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

[Feat] Supports Aclgraph for bge-m3 (#3171)

Browse Source 
### What this PR does / why we need it?
[Feat] Supports Aclgraph for bge-m3

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
```
pytest -s tests/e2e/singlecard/test_embedding.py
pytest -s tests/e2e/singlecard/test_embedding_aclgraph.py
```
to start an online server with bs 10, each batch's seq length=8192, we
set --max-num-batched-tokens=8192*10 to ensure encoder is not chunked:
```
vllm serve /home/data/bge-m3 --max_model_len 1024 --served-model-name "bge-m3" --task embed --host 0.0.0.0 --port 9095 --max-num-batched-tokens 81920 --compilation-config '{"cudagraph_capture_sizes":[8192, 10240, 20480, 40960, 81920]}'
```
For bs10, each batch's seq length=8192, QPS is improved from 85 to 104,
which is a 22% improvement, lots of host bound is reduced.


- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: xuyexiong <xuyexiong@huawei.com>
Co-authored-by: wangyongjun <1104133197@qq.com>
This commit is contained in:
xuyexiong
2025-10-14 23:07:45 +08:00
committed by GitHub
parent 434059e417
commit 02c26dcfc7
11 changed files with 307 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
vllm_ascend/attention/attention_v1.py
View File

@@ -606,9 +606,8 @@ class AscendAttentionBackendImpl(AttentionImpl):
num_actual_tokens = attn_metadata.num_actual_tokens
assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0
attn_type = self.attn_type
if attn_type != AttentionType.DECODER:
raise NotImplementedError("Encoder self-attention and "
"encoder/decoder cross-attention "
if attn_type != AttentionType.DECODER and attn_type != AttentionType.ENCODER_ONLY:
raise NotImplementedError("Encoder/decoder cross-attention "
"are not implemented for "
"PallasAttentionBackendImpl")
# View q k v to BSH.
@@ -628,9 +627,25 @@ class AscendAttentionBackendImpl(AttentionImpl):
key_cache=self.key_cache,
value_cache=self.value_cache,
slot_indices=slots)
if attn_type == AttentionType.ENCODER_ONLY:
cum_seq_len = attn_metadata.query_start_loc[1:].tolist()
attn_out = torch_npu.npu_fusion_attention(
query,
key,
value,
head_num=self.num_heads,
input_layout="TND",
scale=self.scale,
sparse_mode=4,
atten_mask=attn_metadata.attn_mask,
pre_tockens=attn_metadata.max_query_len,
next_tockens=attn_metadata.max_query_len,
actual_seq_qlen=cum_seq_len,
actual_seq_kvlen=cum_seq_len,
)
output = attn_out[0]
# V0-Style scheduler situation.
if attn_metadata.attn_state == AscendAttentionState.PrefillNoCache:
elif attn_metadata.attn_state == AscendAttentionState.PrefillNoCache:
output = self._forward_prefill_no_cache(
query, key, value, attn_metadata, output, num_tokens)
elif attn_metadata.attn_state == \