MoonshotAI/FlashKDA

View on GitHub

FlashKDA: high-performance Kimi Delta Attention kernels

Stars: 1.3k
Forks: 122
Open beginner issues: 0
Indexed issues: 14

Dominant language: Cuda
License: MIT
Last GitHub push: Sep 1, 2026
Latest indexed: Sep 20, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed
PR merge metrics: No merged PRs in 30d

Beginner friendly Everything

Open All

14 open issues indexed

What is the motivation behind replacing the fp16 Neumann inverse with an 8x8 fp32 forward substitution Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

MoonshotAI/FlashKDA#34 · 1 comment · Opened Sep 11, 2026
Add a differentiable CPU backend for KDA research and prototyping Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

MoonshotAI/FlashKDA#26 · Opened Jul 29, 2026
Add Pythonic API wrapper with automatic workspace management Open

Difficulty 3/5 1-2 days Newbie friendliness 68/100

MoonshotAI/FlashKDA#23 · Opened Jul 28, 2026
Add random seed to bench_fwd.py for reproducible benchmarks Open

Difficulty 1/5 Under an hour Newbie friendliness 88/100

MoonshotAI/FlashKDA#22 · Opened Jul 28, 2026
Add standalone usage example and inference-only note Open

Difficulty 2/5 1-3 hours Newbie friendliness 85/100

MoonshotAI/FlashKDA#21 · Opened Jul 28, 2026
Centralize CHUNK and D constants as single source of truth Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

MoonshotAI/FlashKDA#20 · Opened Jul 28, 2026
Surface CUTLASS initialization and CUDA launch-configuration failures Open

Difficulty 4/5 3-5 days Newbie friendliness 58/100

MoonshotAI/FlashKDA#15 · Opened Jul 27, 2026
Partial output tiles serialize stores through one lane Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

MoonshotAI/FlashKDA#14 · Opened Jul 27, 2026
[Bug] test_fwd failed with CUDA 13 Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

MoonshotAI/FlashKDA#12 · Opened Jul 17, 2026
Publish FlashKDA kernels on Hugging Face Kernels? Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

MoonshotAI/FlashKDA#11 · 6 comments · Opened Jul 15, 2026
[Feature] Intra-Card Context Parallelism to improve B=1 long-sequence performance Open

Difficulty 5/5 Over a week Newbie friendliness 45/100

MoonshotAI/FlashKDA#8 · 3 reactions · Opened May 29, 2026
[Question] Roadmap for supporting additional K/V head dimensions and backward kernels Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

MoonshotAI/FlashKDA#4 · 3 comments · Opened Apr 24, 2026
gate mode diff Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

MoonshotAI/FlashKDA#3 · 4 comments · Opened Apr 23, 2026
intermidiate ssm state store support Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

MoonshotAI/FlashKDA#2 · Opened Apr 23, 2026

MoonshotAI/FlashKDA

Get new issues in your inbox