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

Upgrade CANN to 8.3.rc1 (#3945)

Browse Source 
### What this PR does / why we need it?
This PR upgrade CANN from 8.2rc1 to 8.3rc1 and remove the CANN version
check logic.

TODO: we notice that UT runs failed with CANN 8.3 image. So the base
image for UT is still 8.2. We'll fix it later.


- vLLM version: v0.11.0
- vLLM main:
83f478bb19

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-11-03 20:21:07 +08:00
committed by GitHub
parent 49d74785c4
commit cc2cd42ad3
39 changed files with 119 additions and 213 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
vllm_ascend/attention/attention_mask.py
View File

@@ -47,11 +47,10 @@ class AttentionMaskBuilder:
self.attn_mask_cache = attn_mask
self.device = device
self.pooling_mask = None
if torch.version.cann.startswith("8.3"):
assigned_mask_dim = 2048
self.chunked_prefill_attn_mask = torch.triu(
torch.ones(assigned_mask_dim, assigned_mask_dim),
diagonal=1).to(torch.int8).to(device)
assigned_mask_dim = 2048
self.chunked_prefill_attn_mask = torch.triu(
torch.ones(assigned_mask_dim, assigned_mask_dim),
diagonal=1).to(torch.int8).to(device)
@staticmethod
def get_mask_scale_factor(dtype: torch.dtype = torch.float16):
@@ -68,7 +67,7 @@ class AttentionMaskBuilder:
def get_attn_mask(self, max_seq_len: int, dtype: torch.dtype,
device: torch.device):
if max_seq_len == 2048 and torch.version.cann.startswith("8.3"):
if max_seq_len == 2048:
return self.chunked_prefill_attn_mask.to(torch.bool)
self._update_attn_cache(max_seq_len, dtype)
return self.attn_mask_cache[:max_seq_len, :max_seq_len].contiguous(
@@ -89,23 +88,7 @@ class AttentionMaskBuilder:
dtype: torch.dtype = None,
device: torch.device = None,
) -> torch.Tensor:
if torch.version.cann.startswith("8.3"):
return self.chunked_prefill_attn_mask
else:
if dtype not in [torch.float16, torch.bfloat16]:
raise ValueError(
"splitfuse_attn_mask now only supports bf16 and fp16")
max_seq_len = max(seq_lens, default=0)
self._update_attn_cache(max_seq_len, dtype)
# FIXME: Currently the mask value of chunked-prefill situation and Prefill-Only situation
# is not the same. Fix this in the future when kernel is ready.
mask_scale_factor = AttentionMaskBuilder.get_mask_scale_factor(
dtype)
attn_mask = torch.index_select(self.attn_mask_cache,
dim=0,
index=position)[:, :max_seq_len]
attn_mask *= mask_scale_factor
return attn_mask.contiguous().to(device, non_blocking=True)
return self.chunked_prefill_attn_mask
def _update_attn_cache(self, seqlen: int, dtype: torch.dtype):
if seqlen > self._seq_len_cached: