[CPU] Refactor CPU unquantized linear (#24150)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Li, Jiang
2025-09-04 14:28:45 +08:00
committed by GitHub
parent cb55ad86fe
commit 57b1ce94f7
9 changed files with 466 additions and 26 deletions
+29
View File
@@ -1928,6 +1928,35 @@ class CPUDNNLGEMMHandler:
torch.ops._C.release_dnnl_matmul_handler(self.handler)
if hasattr(torch.ops._C, "create_onednn_mm_handler"):
_supports_onednn = True
else:
_supports_onednn = False
def create_onednn_mm(
weight: torch.Tensor, # [K, N]
primitive_cache_size: int = 128,
) -> CPUDNNLGEMMHandler:
handler = CPUDNNLGEMMHandler()
handler.k, handler.n = weight.size()
handler.handler = torch.ops._C.create_onednn_mm_handler(
weight, primitive_cache_size)
return handler
def onednn_mm(
dnnl_handler: CPUDNNLGEMMHandler,
x: torch.Tensor,
bias: Optional[torch.Tensor],
) -> torch.Tensor:
output = torch.empty((*x.shape[0:-1], dnnl_handler.n), dtype=x.dtype)
torch.ops._C.onednn_mm(output, x.reshape(-1, dnnl_handler.k), bias,
dnnl_handler.handler)
return output
def create_onednn_scaled_mm(
weight: torch.Tensor, # [K, N]
weight_scales: torch.Tensor,