[Refactor] Clean up unused variables & func (#32692)

Signed-off-by: yewentao256 <zhyanwentao@126.com>

vllm/entrypoints/openai/engine/protocol.py

@@ -7,7 +7,6 @@ import time
 from typing import Any, ClassVar, Literal, TypeAlias
 
 import regex as re
-import torch
 from pydantic import (
     BaseModel,
     ConfigDict,
@@ -25,8 +24,6 @@ from vllm.utils.import_utils import resolve_obj_by_qualname
 
 logger = init_logger(__name__)
 
-_LONG_INFO = torch.iinfo(torch.long)
-
 
 class OpenAIBaseModel(BaseModel):
     # OpenAI API does allow extra fields

vllm/model_executor/layers/quantization/utils/petit_utils.py

@@ -38,10 +38,6 @@ def _import_petit_kernel() -> "ModuleType":
         raise ImportError(_PETIT_INSTALL_MSG) from None
 
 
-# The _require_petit function can now be a simple alias for consistency.
-_require_petit = _import_petit_kernel
-
-
 def _check_petit_nvfp4_supported(
     quant_method: str, group_size: int | None
 ) -> tuple[bool, str | None]:

vllm/model_executor/models/glmasr_utils.py

@@ -166,23 +166,3 @@ def _extract_mask_for_item(
         return feature_attention_mask[start_idx:end_idx]
     mask_slice = feature_attention_mask[start_idx:end_idx]
     return _normalize_to_tensor(mask_slice)
-
-
-def _get_num_features_for_item(
-    feature_attention_mask: torch.Tensor | None,
-    chunk_counts: torch.Tensor | list[int] | None,
-    item_idx: int,
-    audio_embeds: list[torch.Tensor] | None,
-    merge_factor: int,
-    conv_params: list[tuple[int, int, int]],
-) -> int:
-    """Get number of features for a specific audio item."""
-    if feature_attention_mask is not None:
-        mask = _extract_mask_for_item(feature_attention_mask, chunk_counts, item_idx)
-        audio_output_lengths = _get_audio_output_lengths_from_mask(
-            mask, merge_factor, conv_params
-        )
-        return audio_output_lengths.sum().item()
-    if audio_embeds is not None:
-        return audio_embeds[item_idx].shape[0]
-    raise ValueError("Either feature_attention_mask or audio_embeds must be provided")

vllm/model_executor/models/phi4mm_audio.py

@@ -33,8 +33,6 @@ from vllm.model_executor.models.phi4mm_utils import (
     unfold_tensor,
 )
 
-_AUDIO_PLACEHOLDER_TOKEN_ID = 200011  # <|endoftext11|>
-
 
 class ConformerEncoderLayer(nn.Module):
     """ConformerEncoder Layer module.

vllm/platforms/rocm.py

@@ -48,7 +48,6 @@ _ROCM_UNSUPPORTED_MODELS: list[str] = []
 
 # Models partially supported by ROCm.
 # Architecture -> Reason.
-_ROCM_SWA_REASON = ()
 _ROCM_PARTIALLY_SUPPORTED_MODELS: dict[str, str] = {}
 _ROCM_DEVICE_ID_NAME_MAP: dict[str, str] = {
     "0x74a0": "AMD_Instinct_MI300A",