, round_up
 from vllm.utils.torch_utils import current_stream, direct_register_custom_op
 from vllm.v1.worker.ubatching import dbo_current_ubatch_id
 

vllm/model_executor/layers/fused_moe/modular_kernel.py

@@ -16,7 +16,7 @@ from vllm.model_executor.layers.fused_moe.utils import (
     count_expert_num_tokens,
     disable_inplace,
 )
-from vllm.utils import cdiv
+from vllm.utils.math_utils import cdiv
 from vllm.v1.worker.ubatching import (
     dbo_current_ubatch_id,
     dbo_enabled,

vllm/model_executor/layers/fused_moe/moe_align_block_size.py

@@ -5,7 +5,7 @@ import torch
 
 from vllm import _custom_ops as ops
 from vllm.triton_utils import triton
-from vllm.utils import round_up
+from vllm.utils.math_utils import round_up
 
 
 def moe_align_block_size(

vllm/model_executor/layers/fused_moe/pplx_prepare_finalize.py

@@ -15,7 +15,7 @@ from vllm.model_executor.layers.fused_moe.utils import (
     _validate_scale_shape,
     moe_kernel_quantize_input,
 )
-from vllm.utils import cdiv, round_up
+from vllm.utils.math_utils import cdiv, round_up
 
 logger = init_logger(__name__)
 

vllm/model_executor/layers/fused_moe/utils.py

@@ -23,8 +23,8 @@ 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
 from vllm.utils.flashinfer import flashinfer_fp4_quantize
+from vllm.utils.math_utils import cdiv
 from vllm.utils.torch_utils import is_torch_equal_or_newer
 
 

vllm/model_executor/model_loader/tensorizer.py

@@ -26,7 +26,7 @@ from vllm.config import ModelConfig, ParallelConfig, VllmConfig, set_current_vll
 from vllm.logger import init_logger
 from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.import_utils import PlaceholderModule
 
 if TYPE_CHECKING: