Skip to content

change ascend paged attention from BSH format to TND format for better performace #4295

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
jinminxi104 wants to merge 1 commit into
base: main
Choose a base branch
from
+8 −4
Draft

change ascend paged attention from BSH format to TND format for better performace #4295

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion lmdeploy/pytorch/backends/dlinfer/ascend/op_backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,7 +158,8 @@ def get_cpu_seqlens(is_decoding, is_unpaged_prefill):
repeat_interleave, used for attention metadata.
"""
if is_decoding:
q_seqlens_cpu = None
q_seqlens_cpu = None if AscendOpsBackend.enable_graph else torch.arange(1, step_context.kv_seqlens.size(0) + 1,
dtype=torch.int32)
kv_seqlens_cpu = kv_seqlens_expanded = step_context.kv_seqlens.cpu()
elif is_unpaged_prefill:
q_seqlens_cpu = step_context.q_seqlens.cpu()
Expand Down
3 changes: 3 additions & 0 deletions lmdeploy/pytorch/kernels/dlinfer/pagedattention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,6 +78,7 @@ def paged_token_attention(
k_cache,
v_cache,
attn_output,
q_seqlens,
kv_seq_len,
max_kv_seq_len,
block_offsets,
Expand All @@ -96,6 +97,7 @@ def paged_token_attention(
v_cache,
block_offsets,
block_size,
q_seqlens,
kv_seq_len,
max_kv_seq_len,
num_q_heads,
Expand Down Expand Up @@ -167,6 +169,7 @@ def paged_attention_fwd(
key_cache,
value_cache,
attn_output,
q_seqlens,
kv_seqlens,
max_kv_seq_len,
block_offsets,
Expand Down
6 changes: 3 additions & 3 deletions requirements/runtime_ascend.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,9 +22,9 @@ safetensors
sentencepiece
shortuuid
tiktoken
torch>=2.3.1,<2.9.0
torch-npu>=2.3.1,<2.9.0
torchvision>=0.18.1,<0.24.0
torch>=2.3.1,<2.10.0
torch-npu>=2.3.1,<2.10.0
torchvision>=0.18.1,<0.25.0
transformers
uvicorn
xgrammar
Loading