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

[CP&SP] Integrate FIA operator in mla_cp._forward_decode (#5641)

Browse Source 
### What this PR does / why we need it?
Replace the npu_multi_head_latent_attention with FIA operator in
mla_cp.py _forward_decode.
Adjust mla_attn_dpc_pcp in acl_graph.py

### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef

---------

Signed-off-by: 白永斌 <baiyongbin3@h-partners.com>
Signed-off-by: Bai Yongbin <845473182@qq.com>
Signed-off-by: tongyuzhou <t00886357@china.huawei.com>
Co-authored-by: 白永斌 <baiyongbin3@h-partners.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: tongyuzhou <t00886357@china.huawei.com>
This commit is contained in:
Bai Yongbin
2026-01-22 20:02:30 +08:00
committed by GitHub
parent 88632cf976
commit 7f91ac2649
4 changed files with 123 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/ut/attention/test_mla_cp.py
View File

@@ -450,11 +450,11 @@ class TestAscendMLAImpl(TestBase):
self.assertEqual(result.shape[2], self.impl.kv_lora_rank + 1)
@patch('vllm_ascend.attention.context_parallel.mla_cp.get_forward_context')
@patch("torch_npu.atb.npu_multi_head_latent_attention")
@patch("torch_npu.npu_fused_infer_attention_score")
@patch('torch_npu.npu_attention_update')
@patch_distributed_groups(dcp_size=2, pcp_size=2, needs_mocks=False)
def test_forward_decode_pcp_dcp(self, mock_npu_attention_update,
mock_npu_multi_head_latent_attention,
mock_npu_fused_infer_attention_score,
mock_get_forward_context):
self.impl.dcp_size = 2
self.impl.pcp_size = 2
@@ -470,8 +470,8 @@ class TestAscendMLAImpl(TestBase):
q_nope = torch.randn(B, N, self.impl.qk_nope_head_dim)
q_pe = torch.randn(B, N, self.impl.qk_rope_head_dim)
k_nope = torch.randn(NB, BS, 1, self.impl.kv_lora_rank)
k_pe = torch.randn(NB, BS, 1, self.impl.qk_rope_head_dim)
k_nope = torch.randn(NB, 1, BS, self.impl.kv_lora_rank)
k_pe = torch.randn(NB, 1, BS, self.impl.qk_rope_head_dim)
attn_metadata = MagicMock()
attn_metadata.attn_state = AscendAttentionState.SpecDecoding
@@ -485,7 +485,7 @@ class TestAscendMLAImpl(TestBase):
mock_npu_attention_update.return_value = (torch.randn(
B, self.impl.num_heads, self.impl.kv_lora_rank), None)
mock_npu_multi_head_latent_attention.return_value = [
mock_npu_fused_infer_attention_score.return_value = [
torch.randn(B, N, self.impl.kv_lora_rank),
torch.randn(B, N, 1)
]