v2.0.1

What's Changed

• Add documentation for MkDocs setup and API reference by @LoserCheems in #257
• Remove Public Package Exports section from API reference documentation by @LoserCheems in #258
• Update docstrings in attention functions for consistency by @LoserCheems in #259
• Add return type annotations for attention functions by @LoserCheems in #260
• Init cute version by @LoserCheems in #261
• Update CuTe namespace and enhance dependencies by @LoserCheems in #262
• Add support for upstream split reference in sync scripts by @LoserCheems in #263
• [BUG FIX] Refactor CuTe namespace and enhance sync scripts by @LoserCheems in #264
• [BUG FIX] Optimize LSE computation in forward combine kernel by @LoserCheems in #265
• Update CuTe namespace and functionality by @LoserCheems in #266
• Enhance sync script with cherry-pick functionality and improve merge conflict handling Co-authored-by: Copilot copilot@github.com by @LoserCheems in #267
• Rename triton function by @LoserCheems in #268
• Revert "Rename triton function" by @LoserCheems in #269
• Rename forward combine functions and clarify comments by @LoserCheems in #270
• [FEATURE SUPPORT] Add Triton decode support with KV-cache APIs by @LoserCheems in #271
• Enhance decoding functions with FP8 and quantization support by @LoserCheems in #272
• Fix bug for decode benchmark by @LoserCheems in #273
• Cache optim by @LoserCheems in #274
• [PERFORMANCE OPTIMIZATION] Add compile-time CHECK_NAN toggle to finalize for decode kernel fast-path by @LoserCheems in #275
• [FEATURE SUPPORT] Add HuggingFace Kernel Hub support by @LoserCheems in #276
• Bump version to 2.0.1 by @LoserCheems in #277

Full Changelog: v2.0.0...v2.0.1

What's Changed

• Add documentation for MkDocs setup and API reference by @LoserCheems in #257
• Remove Public Package Exports section from API reference documentation by @LoserCheems in #258
• Update docstrings in attention functions for consistency by @LoserCheems in #259
• Add return type annotations for attention functions by @LoserCheems in #260
• Init cute version by @LoserCheems in #261
• Update CuTe namespace and enhance dependencies by @LoserCheems in #262
• Add support for upstream split reference in sync scripts by @LoserCheems in #263
• [BUG FIX] Refactor CuTe namespace and enhance sync scripts by @LoserCheems in #264
• [BUG FIX] Optimize LSE computation in forward combine kernel by @LoserCheems in #265
• Update CuTe namespace and functionality by @LoserCheems in #266
• Enhance sync script with cherry-pick functionality and improve merge conflict handling Co-authored-by: Copilot copilot@github.com by @LoserCheems in #267
• Rename triton function by @LoserCheems in #268
• Revert "Rename triton function" by @LoserCheems in #269
• Rename forward combine functions and clarify comments by @LoserCheems in #270
• [FEATURE SUPPORT] Add Triton decode support with KV-cache APIs by @LoserCheems in #271
• Enhance decoding functions with FP8 and quantization support by @LoserCheems in #272
• Fix bug for decode benchmark by @LoserCheems in #273
• Cache optim by @LoserCheems in #274
• [PERFORMANCE OPTIMIZATION] Add compile-time CHECK_NAN toggle to finalize for decode kernel fast-path by @LoserCheems in #275
• [FEATURE SUPPORT] Add HuggingFace Kernel Hub support by @LoserCheems in #276
• Bump version to 2.0.1 by @LoserCheems in #277

