V0.1.6: Add SCBench

What's Changed

Feature

• [PreRelease]: SCBench by @iofu728 in #96
• Feature(MInference): support transformers>=4.46.0, add chunk mlp by @iofu728 in #133, #113
• Feature(MInference): update release cuda version by @iofu728 in #134
• Fix(MInference): fix the search pattern feature by @iofu728 in #156

Ops

• Feature(FlexPrefill): add flex-prefill by @liyucheng09 in #100
• Feature(MInference): add xAttention by @iofu728 in #149
• Feature(MInference): support SGLang and vLLM vertical_and_slash flash attention and index kernels by @iofu728 in #153

Model support

• Feature(MInference): support LLaMA-3-70B-1M and multi-gpu PP by @iofu728 in #59
• Feature(MInference): add Qwen-Turbo-1M by @iofu728 in #115, #117, #135
• Feature(MInference): add kv_type unittest by @iofu728 in #132

Bug Fix

• Fix(MInference): fix e2e benchmark guideline & fix A-shape multi gpu by @iofu728 in #66
• Fix(MInference): fix the vs pattern loss / sqrt(dk) by @iofu728 in #70
• Fix(SCBench): fix the pipeline and load dataset by @iofu728 in #101
• Fix(SCBench): fix default seq length by @iofu728 in #104
• Feature(SCBench): update scbench scripts by @iofu728 in #105
• Fix(SCBench): bug fix when use_cache == False by @liyucheng09 in #106
• Feature(MInference): support 192 dim for streaming_kernel by @iofu728 in #125
• Fix the Qwen config path in model2path.py by @GuoYiFantastic in #130
• Fix(MInference): fix get config by @iofu728 in #121
• Fix(MInference): fix the residual by @iofu728 in #136
• Fix(MInference): fix multi worker by @iofu728 in #152
• fix typo by @DefTruth in #141

New Contributors

• @GuoYiFantastic made their first contribution in #130
• @DefTruth made their first contribution in #141

Full Changelog: v0.1.5.post1...v0.1.6