Skip to content

[SW-225565] Enable triangular softmax with merged prefill #197

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
wants to merge 4 commits into
base: main
Choose a base branch
from
Draft

[SW-225565] Enable triangular softmax with merged prefill #197

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
9a99d34
Enable triangular softmax with merged prefill
kamil-kaczor May 26, 2025
b6f7e94
Change to generic name
kamil-kaczor Jun 3, 2025
0178798
Change to generic name
kamil-kaczor Jun 3, 2025
06756a9
Enable valid_seq_len in triangular merged
kamil-kaczor Jul 14, 2025
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
5 changes: 5 additions & 0 deletions vllm_hpu_extension/flags.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,6 +156,11 @@ def enabled_flags():
& Kernel(fsdpa)
& EnvFlag("VLLM_PROMPT_USE_FUSEDSDPA",
Not(ModelType('mllama')))),
"triangular_mask": (Not(Hardware("cpu"))
Copy link
Contributor

@michalkuligowski michalkuligowski Jun 27, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please fix conflict

Copy link
Contributor Author

@kamil-kaczor kamil-kaczor Jun 27, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@michalkuligowski converting to draft - there were issues with triangular and I will only merge it once I confirm its stable.

& Kernel(fsdpa)
& VersionRange(">=1.22.0.294")
& EnvFlag("VLLM_PROMPT_USE_FUSEDSDPA",
Not(ModelType('mllama')))),
"compile_one_hot": (VersionRange(">=1.20.0.370") & Not(EnvFlag("PT_HPU_LAZY_MODE", "1"))),
"flex_attention": (Not(Hardware("cpu")) & Not(EnvFlag("PT_HPU_LAZY_MODE", "1"))
& ModelType("llama")
Expand Down
6 changes: 3 additions & 3 deletions vllm_hpu_extension/ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -274,9 +274,9 @@ def _fsdpa_prompt_attention(
assert attn_bias is not None or valid_seq_lengths is not None, \
'Either attn_bias or valid_seq_lengths must be != None'
if is_causal and attn_bias is not None:
# TODO: causal + attn_bias is not yet supported
is_causal = False
valid_seq_lengths = None
if 'triangular_mask' not in enabled_flags():
is_causal = False
valid_seq_lengths = torch.sum(valid_seq_lengths)
attn_weights = fsdpa_op(query, key, value, attn_bias, 0.0, is_causal,
scale, softmax_mode, recompute_mode,
valid_seq_lengths, 'right')
Expand Down