Search results

[Feature Request] Support head_dim=256 on SM100 (Blackwell)

Status: Open.

#309 In SandAI-org/MagiAttention;

· LucienXian opened on Apr 7, 2026

Does magi support zigzag dispatch pattern?

Status: Open.

#196 In SandAI-org/MagiAttention;

· jordane95 opened on Dec 29, 2025

TypeError: flex_flash_attn_func() got an unexpected keyword argument 'max_seqlen_q'

Status: Open.

#193 In SandAI-org/MagiAttention;

· hengrui0516 opened on Dec 16, 2025

Are there plans to support sparse attention inference on 4090 and 5090?

Status: Open.

#184 In SandAI-org/MagiAttention;

· helloyongyang opened on Dec 8, 2025

Using dummy/padded attention slices with flex_flash_attn_func for torch.compile static shapes

Status: Open.

#180 In SandAI-org/MagiAttention;

· keunhong opened on Dec 2, 2025

Support 256 Head Dim

Status: Open.

#165 In SandAI-org/MagiAttention;

· lin-ht opened on Nov 5, 2025

[Support for Cross-Attention] How to use cross-attention with magiattention

Status: Open.

#123 In SandAI-org/MagiAttention;

· zelingao98 opened on Aug 26, 2025

NSA Support

Status: Open.

#114 In SandAI-org/MagiAttention;

· jensenzhoujh opened on Aug 11, 2025

Blackwell support

Status: Open.

#105 In SandAI-org/MagiAttention;

· littsk opened on Aug 2, 2025

Dynamic Load Blancer

Status: Open.

#104 In SandAI-org/MagiAttention;

· littsk opened on Aug 2, 2025

Nvshmem Based GroupCast & GroupReduce

Status: Open.

#103 In SandAI-org/MagiAttention;

· littsk opened on Aug 2, 2025

Performance Optimize For Sparse Attention

Status: Open.

#99 In SandAI-org/MagiAttention;

· littsk opened on Aug 1, 2025