[Quantization][INC][ARK] Support INT2 XPU WOQ Linear (#47521)

Signed-off-by: Zhenzhong1 <zhenzhong.xu@intel.com>
Signed-off-by: Zhenzhong Xu <zhenzhong.xu@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Zhenzhong Xu
2026-07-14 08:29:45 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent 21472f32ea
commit 0b0ef8d7eb
3 changed files with 158 additions and 5 deletions
@@ -20,6 +20,8 @@ if TYPE_CHECKING:
logger = init_logger(__name__)
XPU_WNA16_SUPPORTED_BITS = {2, 4}
class INCWna16Scheme(INCScheme):
@staticmethod
@@ -35,7 +37,7 @@ class INCWna16Scheme(INCScheme):
):
del config, layer
if current_platform.is_xpu():
if layer_config.bits == 4 and layer_config.sym:
if layer_config.bits in XPU_WNA16_SUPPORTED_BITS and layer_config.sym:
from .inc_ark_ops import get_ark_state
from .inc_wna16_linear import (
INCARKLinearMethod,
@@ -45,6 +47,13 @@ class INCWna16Scheme(INCScheme):
is_ark_available, ark_error, _, _ = get_ark_state()
if is_ark_available:
return INCLinearMethod(INCARKLinearMethod(layer_config))
elif layer_config.bits == 2:
raise NotImplementedError(
"INC int2 on XPU requires the ARK backend. "
f"Layer: {prefix}. "
f"auto_round_kernel unavailable: "
f"{ark_error or 'unknown error'}"
)
logger.debug(
"ARK backend is unavailable for layer %s; "