[Misc] Remove `override_attention_dtype` (#48684)
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
wangxiyuan committed
Aug 14, 2026 at 23:15 UTC
cdc4824a21eaa986d4d1fee90a7e6465c9f706e6
2 files changed
-15
vllm/config/model.py
-10
@@ -1,7 +1,6 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
-import warnings
4
from collections.abc import Callable
5
from dataclasses import InitVar, field
6
from functools import cached_property
@@ -349,8 +348,6 @@ class ModelConfig:
348
- "transformers" will use the Transformers model implementation.
349
- "terratorch" will use the TerraTorch model implementation.
350
"""
352
- override_attention_dtype: str | None = None
353
- """Override dtype for attention"""
351
logits_processors: list[str | type[LogitsProcessor]] | None = None
352
"""One or more logits processors' fully-qualified class names or class
353
definitions"""
@@ -433,7 +430,6 @@ class ModelConfig:
430
"config_format",
431
"hf_token",
432
"hf_overrides",
436
- "override_attention_dtype",
433
"logits_processors",
434
"io_processor_plugin",
435
"pooler_config",
@@ -595,12 +591,6 @@ class ModelConfig:
591
self.hf_token,
592
)
593
598
- if self.override_attention_dtype is not None and not current_platform.is_rocm():
599
- warnings.warn(
600
- "override-attention-dtype is set but not using ROCm platform",
601
- stacklevel=2,
602
- )
603
-
594
if self.enable_sleep_mode:
595
if not current_platform.is_sleep_mode_available():
596
raise ValueError("Sleep mode is not supported on current platform.")
vllm/engine/arg_utils.py
-5
@@ -700,7 +700,6 @@ class EngineArgs:
700
ModelConfig, "override_generation_config"
701
)
702
model_impl: str = ModelConfig.model_impl
703
- override_attention_dtype: str | None = ModelConfig.override_attention_dtype
703
attention_backend: AttentionBackendEnum | None = AttentionConfig.backend
704
705
kv_cache_dtype_skip_layers: list[str] = get_field(
@@ -917,9 +916,6 @@ class EngineArgs:
916
"--enable-cumem-allocator", **model_kwargs["enable_cumem_allocator"]
917
)
918
model_group.add_argument("--model-impl", **model_kwargs["model_impl"])
920
- model_group.add_argument(
921
- "--override-attention-dtype", **model_kwargs["override_attention_dtype"]
922
- )
919
model_group.add_argument(
920
"--logits-processors", **model_kwargs["logits_processors"]
921
)
@@ -1787,7 +1783,6 @@ class EngineArgs:
1783
enable_sleep_mode=self.enable_sleep_mode,
1784
enable_cumem_allocator=self.enable_cumem_allocator,
1785
model_impl=self.model_impl,
1790
- override_attention_dtype=self.override_attention_dtype,
1786
logits_processors=self.logits_processors,
1787
video_pruning_rate=self.video_pruning_rate,
1788
video_pruning_method=self.video_pruning_method,