Fix documentation of m_grouped_gemm_fp8_fp8_bf16_nt_contiguous in m_grouped_gemm.py

This commit is contained in:
Liang 2025-03-04 11:26:23 +08:00 committed by GitHub
parent dff6bb6f0b
commit ded740f736
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -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