Skip to content

[MLA] Address Triton decode KV rows with int64 ids #2533

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
zejunchen-zejun merged 2 commits into from
Oct 16, 2026
Merged

[MLA] Address Triton decode KV rows with int64 ids #2533

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4229c75
[MLA] Address Triton decode KV rows with int64 ids
zejunchen-zejun Oct 11, 2026
ef06615
Merge branch 'main' into zejun/triton-mla-int64-block-table
zejunchen-zejun Oct 16, 2026
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 atom/model_ops/attentions/triton_mla.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,10 +40,14 @@ def __init__(self, model_runner):
kv_lora_rank = hf.kv_lora_rank
num_kv_splits = 4
triton_mla_buffers = {
# int64, not int32: aiter's Triton decode multiplies each KV row id by
# the row stride (576) in the id's own dtype. An int32 offset wraps past
# row 2**31 / 576 = 3,728,270 (block 233,016 of 16 tokens), and the
# decode then reads outside the KV cache.
"triton_block_table": torch.zeros(
self.max_bs,
self.max_num_blocks_per_seq,
dtype=torch.int32,
dtype=torch.int64,

@zejunchen-zejun zejunchen-zejun Oct 14, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

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

int64 covers the range that int32 express.

device=self.device,
),
"triton_attn_logits": torch.empty(
Expand Down