Skip to content

[Feat][Linear-Attn] Update GDN prefill manifest contract #1601

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
lcy-seso merged 2 commits into from
Jun 21, 2026
+4 −2
Merged

[Feat][Linear-Attn] Update GDN prefill manifest contract #1601

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
efaefb5
[Spec] Update GDN prefill manifest contract
superAngGao Jun 21, 2026
af4aedd
[Chore][Linear-Attn] Retrigger manifest PR checks
superAngGao Jun 21, 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: 4 additions & 2 deletions tileops/manifest/linear_attention.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,8 @@ GatedDeltaNetPrefillFwdOp:
o: {dtype: "same_as(q)"}
final_state: {dtype: "same_as(q)"}
params:
chunk_size: {type: int, default: 64}
chunk_size: {type: "int | None", default: null}
layout: {type: str, default: "bhtd"}
Comment thread
lcy-seso marked this conversation as resolved.
shape_rules:
- "q.shape == (B, H, S, DK)"
- "k.shape == (B, H, S, DK)"
Expand All @@ -32,13 +33,14 @@ GatedDeltaNetPrefillFwdOp:
- "beta.shape == (B, H, S)"
- "o.shape == (B, H, S, DV)"
- "final_state.shape == (B, H, DK, DV)"
- "S % chunk_size == 0"
- "chunk_size is None or S % chunk_size == 0"

workloads:
# Inference-oriented BHSD rows. Unit tests should use smaller
# branch-covering shapes in tests/ops, not these benchmark workloads.
- {q_shape: [1, 16, 512, 128], k_shape: [1, 16, 512, 128], v_shape: [1, 16, 512, 128], g_shape: [1, 16, 512], beta_shape: [1, 16, 512], chunk_size: 64, dtypes: [float16, bfloat16], label: "gdn-prefill-b1-s512-h16-d128"}
- {q_shape: [4, 16, 2048, 64], k_shape: [4, 16, 2048, 64], v_shape: [4, 16, 2048, 128], g_shape: [4, 16, 2048], beta_shape: [4, 16, 2048], chunk_size: 64, dtypes: [float16, bfloat16], label: "gdn-prefill-b4-s2k-h16-dk64-dv128"}
- {q_shape: [1, 16, 2048, 128], k_shape: [1, 16, 2048, 128], v_shape: [1, 16, 2048, 128], g_shape: [1, 16, 2048], beta_shape: [1, 16, 2048], chunk_size: 64, dtypes: [float16, bfloat16], label: "gdn-prefill-b1-s2k-h16-d128"}
Comment thread
lcy-seso marked this conversation as resolved.
- {q_shape: [1, 16, 8192, 128], k_shape: [1, 16, 8192, 128], v_shape: [1, 16, 8192, 128], g_shape: [1, 16, 8192], beta_shape: [1, 16, 8192], chunk_size: 64, dtypes: [float16, bfloat16], label: "gdn-prefill-b1-s8k-h16-d128"}

roofline:
Expand Down
Loading