biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ez] Remove checks for torch version <= 2.8 (#33209)

Browse Source 
Signed-off-by: angelayi <yiangela7@gmail.com>
This commit is contained in:
Angela Yi
2026-01-28 13:03:56 -08:00
committed by GitHub
parent 59bcc5b6f2
commit 4197168ea5
11 changed files with 30 additions and 139 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/layers/fused_moe/fused_moe.py
View File

@@ -52,7 +52,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
)
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils.torch_utils import direct_register_custom_op, is_torch_equal_or_newer
from vllm.utils.torch_utils import direct_register_custom_op
logger = init_logger(__name__)
@@ -1406,11 +1406,6 @@ direct_register_custom_op(
op_func=inplace_fused_experts,
mutates_args=["hidden_states"],
fake_impl=inplace_fused_experts_fake,
tags=(
()
if is_torch_equal_or_newer("2.7.0")
else (torch.Tag.needs_fixed_stride_order,)
),
)
@@ -1501,11 +1496,6 @@ direct_register_custom_op(
op_name="outplace_fused_experts",
op_func=outplace_fused_experts,
fake_impl=outplace_fused_experts_fake,
tags=(
()
if is_torch_equal_or_newer("2.7.0")
else (torch.Tag.needs_fixed_stride_order,)
),
)

3
vllm/model_executor/layers/quantization/mxfp4.py
View File

@@ -56,7 +56,6 @@ from vllm.scalar_type import scalar_types
from vllm.utils.flashinfer import has_flashinfer
from vllm.utils.import_utils import has_triton_kernels
from vllm.utils.math_utils import round_up
from vllm.utils.torch_utils import is_torch_equal_or_newer
logger = init_logger(__name__)
@@ -89,7 +88,6 @@ def get_mxfp4_backend_with_lora() -> Mxfp4Backend:
# If FlashInfer is not available, try either Marlin or Triton
triton_kernels_supported = (
has_triton_kernels()
and is_torch_equal_or_newer("2.8.0")
# NOTE: triton_kernels are only confirmed to work on SM90 and SM100
# SM110 fails with this error: https://github.com/vllm-project/vllm/issues/29317
# SM120 needs this fix: https://github.com/triton-lang/triton/pull/8498
@@ -151,7 +149,6 @@ def get_mxfp4_backend(with_lora_support: bool) -> Mxfp4Backend:
# If FlashInfer is not available, try either Marlin or Triton
triton_kernels_supported = (
has_triton_kernels()
and is_torch_equal_or_newer("2.8.0")
# NOTE: triton_kernels are only confirmed to work on SM90 and SM100
# SM110 fails with this error: https://github.com/vllm-project/vllm/issues/29317
# SM120 needs this fix: https://github.com/triton-lang/triton/pull/8498

14
vllm/model_executor/layers/quantization/torchao.py
View File

@@ -108,20 +108,6 @@ class TorchAOConfig(QuantizationConfig):
skip_modules: list[str] | None = None,
is_checkpoint_torchao_serialized: bool = False,
) -> None:
"""
# TorchAO quantization relies on tensor subclasses. In order,
# to enable proper caching this needs standalone compile
if is_torch_equal_or_newer("2.8.0.dev"):
os.environ["VLLM_TEST_STANDALONE_COMPILE"] = "1"
logger.info(
"Using TorchAO: Setting VLLM_TEST_STANDALONE_COMPILE=1")
# TODO: remove after the torch dependency is updated to 2.8
if is_torch_equal_or_newer(
"2.7.0") and not is_torch_equal_or_newer("2.8.0.dev"):
os.environ["VLLM_DISABLE_COMPILE_CACHE"] = "1"
logger.info("Using TorchAO: Setting VLLM_DISABLE_COMPILE_CACHE=1")
"""
super().__init__()
self.torchao_config = torchao_config
self.skip_modules = skip_modules or []