[NOTICE] Deprecation of old flow

[tqchen]

As we are transitioning to the SLM flow, we plan to phase out the old compilation flow in mlc_llm folder. The new flow would enable a single unified model definition across all hardware backends, including server and local settings.

This is a notice that we plan to do it in about a week(around Mar 10th).

Generally all models supported in the old flow are now ready in the new flow. We will create a backup branch(legacy) so the old flow can remain visible for a while. This also allows us to be able to reuse mlc_llm package name for future new packages. Thanks everyone for working with us through this process. Feel free to make use of this issue for followup questions

[tqchen]

Update on this issue: over the past week, we looked into all the currently known gaps. Here are the status

• FT optimizations should now align when the cudagraph optimization lands in [Relax] CUDA graph rewrite treating StringImm as static apache/tvm#16691 and [Runtime] PagedKVCache execute data copy on a separate stream apache/tvm#16692. Both should lands into the wheel nightly soon.
• Android we have aligned the attention perf in [Question] "import tvm" error in python #915
• We also know that there are manual dispatch kernels in q4f16_0 by @srkreddy1238 that can be migrated into the new flow

Overall, we find that new flow have potential to do better in longer context due to the flashinfer style recursive attention, and unifying the serving and single batch case.

As planned, we will remove the old flow. This would also free up the package name mlc_llm which can be used for the package in the new flow.

Thanks everyone for bearing through this transition process, we are getting close to the stablized interface with a lot more fun cabilities ahead.

[tqchen]

• Backup branch created https://github.com/mlc-ai/mlc-llm/tree/backup-before-old-flow-deprecation
• PR that runs deprecation Deprecate old flow #1928

[tqchen]

Followup #1929