[Quant][Kernel] Remove GPTQ Group/Dynamic Activation Ordering (#54809)
Signed-off-by: Roderick-Wu <roderickwu2003@gmail.com> Signed-off-by: Roderick Wu <roderickwu2003@gmail.com> Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -106,7 +106,7 @@ class INCXPUW4A8LinearMethod(INCXPULinearMethod):
|
||||
layer.scales_fp16,
|
||||
layer.qzeros,
|
||||
self.group_size,
|
||||
None,
|
||||
None, # Retained by the external XPU op ABI.
|
||||
bias,
|
||||
)
|
||||
return out.to(x.dtype).reshape(out_shape)
|
||||
|
||||
@@ -14,7 +14,6 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
|
||||
from vllm.model_executor.parameter import (
|
||||
GroupQuantScaleParameter,
|
||||
PackedvLLMParameter,
|
||||
RowvLLMParameter,
|
||||
)
|
||||
from vllm.scalar_type import scalar_types
|
||||
|
||||
@@ -254,16 +253,6 @@ class INCXPULinearBase(INCLinearScheme):
|
||||
layer.register_parameter("scales", scales)
|
||||
layer.register_parameter("qzeros", qzeros)
|
||||
|
||||
g_idx = RowvLLMParameter(
|
||||
data=torch.tensor(
|
||||
[i // self.group_size for i in range(input_size_per_partition)],
|
||||
dtype=torch.int32,
|
||||
),
|
||||
input_dim=0,
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
layer.register_parameter("g_idx", g_idx)
|
||||
|
||||
def _convert_awq_qweight_to_gptq(self, qw: torch.Tensor) -> torch.Tensor:
|
||||
"""Convert AWQ qweight [K, N // pf] to GPTQ qweight [K // pf, N].
|
||||
|
||||
@@ -356,7 +345,7 @@ class INCXPULinearMethod(INCXPULinearBase):
|
||||
layer.scales,
|
||||
layer.qzeros,
|
||||
self.group_size,
|
||||
None,
|
||||
None, # Retained by the external XPU op ABI.
|
||||
)
|
||||
return out.reshape(out_shape)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user