[Platform] Add current_platform.num_compute_units interface (#35042)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Signed-off-by: Kunshang Ji <jikunshang95@gmail.com>
This commit is contained in:
Kunshang Ji
2026-02-24 22:22:49 -08:00
committed by GitHub
parent 92510edc32
commit 8ad54a991b
24 changed files with 72 additions and 52 deletions
@@ -11,6 +11,7 @@ from vllm.model_executor.layers.quantization.utils.allspark_utils import (
check_allspark_supported_dtype_shape,
)
from vllm.model_executor.parameter import BasevLLMParameter, permute_param_layout_
from vllm.utils.platform_utils import num_compute_units
from .MPLinearKernel import MPLinearKernel, MPLinearLayerConfig
@@ -45,7 +46,7 @@ class AllSparkLinearKernel(MPLinearKernel):
# prepare the parameters required for the kernel
properties = torch.cuda.get_device_properties(device.index)
sm_count = properties.multi_processor_count
sm_count = num_compute_units(device.index)
sm_version = properties.major * 10 + properties.minor
gemm_args = {}
gemm_args["sm_count"] = sm_count
@@ -7,7 +7,7 @@ import torch
import vllm.envs as envs
from vllm import _custom_ops as ops
from vllm.platforms import current_platform
from vllm.utils.platform_utils import get_cu_count
from vllm.utils.platform_utils import num_compute_units
from vllm.utils.torch_utils import direct_register_custom_op
from .ScaledMMLinearKernel import (
@@ -36,7 +36,7 @@ def rocm_per_tensor_float_w8a8_scaled_mm_impl(
out_dtype,
As,
Bs,
get_cu_count(),
num_compute_units(),
bias,
)
# Fallback