A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 17:06 UTC
[Refactor] Clean up unused variables & func (#32692)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Jan 23, 2026 at 17:04 UTC
37c9859fab60bbc346be20a662387479eb0760de
5 files changed
-30
vllm/entrypoints/openai/engine/protocol.py
-3
@@ -7,7 +7,6 @@ import time
7
from typing import Any, ClassVar, Literal, TypeAlias
8
9
import regex as re
10
-import torch
10
from pydantic import (
11
BaseModel,
12
ConfigDict,
@@ -25,8 +24,6 @@ from vllm.utils.import_utils import resolve_obj_by_qualname
24
25
logger = init_logger(__name__)
26
28
-_LONG_INFO = torch.iinfo(torch.long)
29
-
27
28
class OpenAIBaseModel(BaseModel):
29
# OpenAI API does allow extra fields
vllm/model_executor/layers/quantization/utils/petit_utils.py
-4
@@ -38,10 +38,6 @@ def _import_petit_kernel() -> "ModuleType":
38
raise ImportError(_PETIT_INSTALL_MSG) from None
39
40
41
-# The _require_petit function can now be a simple alias for consistency.
42
-_require_petit = _import_petit_kernel
43
-
44
-
41
def _check_petit_nvfp4_supported(
42
quant_method: str, group_size: int | None
43
) -> tuple[bool, str | None]:
vllm/model_executor/models/glmasr_utils.py
-20
@@ -166,23 +166,3 @@ def _extract_mask_for_item(
166
return feature_attention_mask[start_idx:end_idx]
167
mask_slice = feature_attention_mask[start_idx:end_idx]
168
return _normalize_to_tensor(mask_slice)
169
-
170
-
171
-def _get_num_features_for_item(
172
- feature_attention_mask: torch.Tensor | None,
173
- chunk_counts: torch.Tensor | list[int] | None,
174
- item_idx: int,
175
- audio_embeds: list[torch.Tensor] | None,
176
- merge_factor: int,
177
- conv_params: list[tuple[int, int, int]],
178
-) -> int:
179
- """Get number of features for a specific audio item."""
180
- if feature_attention_mask is not None:
181
- mask = _extract_mask_for_item(feature_attention_mask, chunk_counts, item_idx)
182
- audio_output_lengths = _get_audio_output_lengths_from_mask(
183
- mask, merge_factor, conv_params
184
- )
185
- return audio_output_lengths.sum().item()
186
- if audio_embeds is not None:
187
- return audio_embeds[item_idx].shape[0]
188
- raise ValueError("Either feature_attention_mask or audio_embeds must be provided")
vllm/model_executor/models/phi4mm_audio.py
-2
@@ -33,8 +33,6 @@ from vllm.model_executor.models.phi4mm_utils import (
33
unfold_tensor,
34
)
35
36
-_AUDIO_PLACEHOLDER_TOKEN_ID = 200011 # <|endoftext11|>
37
-
36
37
class ConformerEncoderLayer(nn.Module):
38
"""ConformerEncoder Layer module.
vllm/platforms/rocm.py
-1
@@ -48,7 +48,6 @@ _ROCM_UNSUPPORTED_MODELS: list[str] = []
48
49
# Models partially supported by ROCm.
50
# Architecture -> Reason.
51
-_ROCM_SWA_REASON = ()
51
_ROCM_PARTIALLY_SUPPORTED_MODELS: dict[str, str] = {}
52
_ROCM_DEVICE_ID_NAME_MAP: dict[str, str] = {
53
"0x74a0": "AMD_Instinct_MI300A",