Add Flash Attention Support to Deepseek v4 - #4488

Open

octatrifan wants to merge 15 commits into

mainfrom

octatrifan-dsv4-flash-attention

Commits on Jul 17, 2026

Support dynamic Tokamax Flash Attention for DeepSeek-V4 HCA and CSA layers
octatrifan
committed
Remove deepseek4-small config and registration
octatrifan
committed
Add unit tests for CompressedAttention (DeepSeek-V4) compilation and execution
octatrifan
committed

Commits on Jul 18, 2026

Enable Flash Attention support for DeepSeek-V4 (MLA) and add HCA/CSA unit tests
octatrifan
committed

Commits on Jul 20, 2026

Commits on Jul 24, 2026

Support non-multiple-of-128 sequence lengths in DeepSeek-V4 HCA/CSA Flash Attention
octatrifan
committed

Commits on Jul 25, 2026

Revert experimental rotary_embedding.head_dim check in favor of stock PR 4488 config.head_dim
octatrifan
committed

Commits on Jul 26, 2026

Commits on Jul 27, 2026

Commits on Jul 28, 2026