[Question] Support head_dim=256 for attention ops (Qwen3.5/3.6-style GDN-hybrid models) ？
Nobody has claimed this yet.
Assessment
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Newbie friendliness
- 48/100
- Issue type
- Feature
- Clarity
- Mostly clear
- Activity status
- Quiet
- Tech stack
- cpp
- Domain
- machine-learning, performance
Research direction
Start by locating the attention operator implementation and the existing head_dim=80/128 handling. Trace how attention configurations are validated, then check whether a head_dim=256 configuration can run for the cited Qwen3.5/3.6-style full-attention layers. Done means the new configuration is supported and its end-to-end latency impact is benchmarked.
Written by the indexing model from the issue text.
Description
-
The attention ops currently support only head_dim = 80 / 128.Could head_dim=256 be added?
-
Recent GDN-hybrid models (Qwen3.5 / 3.6) use head_dim=256 in their full-attention layers, so this would unblock those configs.
-
One question: since GDN (linear-attention) layers dominate these models,how much end-to-end latency benefit can we realistically expect from optimizing the full-attention op here? Has anyone benchmarked hpc_ops attention on a Qwen3.5/3.6-class model?
-
Happy to help test if a branch is available.
- Dominant language
- C++
- Stars
- 1.2k
- Forks
- 142
- Avg merge
- 23h 15m
- Merged PRs (30d)
- 8
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from Tencent/hpc-ops
-
Difficulty 5/5 Over a week Newbie friendliness 35/100
-
Difficulty 5/5 Over a week Newbie friendliness 25/100
-
[Question] How to Understand the thr_layout and val_layout in make_tiled_copy_V_interleave_trans? Open
Difficulty 4/5 3-5 days Newbie friendliness 35/100
-
Difficulty 4/5 3-5 days Newbie friendliness 45/100
-
Difficulty 4/5 3-5 days Newbie friendliness 48/100
Similar issues
-
Difficulty 2/5 1-3 hours Newbie friendliness 86/100
-
Sensor initialization takes very long when `--initial-sim-time` is set to current UNIX timestamp Open
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
gazebosim/gz-sensors#662 · 1 comment ·
-
enhancement
Difficulty 2/5 1-3 hours Newbie friendliness 76/100
-
comp-datalake
Difficulty 2/5 1-3 hours Newbie friendliness 88/100
ClickHouse/ClickHouse#121222 ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
LadybirdBrowser/ladybird#12123 ·