EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

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

Browse Source 
### 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 c68ddc11ce

### 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
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm_ascend/attention/mla_v1.py
View File

@@ -565,14 +565,15 @@ class AscendMLAImpl(MLAAttentionImpl):
self.enable_mlapo = envs.VLLM_ASCEND_ENABLE_MLAPO
def _v_up_proj(self, x):
if self.W_UV.shape[0] * self.W_UV.shape[1] < 65536:
if x.dtype in [torch.float16, torch.bfloat16] \
and hasattr(torch.ops._C_ascend, "batch_matmul_transpose"):
x = x.view(-1, self.num_heads, self.kv_lora_rank)
x = torch_npu.npu_transpose_batchmatmul(x,
self.W_UV,
perm_x1=[1, 0, 2],
perm_x2=[0, 1, 2],
perm_y=[1, 0, 2])
x = x.reshape(-1, self.num_heads * self.v_head_dim)
b, _, _ = x.shape
res = torch.empty((b, self.num_heads, self.v_head_dim),
dtype=x.dtype,
device=x.device)
torch.ops._C_ascend.batch_matmul_transpose(x, self.W_UV, res)
x = res.reshape(-1, self.num_heads * self.v_head_dim)
else:
# Convert from (B, N, L) to (N, B, L)
x = x.view(-1, self.num_heads, self.kv_lora_rank).transpose(0, 1)