Skip to content

[v1] Add comments to the new ragged paged attention Pallas kernel #14155

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

Merged
merged 3 commits into from
Mar 3, 2025
Merged

[v1] Add comments to the new ragged paged attention Pallas kernel #14155

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1c511eb
Resolve Alex's remaining PR
vanbasten23 Mar 3, 2025
e28828d
fix linter
vanbasten23 Mar 3, 2025
1ce8b58
Fix comment
mgoin Mar 3, 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
6 changes: 5 additions & 1 deletion vllm/v1/attention/backends/pallas.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@
AttentionLayer, AttentionType)
from vllm.attention.backends.utils import CommonAttentionState

# These are the 2 tunable parameters of the paged attention Pallas kernel.
NUM_QUERIES_PER_BLOCK = 16
NUM_KV_PAGES_PER_BLOCK = 128

Expand Down Expand Up @@ -154,6 +155,9 @@ def forward(
write_to_kv_cache(key, value, key_cache, value_cache, slot_mapping)

query = query * self.scale
# use_kernel switches between using kernel or reference implementation
# (non kernel: https://github.com/pytorch/xla/blob/cee0820e78fc9675e2d0511db891fd44342e890d/torch_xla/experimental/custom_kernel.py#L890).
use_kernel = False
output = torch.ops.xla.ragged_paged_attention(
query,
key_cache,
Expand All @@ -164,7 +168,7 @@ def forward(
attn_metadata.num_seqs,
num_kv_pages_per_block=NUM_KV_PAGES_PER_BLOCK,
num_queries_per_block=NUM_QUERIES_PER_BLOCK,
use_kernel=False,
use_kernel=use_kernel,
)

return output.reshape(num_tokens, hidden_size)
Expand Down