[RFE] 🦗 SDPA Support Matrix: Hopper (SM90) — open-source kernels wanted!

[vedaanta]

🦗 Mission board: Hopper (SM90)

We're building open-source SDPA (flash attention) kernels for cuDNN frontend, and this is the coverage board for Hopper (SM90).

Right now the board is dark — every cell is up for grabs. Claim one, land it, and light it up. 🔦

Legend

Symbol	Meaning
✅	Shipped & validated
🚧	In progress — someone's on it
❌	Open — claim it!
🚫	Not applicable on this architecture

The board

Feature	Forward	Backward
— Data types —
FP16	❌	❌
BF16	❌	❌
FP8 (E4M3 / E5M2)	❌	❌
MXFP8	❌	❌
— Head dims —
d = 64	❌	❌
d = 128	❌	❌
d = 192/128	❌	❌
d = 256	❌	❌
d = 512	❌	❌
— Masks —
Causal	❌	❌
Causal bottom-right (diagonal_alignment)	❌	❌
Sliding window (SWA)	❌	❌
Padding mask (per-batch seq lens)	❌	❌
— Other —
GQA / MQA	❌	❌
Attention bias (additive)	❌	❌
ALiBi	❌	❌
Attention sink	❌	❌
THD / varlen (packed cu_seqlens layout)	❌	❌
Dropout	❌	❌
Paged attention (KV cache)	❌	❌

How to play

1. 💬 Comment below with the cell(s) you want to tackle.
2. 🚧 We'll assign you and flip the cell to in progress.
3. ✅ Land the PR with tests — cell lights up, scoreboard ticks. 🎉

Scoreboard: 0 / 40 cells lit.

Sister boards: ⚡ Ampere (SM80) #377 · 🚀 Blackwell (SM100) #379 · 💫 Blackwell Ultra (SM103) #380 · ✨ Blackwell GeForce / DGX Spark (SM120/SM121) #381