Bump v2.6.0

README.md

@@ -314,6 +314,11 @@ Implement deterministic backward pass. Thanks to engineers from [Meituan](www.me
 Support paged KV cache (i.e., [PagedAttention](https://arxiv.org/abs/2309.06180)).
 Thanks to @beginlner for this contribution.
 
+### 2.6: Softcapping.
+
+Support attention with softcapping, as used in Gemma-2 and Grok models.
+Thanks to @Narsil for this contribution.
+
 ## Performance
 
 We present expected speedup (combined forward + backward pass) and memory savings from using FlashAttention against PyTorch standard attention, depending on sequence length, on different GPUs (speedup depends on memory bandwidth - we see more speedup on slower GPU memory).

flash_attn/__init__.py

-__version__ = "2.5.9.post1"
+__version__ = "2.6.0"
 
 from flash_attn.flash_attn_interface import (
     flash_attn_func,

training/Dockerfile

@@ -85,7 +85,7 @@ RUN pip install transformers==4.25.1 datasets==2.8.0 pytorch-lightning==1.8.6 tr
 RUN pip install git+https://github.com/mlcommons/logging.git@2.1.0
 
 # Install FlashAttention
-RUN pip install flash-attn==2.5.9.post1
+RUN pip install flash-attn==2.6.0
 
 # Install CUDA extensions for fused dense
-RUN pip install git+https://github.com/HazyResearch/flash-attention@v2.5.9.post1#subdirectory=csrc/fused_dense_lib
+RUN pip install git+https://github.com/HazyResearch/flash-attention@v2.6.0#subdirectory=csrc/fused_dense_lib