Issues · meta-pytorch/attention-gym · GitHub

Labels Milestones New issue

Better Flex Attention BlockMask docs

#197

· man2machine opened

on Mar 31, 2026

FlexAttn Autograd

#194

· jeromeku opened

on Jan 26, 2026

`get_mask_mod` and `get_score_mod` produces incorrect logical to physical translation.

#189

· man2machine opened

on Jan 6, 2026

routing to sdpa_dense and flex_attention_autograd

#179

· LJHG opened

on Nov 13, 2025

Is it possible to output the values for multiple scores with flex attention?

#174

· t0278611 opened

on Nov 3, 2025

Flex Attention runs slower than FlashAttention-2

#172

· Yangchenchen-Jin opened

on Oct 29, 2025

Flex Attention does not support learnable scalar?

#171

· DylanLIiii opened

on Oct 27, 2025

Support for differentiable soft attention masks (float mask in [0,1], not in log-space)

#167

· enszhou opened

on Oct 7, 2025

FlexAttention + custom score_mod fails under torch.compile (FlexibleLayout / NoValidChoicesError)

#166

· v1kstrand opened

on Sep 18, 2025

[Question] src/tcmalloc.cc:283] Attempt to free invalid pointer 0x13d801860

#161

· JasperChennn opened

on Aug 14, 2025

`_compile=True` blows up GPU memory when seq length / block_mask change, but is best when both are fixed (FlexAttention)

#159

· wileewang opened

on Jul 22, 2025

[Question] sliding window attention + dynamic input shapes

#138

· pfeatherstone opened

on May 6, 2025