v0.1.8

What's Changed

• kernel: use f32 storage & expf for better precision by @DefTruth in #162
• feat: support attn-mask for triton fwd/bwd by @DefTruth in #163
• feat: support dropout_p for triton fwd/bwd by @DefTruth in #164
• feat: support flash decode for triton bwd Nq=1 by @DefTruth in #165
• kernel: fix philox state for dropout by @DefTruth in #166
• chore: fix docs broken by @DefTruth in #167
• chore: fix docs typo by @DefTruth in #168
• CUDA: clean up cuda kernel codebase by @DefTruth in #169
• kernel: optimize the precision for attn_mask by @DefTruth in #170
• kernel: clean up triton bwd v1 kernel by @DefTruth in #171
• chore: add perf script by @DefTruth in #172
• autotune: fine-grained bucket size for max tune by @DefTruth in #173
• kernel: fix tl.dot dtype for bwd by @DefTruth in #174
• kernel: better load/compute overlap for bwd by @DefTruth in #175
• kernel: optional v grad fp32 buffer for better precision by @DefTruth in #176
• kernel: remove unused num_k_blocks by @DefTruth in #177
• [1/N] autotune: support persistent tuned configs by @DefTruth in #178
• [2/N] autotune: support persistent tuned configs by @DefTruth in #179
• perf: add tflops perf plots by @DefTruth in #180
• [2/N] perf: add tflops perf plots by @DefTruth in #181

Full Changelog: v0.1.7...v0.1.8