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 7, 11:30 UTC
[Misc] Cleanup useless `current_platform` import (#35715)
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
wangxiyuan committed
Mar 2, 2026 at 17:36 UTC
510bc9e1df082fc58fef1867399cd02baa9ebb47
5 files changed
-17
vllm/compilation/passes/fusion/sequence_parallelism.py
-4
@@ -18,7 +18,6 @@ from vllm.logger import init_logger
18
from vllm.model_executor.layers.quantization.utils.quant_utils import (
19
kFp8StaticTensorSym,
20
)
21
-from vllm.platforms import current_platform
21
22
from ..inductor_pass import enable_fake_mode
23
from ..utility.noop_elimination import NoOpEliminationPass
@@ -215,9 +214,6 @@ class MiddleAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
214
)
215
216
218
-FP8_DTYPE = current_platform.fp8_dtype()
219
-
220
-
217
class FirstAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
218
def __init__(
219
self,
vllm/config/model.py
-6
@@ -461,8 +461,6 @@ class ModelConfig:
461
462
self.maybe_pull_model_tokenizer_for_runai(self.model, self.tokenizer)
463
464
- from vllm.platforms import current_platform
465
-
464
if self.override_attention_dtype is not None and not current_platform.is_rocm():
465
warnings.warn(
466
"override-attention-dtype is set but not using ROCm platform",
@@ -940,8 +938,6 @@ class ModelConfig:
938
f"Unknown quantization method: {self.quantization}. Must "
939
f"be one of {supported_quantization}."
940
)
943
- from vllm.platforms import current_platform
944
-
941
current_platform.verify_quantization(self.quantization)
942
943
if self.quantization in me_quant.DEPRECATED_QUANTIZATION_METHODS:
@@ -1811,8 +1807,6 @@ def _resolve_auto_dtype(
1807
*,
1808
is_pooling_model: bool,
1809
):
1814
- from vllm.platforms import current_platform
1815
-
1810
supported_dtypes = [
1811
dtype
1812
for dtype in current_platform.supported_dtypes
vllm/distributed/parallel_state.py
-2
@@ -385,8 +385,6 @@ class GroupCoordinator:
385
self.cpu_group, 1 << 22, 6
386
)
387
388
- from vllm.platforms import current_platform
389
-
388
self.use_custom_op_call = (
389
current_platform.is_cuda_alike() or current_platform.is_tpu()
390
)
vllm/v1/attention/backends/fa_utils.py
-3
@@ -55,9 +55,6 @@ elif current_platform.is_rocm():
55
def get_flash_attn_version(
56
requires_alibi: bool = False, head_size: int | None = None
57
) -> int | None:
58
- # import here to avoid circular dependencies
59
- from vllm.platforms import current_platform
60
-
58
if current_platform.is_xpu():
59
return 2
60
if current_platform.is_rocm():
vllm/v1/attention/backends/flashinfer.py
-2
@@ -374,8 +374,6 @@ class FlashInferBackend(AttentionBackend):
374
375
@classmethod
376
def get_required_kv_cache_layout(cls) -> KVCacheLayoutType | None:
377
- from vllm.platforms import current_platform
378
-
377
capability = current_platform.get_device_capability()
378
if capability is not None and capability.major == 10:
379
return "HND"