v2.0.2

What's Changed

• Refactor cache management in cache_utils.py by @LoserCheems in #278
• [BUG FIX] Update softmax threshold parameter in sparse attention functions by @LoserCheems in #279
• [PERFORMANCE OPTIMIZATION] Optimize alpha/delta layout to eliminate transpose copy in gated attention by @LoserCheems in #280
• [BUG FIX] Varlen Decode Reference Bug Fix by @LoserCheems in #281
• Fix stride bug by @LoserCheems in #282
• Refactor activation functions for optimization by @LoserCheems in #285
• FEATURE] Optimize kernel representations for Triton by @LoserCheems in #286
• Enhance attention kernel performance and quantization support by @LoserCheems in #287
• Refactor attention functions and add autotuning support in Triton by @LoserCheems in #288
• Add optional output and logsumexp tensors to attention functions by @LoserCheems in #289
• Quant backward by @LoserCheems in #290
• Update autotune cache system and local attention by @LoserCheems in #293
• Refactor MASK_CAUSAL parameter to use IS_CAUSAL by @LoserCheems in #295
• Refactor flash_dec_combine to flash_fwd_combine in forward attention … by @LoserCheems in #296
• Fix _reference_scores launch config lookup for sparse/gated and GQA by @LoserCheems in #300
• Fix window_sizes parameter handling and falsy-value default bugs by @LoserCheems in #302

Full Changelog: v2.0.1...v2.0.2

What's Changed

• Refactor cache management in cache_utils.py by @LoserCheems in #278
• [BUG FIX] Update softmax threshold parameter in sparse attention functions by @LoserCheems in #279
• [PERFORMANCE OPTIMIZATION] Optimize alpha/delta layout to eliminate transpose copy in gated attention by @LoserCheems in #280
• [BUG FIX] Varlen Decode Reference Bug Fix by @LoserCheems in #281
• Fix stride bug by @LoserCheems in #282
• Refactor activation functions for optimization by @LoserCheems in #285
• FEATURE] Optimize kernel representations for Triton by @LoserCheems in #286
• Enhance attention kernel performance and quantization support by @LoserCheems in #287
• Refactor attention functions and add autotuning support in Triton by @LoserCheems in #288
• Add optional output and logsumexp tensors to attention functions by @LoserCheems in #289
• Quant backward by @LoserCheems in #290
• Update autotune cache system and local attention by @LoserCheems in #293
• Refactor MASK_CAUSAL parameter to use IS_CAUSAL by @LoserCheems in #295
• Refactor flash_dec_combine to flash_fwd_combine in forward attention … by @LoserCheems in #296
• Fix _reference_scores launch config lookup for sparse/gated and GQA by @LoserCheems in #300
• Fix window_sizes parameter handling and falsy-value default bugs by @LoserCheems in #302

Full Changelog: v2.0.1...v2.0.2