[Perf] Optimize moe_align_block_size CUDA kernel (#19572)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -1524,15 +1524,6 @@ def moe_align_block_size(topk_ids: torch.Tensor, num_experts: int,
|
||||
num_tokens_post_pad)
|
||||
|
||||
|
||||
def sgl_moe_align_block_size(topk_ids: torch.Tensor, num_experts: int,
|
||||
block_size: int, sorted_token_ids: torch.Tensor,
|
||||
experts_ids: torch.Tensor,
|
||||
num_tokens_post_pad: torch.Tensor) -> None:
|
||||
torch.ops._moe_C.sgl_moe_align_block_size(topk_ids, num_experts,
|
||||
block_size, sorted_token_ids,
|
||||
experts_ids, num_tokens_post_pad)
|
||||
|
||||
|
||||
def moe_wna16_gemm(input: torch.Tensor, output: torch.Tensor,
|
||||
b_qweight: torch.Tensor, b_scales: torch.Tensor,
|
||||
b_qzeros: Optional[torch.Tensor],
|
||||
|
||||
Reference in New Issue
Block a user