Full Changelog: v2.0.0...v2.0.1

v2.0.0

What's Changed

• Improve numerical stability in sparse attention with sink auxiliary logits by @LoserCheems in #220
• [PERFORMANCE OPTIMIZATION] Flash Sparse Attention by @LoserCheems in #221
• [BUG FIX] Refactor block min/max calculations by @LoserCheems in #223
• [BUG FIX] Improve packed GQA handling by @LoserCheems in #224
• Add utility functions for device management and input validation by @LoserCheems in #225
• [PERFORMANCE OPTIMIZATION] Triton Sparse Base Forward Kernel with Gate-Based Sparsity by @LoserCheems in #226
• [FEATURE] Enhance forward combine kernel and split attention by @LoserCheems in #227
• Improves softmax stability with log2 scaling by @LoserCheems in #228
• Renames variables and refactors functions for clarity by @LoserCheems in #229
• Improve performance and configuration for SM90 forward path by @LoserCheems in #231
• Refactor rescaling logic in online_softmax and rescale_o functions by @LoserCheems in #232
• [BUG FIX] Improve forward kernel configuration and validation by @LoserCheems in #233
• Refactor qheads_per_kvhead calculations for clarity by @LoserCheems in #234
• [FEATURE SUPPORT] Add Triton backward support by @LoserCheems in #235
• [FEATURE SUPPORT] Add Configurable Sparse Gate Modes and Adaptive Thresholding in Triton Forward Kernel by @LoserCheems in #236
• Refactor log_sigmoid function for improved performance and accuracy by @LoserCheems in #237
• [FEATURE SUPPORT] Add Configurable Sparse Gate Modes and Adaptive Thresholding in Triton Backward Kernel by @LoserCheems in #238
• Enhance forward kernel for block range and masking logic by @LoserCheems in #239
• Refactor backward kernels for clarity and optimization by @LoserCheems in #240
• [BUG FIX] Update launch configuration for RTX Pro 6000 by @LoserCheems in #241
• Add benchmark functions for Triton attention operations by @LoserCheems in #242
• [FEATURE SUPPORT] Enable Softmax-Threshold Block Skipping in Triton Dense/Sparse Forward Attention by @LoserCheems in #243
• [BUG FIX] Improve clarity and accuracy in gating mechanisms by @LoserCheems in #244
• [BUG FIX] Update stride parameters for consistency by @LoserCheems in #245
• Add softmax threshold parameter for enhanced flexibility by @LoserCheems in #246
• [FEATURE] Implement dense attention with masking support by @LoserCheems in #247
• Enhance sparse attention implementation and documentation by @LoserCheems in #248
• [FEATURE] Implement gated attention mechanism and enhance performance by @LoserCheems in #249
• Update project structure and dependencies by @LoserCheems in #250
• [BUG FIX] Improve error reporting and occupancy in benchmarks by @LoserCheems in #251
• Update repository URLs and improve documentation by @LoserCheems in #252
• Refactor benchmark tests to simplify tensor initialization by @LoserCheems in #253
• Refactor test utilities and add CUDA tensor operation tests by @LoserCheems in #254
• Refactor masking logic in backward kernel functions by @LoserCheems in #255
• Refactor GitHub Actions workflows for package building and publishing by @LoserCheems in #256

Full Changelog: v1.2.4...v2.0.0

What's Changed

• Improve numerical stability in sparse attention with sink auxiliary logits by @LoserCheems in #220
• [PERFORMANCE OPTIMIZATION] Flash Sparse Attention by @LoserCheems in #221
• [BUG FIX] Refactor block min/max calculations by @LoserCheems in #223
• [BUG FIX] Improve packed GQA handling by @LoserCheems in #224
• Add utility functions for device management and input validation by @LoserCheems in #225
• [PERFORMANCE OPTIMIZATION] Triton Sparse Base Forward Kernel with Gate-Based Sparsity by @LoserCheems in #226
• [FEATURE] Enhance forward combine kernel and split attention by @LoserCheems in #227
• Improves softmax stability with log2 scaling by @LoserCheems in #228
• Renames variables and refactors functions for clarity by @LoserCheems in #229
• Improve performance and configuration for SM90 forward path by @LoserCheems in #231
• Refactor rescaling logic in online_softmax and rescale_o functions by @LoserCheems in #232
• [BUG FIX] Improve forward kernel configuration and validation by @LoserCheems in #233
• Refactor qheads_per_kvhead calculations for clarity by @LoserCheems in #234
• [FEATURE SUPPORT] Add Triton backward support by @LoserCheems in #235
• [FEATURE SUPPORT] Add Configurable Sparse Gate Modes and Adaptive Thresholding in Triton Forward Kernel by @LoserCheems in #236
• Refactor log_sigmoid function for improved performance and accuracy by @LoserCheems in #237
• [FEATURE SUPPORT] Add Configurable Sparse Gate Modes and Adaptive Thresholding in Triton Backward Kernel by @LoserCheems in #238
• Enhance forward kernel for block range and masking logic by @LoserCheems in #239
• Refactor backward kernels for clarity and optimization by @LoserCheems in #240
• [BUG FIX] Update launch configuration for RTX Pro 6000 by @LoserCheems in #241
• Add benchmark functions for Triton attention operations by @LoserCheems in #242
• [FEATURE SUPPORT] Enable Softmax-Threshold Block Skipping in Triton Dense/Sparse Forward Attention by @LoserCheems in #243
• [BUG FIX] Improve clarity and accuracy in gating mechanisms by @LoserCheems in #244
• [BUG FIX] Update stride parameters for consistency by @LoserCheems in #245
• Add softmax threshold parameter for enhanced flexibility by @LoserCheems in #246
• [FEATURE] Implement dense attention with masking support by @LoserCheems in #247
• Enhance sparse attention implementation and documentation by @LoserCheems in #248
• [FEATURE] Implement gated attention mechanism and enhance performance by @LoserCheems in #249
• Update project structure and dependencies by @LoserCheems in #250
• [BUG FIX] Improve error reporting and occupancy in benchmarks by @LoserCheems in #251
• Update repository URLs and improve documentation by @LoserCheems in #252
• Refactor benchmark tests to simplify tensor initialization by @LoserCheems in #253
• Refactor test utilities and add CUDA tensor operation tests by @LoserCheems in #254
• Refactor masking logic in backward kernel functions by @Loserc...

v1.2.4

Last attn_mask version

We will adopt a new strategy to alleviate the memory bottleneck of attn_mask. This is the last version with attn_mask. Future versions will not pass attn_mask.

