mirror of
https://github.com/deepseek-ai/DeepGEMM
synced 2025-05-02 04:23:50 +00:00
Fix documentation of m_grouped_gemm_fp8_fp8_bf16_nt_contiguous
in m_grouped_gemm.py
This commit is contained in:
parent
dff6bb6f0b
commit
ded740f736
@ -54,7 +54,6 @@ def m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(lhs: Tuple[torch.Tensor, torch.Ten
|
||||
`m_indices[i]` records the group which the j-th row of the LHS belong to,
|
||||
which means that the i-th row of the LHS matrix will be multiplied with `rhs[m_indices[i]]`.
|
||||
Values of `m_indices` in every-m-alignment-block must also be the same.
|
||||
`-1` in this tensor indicates no RHS matrix selected, the kernel will skip the computation for that aligned block.
|
||||
"""
|
||||
lhs, lhs_scales = lhs
|
||||
rhs, rhs_scales = rhs
|
||||
|
Loading…
Reference in New Issue
Block a user