[Kernel] fix types used in aqlm and ggml kernels to support dynamo (#7596)
This commit is contained in:
@@ -95,7 +95,7 @@ def generic_dequantize_gemm(
|
||||
codebooks: torch.
|
||||
Tensor, # [num_codebooks, codebook_size, out_group_size, in_group_size]
|
||||
scales: torch.Tensor, # [num_out_groups, 1, 1, 1]
|
||||
output_partition_sizes: torch.IntTensor,
|
||||
output_partition_sizes: List[int],
|
||||
bias: Optional[torch.Tensor],
|
||||
) -> torch.Tensor:
|
||||
output_shape = input.shape[:-1] + (scales.shape[0], )
|
||||
@@ -133,7 +133,7 @@ def optimized_dequantize_gemm(
|
||||
codebooks: torch.
|
||||
Tensor, # [num_codebooks, codebook_size, out_group_size, in_group_size]
|
||||
scales: torch.Tensor, # [num_out_groups, 1, 1, 1]
|
||||
output_partition_sizes: torch.IntTensor,
|
||||
output_partition_sizes: List[int],
|
||||
bias: Optional[torch.Tensor],
|
||||
) -> torch.Tensor:
|
||||
weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
|
||||
@@ -288,10 +288,8 @@ class AQLMLinearMethod(LinearMethodBase):
|
||||
codebooks,
|
||||
{
|
||||
# metadata indicates fixed size concatenated along dim 0
|
||||
"is_metadata":
|
||||
True,
|
||||
"output_partition_sizes":
|
||||
torch.tensor(output_partition_sizes, device='cpu'),
|
||||
"is_metadata": True,
|
||||
"output_partition_sizes": output_partition_sizes
|
||||
},
|
||||
)
|
||||
|
||||
@@ -334,7 +332,7 @@ class AQLMLinearMethod(LinearMethodBase):
|
||||
codes = layer.codes
|
||||
scales = layer.scales
|
||||
output_partition_sizes = getattr(codebooks, "output_partition_sizes",
|
||||
None)
|
||||
[])
|
||||
|
||||
nbooks = codes.shape[2]
|
||||
ingroups = codebooks.shape[3]
|
||||
|
||||
@@ -212,6 +212,7 @@ class GPTQLinearMethod(LinearMethodBase):
|
||||
layer.g_idx.data = torch.argsort(layer.g_idx).to(torch.int)
|
||||
else:
|
||||
layer.g_idx.data = torch.empty((0, ),
|
||||
dtype=torch.int,
|
||||
device=layer.g_idx.device)
|
||||
layer.exllama_state = ExllamaState.READY
|
||||
ops.gptq_shuffle(layer.qweight, layer.g_idx,
|
||||
|
||||
Reference in New Issue
Block a user