[Core] Set linear_weights directly on the layer (#3977)

This commit is contained in:
Antoni Baum
2024-04-11 16:35:51 -04:00
committed by GitHub
parent 8afca50889
commit a10d3056da
8 changed files with 114 additions and 102 deletions
@@ -91,12 +91,14 @@ class MarlinLinearMethod(LinearMethodBase):
def create_weights(
self,
layer: torch.nn.Module,
input_size_per_partition: int,
output_size_per_partition: int,
input_size: int,
output_size: int,
params_dtype: torch.dtype,
) -> Dict[str, Any]:
**extra_weight_attrs,
):
del output_size # Unused.
if params_dtype != torch.float16:
@@ -187,21 +189,22 @@ class MarlinLinearMethod(LinearMethodBase):
dtype=torch.int),
requires_grad=False)
return {
"B": qweight,
"s": scales,
"workspace": workspace,
}
layer.register_parameter("B", qweight)
set_weight_attrs(qweight, extra_weight_attrs)
layer.register_parameter("s", scales)
set_weight_attrs(scales, extra_weight_attrs)
layer.register_parameter("workspace", workspace)
set_weight_attrs(workspace, extra_weight_attrs)
def apply_weights(
self,
weights: Dict[str, Any],
layer: torch.nn.Module,
x: torch.Tensor,
bias: Optional[torch.Tensor] = None,
) -> torch.Tensor:
qweight = weights["B"]
scales = weights["s"]
workspace = weights["workspace"]
qweight = layer.B
scales = layer.s
workspace = layer.workspace
x_2d = x.view(-1, x.shape[-1])