[CPU] Support int8 compute mode in CPU AWQ (#35697)

Signed-off-by: Yintong Lu <yintong.lu@intel.com>
This commit is contained in:
Yintong Lu
2026-03-31 15:27:37 +08:00
committed by GitHub
parent 42318c840b
commit f09daea261
10 changed files with 1197 additions and 11 deletions
+32
View File
@@ -2967,6 +2967,38 @@ if hasattr(torch.ops._C, "int8_scaled_mm_with_quant"):
return torch.empty((M, N), dtype=out_dtype)
if hasattr(torch.ops._C, "convert_weight_packed_scale_zp"):
@register_fake("_C::convert_weight_packed_scale_zp")
def convert_weight_packed_scale_zp_fake(
qweight: torch.Tensor,
qzeros: torch.Tensor,
scales: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
return (
torch.empty_like(qweight),
torch.empty_like(qzeros),
torch.empty_like(scales),
)
if hasattr(torch.ops._C, "int4_scaled_mm_cpu"):
@register_fake("_C::int4_scaled_mm_cpu")
def int4_scaled_mm_cpu_fake(
x: torch.Tensor,
w: torch.Tensor,
w_zeros: torch.Tensor,
w_scales: torch.Tensor,
bias: torch.Tensor | None,
) -> torch.Tensor:
N = w_scales.size(0) * w_scales.size(-1)
return torch.empty((x.size(0), N), dtype=x.dtype, device=x.device)
_supports_cpu_w4a8_int8 = bool(hasattr(torch.ops._C, "convert_weight_packed_scale_zp"))
class CPUDNNLGEMMHandler:
def __init__(self) -> None:
self.handler_tensor: torch.Tensor | None = None