Release Plan of BitBLAS 0.0.1

[LeiWang1999]

Hi all, it's time for us to considering the official release of BitBLAS v0.0.1, here are some todo items before this release:

• Finalize comprehensive test cases and benchmarking scripts.
• Decide on the default policy for the release.
• Performance benchmark results of this release
• Ensure the vLLM PR is either merged or confirm it no longer requires further modifications related to BitBLAS.
• Implement kernel name serialization based on hardware hints and configurations.

[LeiWang1999]

Looking ahead, our future plan for v0.0.2 should include at least support for the Marlin template, quantized Flash Attention, and Group MOE :)

[LeiWang1999]

pr #153 serialized the kernel name with operator config and hint.

[LeiWang1999]

From a policy Perspective, I think we should currently use LOP.3 only for weight propagation, this approach is compatible not only with A100 devices but also with other common devices, such as SM 70 or AMD (even though it’s not currently implemented for AMD, but it could be).

For Stage3 Performance, we can provide option to enable.

Moreover, the incoming stream_k template should share the same weight transformation function with Stage3.

[LeiWang1999]

Think vllm pr vllm-project/vllm#6036 requires no further modifications to BitBLAS, we should consider publishing the formal release.

[LeiWang1999]

PR #249 has successfully passed all test cases; we should now proceed to review the benchmark scripts.