Search results

Large MSE (~4e-3) between Triton and non-Triton outputs on A800

Status: Open.

#31 In mit-han-lab/x-attention;

· mikeallen39 opened on Oct 22, 2025

Clarification Request on the Relationship Between decoding and mode Parameters in find_blocks_chunked

Status: Open.

#30 In mit-han-lab/x-attention;

· mikeallen39 opened on Oct 22, 2025

作者您好，对于 method1.png，即 24x24 的 attention map 计算量变为 24x6，也就是说对于 seq_len 长度的序列，xattn 的保底计算量需求为 seq_len * seq_len / stride * seq_dim，这样理解正确吗？

Status: Open.

#29 In mit-han-lab/x-attention;

· yan-hao-tian opened on Oct 8, 2025

why xattention speed is slower than flash attention

Status: Open.

#28 In mit-han-lab/x-attention;

· DamonsJ opened on Sep 26, 2025

Inference speed test on multi-GPU

Status: Open.

#27 In mit-han-lab/x-attention;

· xwpaul3 opened on Aug 19, 2025

Implementation of Dynamic Programming Approach for Minimum Threshold Prediction (Section 2.3)

Status: Open.

#19 In mit-han-lab/x-attention;

· OrangeEarth15 opened on Jul 13, 2025

prompts about hunyuanvideo running

Status: Open.

#18 In mit-han-lab/x-attention;

· d123f opened on Jul 12, 2025

Errors occur when multi-card testing is adopted

Status: Open.

#17 In mit-han-lab/x-attention;

· qqtang-code opened on Jul 3, 2025

X-Attention Appears to Be Underperforming in Needle-in-a-Haystack Benchmark

Status: Open.

#14 In mit-han-lab/x-attention;

· GITHUBear opened on Jun 13, 2025