[Cherry-pick]bmm_transpose to v011dev (#3995)

### What this PR does / why we need it?
Add a custom op to acclerater the deepseek model. The fusion ops combine
the bmm and transpose together, which is applied to mla module.
Cherry-pick from this commtid c68ddc11ce53334fc9a17bad58342148cbf14e86

### Does this PR introduce _any_ user-facing change?
No

---------

Signed-off-by: hust17yixuan <303660421@qq.com>
This commit is contained in:
Wang Yixuan
2025-12-08 19:22:14 +08:00
committed by GitHub
parent 6391f0625f
commit d412565ec9
15 changed files with 1736 additions and 13 deletions

View File

@@ -115,6 +115,14 @@ std::tuple<at::Tensor &, at::Tensor &, at::Tensor &, at::Tensor &> mla_preproces
}
void batch_matmul_transpose(const at::Tensor &tensor_a, const at::Tensor &tensor_b, at::Tensor &tensor_c,
c10::optional<c10::string_view> format_mode,
c10::optional<c10::string_view> quant_mode)
{
return;
}
} // namespace meta
} // namespace vllm_ascend
@@ -132,5 +140,7 @@ namespace {
ops.impl("sgmv_expand", &vllm_ascend::meta::sgmv_expand_meta);
// MLA preprocess
ops.impl("mla_preprocess", &vllm_ascend::meta::mla_preprocess);
// batch_matmul_transpose
ops.impl("batch_matmul_transpose", &vllm_ascend::meta::batch_matmul_transpose);
}
}