Skip to content

fix gpt_oss/triton/attention.py bug #230

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

Open
xjmxyt wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix gpt_oss/triton/attention.py bug #230

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3be653f
fix attention bug
xjmxyt Dec 3, 2025
d6c5ccb
fix bug
xjmxyt Dec 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: 4 additions & 2 deletions gpt_oss/triton/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,7 @@ def _attn_fwd(
lo, hi = tl.maximum(start_q, start_q + start_m * BLOCK_M - BANDWIDTH), start_q + (start_m + 1) * BLOCK_M
else:
lo, hi = start_q, start_q + (start_m + 1) * BLOCK_M
hi = tl.minimum(hi, N_KV_CTX)

for start_n in range(lo, hi, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
Expand Down Expand Up @@ -181,6 +182,7 @@ def attention_ref(
pos_keys = torch.arange(num_keys, device=query.device)
pos_queries = torch.arange(num_queries, device=query.device) + start_q
mask = pos_keys[None, :] > pos_queries[:, None]
mask = mask | (pos_keys[None, :] < start_q)
mask = mask.float().masked_fill(mask, float("-inf"))
Comment on lines 182 to 185
Copy link

@chatgpt-codex-connector chatgpt-codex-connector bot Dec 3, 2025

Choose a reason for hiding this comment

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

P1 Badge Mask now drops cached prefix tokens

The new mask mask = mask | (pos_keys[None, :] < start_q) causes the reference path to drop every key with index < start_q. During cached decoding we set start_q to the cache offset (see AttentionBlock.forward in gpt_oss/triton/model.py lines 218-253) while passing the full cache as k/v, so this change makes n_ctx==1 or small-context calls ignore all previously cached tokens and attend only to the current block. That silently breaks causal attention whenever offset > 0, producing wrong outputs for generation with a warm cache.

Useful? React with 👍 / 👎.

Copy link
Author

@xjmxyt xjmxyt Dec 3, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

Solve this by matching kernel implementation to reference instead of modifying reference.


if sliding_window:
Expand Down Expand Up @@ -211,7 +213,7 @@ def attention_ref(
@pytest.mark.parametrize("head_dim", [64])
@pytest.mark.parametrize("sm_scale", [0.125])
@pytest.mark.parametrize("sliding_window", [None, 128])
@pytest.mark.parametrize("start_q", [0, 5])
@pytest.mark.parametrize("start_q", [0, 64])
def test_eq(batch_size, num_queries, num_keys, num_key_value_heads, num_key_value_groups, head_dim, sm_scale, sliding_window, start_q):
if num_queries > num_keys:
pytest.skip("too many queries")
Expand All @@ -226,4 +228,4 @@ def test_eq(batch_size, num_queries, num_keys, num_key_value_heads, num_key_valu
o1 = attention(q, k, v, sinks, sm_scale, sliding_window, start_q)
o2 = attention_ref(q, k, v, sinks, sm_scale, sliding_window, start_q)

torch.testing.assert_close(o1, o2)
torch.testing.assert_close(o1, o2)