What's Changed

• Chore/sync after move by @LoserCheems in #208
• [BUG FIX] Unify masking utilities and improve performance by @LoserCheems in #209
• Corrects issue links in README guides by @LoserCheems in #212
• [BUG FIX] Correct causal mask handling for longer KV pairs by @LoserCheems in #213
• Add gradient computation for bias and token-level KV sparsity support by @LoserCheems in #214
• Add rotary-aware attention modules for improved inference by @LoserCheems in #215
• Improve code readability and linting workflow by @LoserCheems in #216
• Simplify attention mechanisms by @LoserCheems in #217
• Refactor create_mask function parameters by @LoserCheems in #218

Full Changelog: v1.2.3...v1.2.4

What's Changed

• Chore/sync after move by @LoserCheems in #208
• [BUG FIX] Unify masking utilities and improve performance by @LoserCheems in #209
• Corrects issue links in README guides by @LoserCheems in #212
• [BUG FIX] Correct causal mask handling for longer KV pairs by @LoserCheems in #213
• Add gradient computation for bias and token-level KV sparsity support by @LoserCheems in #214
• Add rotary-aware attention modules for improved inference by @LoserCheems in #215
• Improve code readability and linting workflow by @LoserCheems in #216
• Simplify attention mechanisms by @LoserCheems in #217
• Refactor create_mask function parameters by @LoserCheems in #218

Full Changelog: v1.2.3...v1.2.4

v1.2.3

What's Changed

• Add selectable masking strategies for attention by @LoserCheems in #204
• Refactor attention block smoothing for consistency by @LoserCheems in #205
• Optimize triton version: GQA, mask/bias broadcasting, skip inactive tiles, and stability fixes by @LoserCheems in #200
• [FEATURE SUPPORT] Triton special compact dynamic-mask attention: 1.6× faster fwd+bwd, numerically equivalent by @LoserCheems in #206
• Fix documentation and references for Flash Sparse Attention by @LoserCheems in #207

Full Changelog: v1.2.2...v1.2.3

What's Changed

• Add selectable masking strategies for attention by @LoserCheems in #204
• Refactor attention block smoothing for consistency by @LoserCheems in #205
• Optimize triton version: GQA, mask/bias broadcasting, skip inactive tiles, and stability fixes by @LoserCheems in #200
• [FEATURE SUPPORT] Triton special compact dynamic-mask attention: 1.6× faster fwd+bwd, numerically equivalent by @LoserCheems in #206
• Fix documentation and references for Flash Sparse Attention by @LoserCheems in #207

Full Changelog: v1.2.2...v1.2.3

v1.2.2

What's Changed

• [FEATURE SUPPORT] Robust dBias accumulation for seqlen_q_bias == 1 by @LoserCheems in #194
• [FEATURE SUPPORT] Centralize dynamic mask creation for FDMA by @LoserCheems in #197
• Update documentation to use mask utility in examples by @LoserCheems in #198
• Fix attention bias calculation and dbias handling by @LoserCheems in #199
• Add block-wise smoothing to attention mask by @LoserCheems in #201
• [FEATURE SUPPORT] Move scaling out of streaming loops, bias-initialized acc_s, and fix dQ double-scaling by @LoserCheems in #203

Full Changelog: v1.2.1...v1.2.2

v1.2.1

What's Changed

• Implement variable-length attention with mask and bias support by @LoserCheems in #185
• Add issue/PR templates by @LoserCheems in #186
• [FEATURE SUPPORT] Variable-Length Attention with Padding-Free Execution by @LoserCheems in #188
• [FEATURE SUPPORT] Broadcastable 4D mask/bias, 128‑rounded key length, stride‑0 broadcasting, and dbias reductions by @LoserCheems in #190
• Refactor bias initialization and enhance bias computation in FlashDMAttnFunc by @LoserCheems in #191
• Fix attention_mask and attention_bias shape descriptions and remove redundant checks by @LoserCheems in #192
• Enhance bias gradient accumulation in backward pass by @LoserCheems in #193

Full Changelog: v1.2.0...v1.2.1

v1.2.0

What's Changed

• [BUG FIX] Fix mask/bias memory access and vectorization issues in kernels by @LoserCheems in #182

Full Changelog: v1.1.9...v1.2.0

v1.1.9

What's Changed

• Refactor attention mask and bias handling for efficiency by @LoserCheems in #177
• [BUG FIX] SM80 NaN in bias.grad when both mask and bias are enabled by @LoserCheems in #179

Full Changelog: v1.1.8...v1.1.9

v1.1.8

What's Changed

• Bump version to 1.1.8 by @LoserCheems in #176

Full Changelog: v1.1.7...v1.1.8

v1.1.7

What's Changed

• Increase GitHub Actions build timeout to 6 hours by @LoserCheems in #175

Full Changelog: v1.1.6...v1.1.7