[Kernel/Quant] Remove AQLM (#22943)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
This commit is contained in:
co-authored by
Wentao Ye
parent
3253ae765e
commit
4fc722eca4
@@ -476,32 +476,6 @@ if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
|
||||
dtype=input.dtype,
|
||||
device=input.device).sum(0)
|
||||
|
||||
@register_fake("_C::aqlm_gemm")
|
||||
def _aqlm_gemm_fake(input: torch.Tensor, codes: torch.Tensor,
|
||||
codebooks: torch.Tensor, scales: torch.Tensor,
|
||||
codebook_partition_sizes: list[int],
|
||||
bias: Optional[torch.Tensor]) -> torch.Tensor:
|
||||
out_features = codes.size(0) * codebooks.size(2)
|
||||
flat_input = input.reshape((-1, input.size(-1)))
|
||||
flat_output = torch.empty((flat_input.size(0), out_features),
|
||||
dtype=input.dtype,
|
||||
device=input.device)
|
||||
|
||||
output_sizes = list(input.shape)
|
||||
output_sizes.pop()
|
||||
output_sizes.append(-1)
|
||||
return flat_output.reshape(tuple(output_sizes))
|
||||
|
||||
@register_fake("_C::aqlm_dequant")
|
||||
def _aqlm_dequant_fake(
|
||||
codes: torch.Tensor, codebooks: torch.Tensor,
|
||||
codebook_partition_sizes: list[int]) -> torch.Tensor:
|
||||
in_features = codes.size(1) * 8
|
||||
out_features = codes.size(0)
|
||||
return torch.empty((out_features, in_features),
|
||||
dtype=codebooks.dtype,
|
||||
device=codebooks.device)
|
||||
|
||||
@register_fake("_C::machete_mm")
|
||||
def machete_mm_fake(
|
||||
a: torch.Tensor,
|
||||
@@ -957,21 +931,6 @@ def cutlass_fp4_moe_mm(out_tensors: torch.Tensor, a_tensors: torch.Tensor,
|
||||
sf_offsets)
|
||||
|
||||
|
||||
# aqlm
|
||||
def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
|
||||
codebooks: torch.Tensor, scales: torch.Tensor,
|
||||
codebook_partition_sizes: list[int],
|
||||
bias: Optional[torch.Tensor]) -> torch.Tensor:
|
||||
return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
|
||||
codebook_partition_sizes, bias)
|
||||
|
||||
|
||||
def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
|
||||
codebook_partition_sizes: list[int]) -> torch.Tensor:
|
||||
return torch.ops._C.aqlm_dequant(codes, codebooks,
|
||||
codebook_partition_sizes)
|
||||
|
||||
|
||||
# gptq_marlin
|
||||
def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
|
||||
size_k: int, size_n: int,
|
||||
|
||||
Reference in New Issue
Block a user