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

[Refactor] 2/N Unify all mask generation methods and cache mask (#4779)

Browse Source 
RFC: https://github.com/vllm-project/vllm-ascend/issues/4629

Reason:

There are various types of masks here, and some of them do not have a
caching mechanism. As a result, the masks need to be initialized for
each layer, leading to waste of video memory.

At the same time, we hope to standardize the management and usage of
masks.

So we have gathered all the masks into the AttentionMaskBuilder class.

Todo:
1. remove spec_attn_mask;  @LICO1314
2. remove pcp_prefill_mask; @LICO1314


- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Signed-off-by: ZYang6263 <zy626375@gmail.com>
Signed-off-by: ZYang6263 <50876451+ZYang6263@users.noreply.github.com>
Signed-off-by: daishixun <dsxsteven@sina.com>
Signed-off-by: lulina <lina.lulina@huawei.com>
Signed-off-by: zengran <zengran2@huawei.com>
Signed-off-by: shiro-zzzz <zhangdianhao@huawei.com>
Signed-off-by: dependabot[bot] <support@github.com>
Signed-off-by: 李少鹏 <lishaopeng21@huawei.com>
Signed-off-by: xuyexiong <xuyexiong@huawei.com>
Signed-off-by: MengqingCao <cmq0113@163.com>
Signed-off-by: lhp-deep <liuhaopeng1@huawei.com>
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
Signed-off-by: wangli <wangli858794774@gmail.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: ZYang6263 <50876451+ZYang6263@users.noreply.github.com>
Co-authored-by: dsxsteven <36877507+dsxsteven@users.noreply.github.com>
Co-authored-by: LuLina <lina.lulina@huawei.com>
Co-authored-by: zengzengran <zengran2@huawei.com>
Co-authored-by: shiro-zzzz <zhangdianhao@huawei.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: shaopeng-666 <lishaopeng21@huawei.com>
Co-authored-by: xuyexiong <xuyexiong@huawei.com>
Co-authored-by: lhp-deep <liuhaopeng1@huawei.com>
Co-authored-by: Canlin Guo <canlinguosdu@gmail.com>
Co-authored-by: Li Wang <wangli858794774@gmail.com>
This commit is contained in:
weijinqian0
2025-12-09 18:51:00 +08:00
committed by GitHub
parent dee00d0de3
commit c331503677
6 changed files with 66 additions and 174 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm_ascend/attention/mla_v1.py
View File

@@ -202,7 +202,6 @@ class AscendMLAMetadataBuilder:
understand this class
"""
# _attn_mask_builder = None
def __init__(self,
kv_cache_spec,
layer_names,
@@ -862,7 +861,6 @@ class AscendMLAImpl(MLAAttentionImpl):
vllm_config = get_current_vllm_config()
self.ring_mla_mask_size = 512
self.prefill_mask = None
self.speculative_config = vllm_config.speculative_config
self.enable_mlapo = envs.VLLM_ASCEND_ENABLE_MLAPO
@@ -1167,10 +1165,7 @@ class AscendMLAImpl(MLAAttentionImpl):
.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
k_pe = k_pe.expand((*k_nope.shape[:-1], -1))
if self.pcp_size > 1:
mask = attn_metadata.prefill.pcp_metadata.pcp_prefill_mask
else:
mask = self.prefill_mask
mask = attn_metadata.attn_mask
torch_npu.atb.npu_ring_mla(
q_nope=q_nope,
q_rope=q_pe,
@@ -1214,24 +1209,12 @@ class AscendMLAImpl(MLAAttentionImpl):
num_tokens,
dtype=torch.float32,
device=q_nope.device)
if self.prefill_mask is None:
if q_nope.dtype == torch.float16:
mask_value = torch.finfo(torch.float32).min
else:
mask_value = 1
prefill_mask = torch.triu(
torch.ones(self.ring_mla_mask_size,
self.ring_mla_mask_size,
device=q_nope.device,
dtype=q_nope.dtype), 1)
self.prefill_mask = torch.where(prefill_mask == 1, mask_value,
0).to(q_nope.dtype)
torch_npu.atb.npu_ring_mla(q_nope=q_nope,
q_rope=q_pe,
k_nope=k_nope,
k_rope=k_pe,
value=value,
mask=self.prefill_mask,
mask=attn_metadata.attn_mask,
seqlen=attn_metadata.prefill.query_lens,
head_num=self.num_heads,
kv_head_num=self.num_heads,