What's Changed

• fix sage_fp16_triton repeat by @ITerydh in #146
• bump to 0.6.3.post1 by @feifeibear in #147
• fix: remove redundant code to avoid error in specific env by @ZDJeffrey in #150
• fix: flashattention3 call by @yuyu5333 in #151
• [feature] adapt for Moore Threads graphics processing unit by @houchen-li in #152
• [bugfix] correct the operator alias in attention.py by @houchen-li in #153
• Fix backward gradient count mismatch by @MartinPernus in #156
• Add AttnType.AITER by @kTorp in #159
• Adapted for Huawei Ascend npu by @endymion-ni in #162
• Adapted Ascend npu for LongContextAttention\UlyssesAttention\RingAttention calulation based on torch_npu.npu_fusion_attention_v2 & npu_fusion_attention_grad_v2 for trainning and inference case by @L4-1024 in #167
• Refine AttnType.TORCH by @genghisun in #165
• bump to 0.6.4 by @feifeibear in #168

New Contributors

• @ITerydh made their first contribution in #146
• @yuyu5333 made their first contribution in #151
• @houchen-li made their first contribution in #152
• @MartinPernus made their first contribution in #156
• @kTorp made their first contribution in #159
• @endymion-ni made their first contribution in #162
• @L4-1024 made their first contribution in #167
• @genghisun made their first contribution in #165

Full Changelog: 0.6.3...0.6.4