[Model] Add DeepSeek-V4 CPU backend (#55355)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Li, Jiang
2026-09-10 14:14:48 +08:00
committed by GitHub
parent 49632fb3ac
commit c3ccc0e957
47 changed files with 9564 additions and 627 deletions
@@ -6,6 +6,9 @@ import torch
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.utils import replace_parameter
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
_upcast_e8m0_to_fp32,
)
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
convert_to_channelwise,
)
@@ -281,10 +284,12 @@ class CPUFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel):
else params.weight_scale
)
assert weight_scale is not None
if weight_scale.dtype in (torch.float8_e8m0fnu, torch.uint8):
weight_scale = _upcast_e8m0_to_fp32(weight_scale.data)
replace_parameter(
layer,
scale_attr,
torch.nn.Parameter(weight_scale.data, requires_grad=False),
torch.nn.Parameter(weight_scale.data.contiguous(), requires_grad=False),
)
def apply_weights(